RevOps runs on rules of thumb and pattern matching, many of them learned under constraints that AI is now changing. Entirely new priors are taking shape right now, but it's still useful to keep tabs on the old ones and see which hold up. I keep this running page as a reminder to myself, so my mental model moves when the evidence does. Each entry gives the verdict, what decides it, and the argument behind it.

In focus 9 Nov 2026

Five priors picked for this edition. The set changes with each check.

  1. Carry 3x pipeline coverage of quotaUpdate

    Nobody can trace the 3x rule. Plan from the coverage ratio your own history implies.

  2. Pay reps on bookings, with accelerators above quotaUpdate

    Where revenue arrives as usage, a signature no longer fixes it, so bookings comp needs a consumption measure there.

  3. Set quota from measured rep capacity, about 4 to 5x OTEKeep

    Raising quotas for an AI lift nobody has measured turns a guess into a pay cut.

  4. The CRM is the system of record: if it's not in Salesforce, it didn't happenUpdate

    The CRM stays the record of amount, stage and owner. It never held the deal itself.

  5. Price per seatUpdate

    Agents break the link between users and value, so seats alone stop fitting.

How a verdict is reached

This isn't first-party proof. The verdicts rest on public sources, and the value is the method, applied the same way every time:

  • Logic before evidence. Each prior is broken into the premises that made it sensible and the premises the challenge needs. The verdict follows from which premise actually changed.
  • Does it still work, not how many do it. Verdicts rest on outcomes: tests, cohort comparisons, before and after. Adoption counts describe the market, not whether something works.
  • Every side, on purpose. Each prior is searched from the people calling it outdated, the people defending it, and the teams that acted and reversed.
  • Sources screened first. Only sources credible on the subject are cited. When a source sells something related, the entry says so.

Every verdict carries the date it was last checked, and each one is re-checked twice a month. A verdict that changes keeps a dated line saying what moved it.

pipeline and forecast

Carry 3x pipeline coverage of quota

Update itNobody can trace the 3x rule, so plan against the coverage ratio your own history implies, not a fixed multiple.

What decides itAbout a third of qualified pipeline closes, so three times quota covers the number. (not shown yet)

Read the argumentHide the argument for Carry 3x pipeline coverage of quota

Verdict basisNever sourced.

Why the prior made sense

  • unprovenBookings come from day-one pipeline at a steady enough rate that one ratio can stand in for the funnel.
  • unprovenAbout a third of qualified pipeline closes, so three times quota covers the number.
  • changedModeling conversion by stage and cohort was too costly, so a single multiple was the practical proxy.

What the challenge needs

  • failsAn all-opportunity win rate measures how day-one pipeline converts.
  • unprovenPipeline quality holds as volume rises, so more pipeline buys more bookings.
  • failsOne industry multiple fits your motion.

Where it came from

Nobody knows. In 2013 Dave Kellogg called 3x the rule of "always, everywhere, every time" and admitted he didn't know its history. He'd heard "the rule of three" before sales-force automation existed. Later searches for an original study turned up nothing. The number has outlived every attempt to source it.

What people say now

The loudest 2026 claim is that 3x is dead because win rates fell, so 5x is the new floor. The math in most of these posts is the same. They take an industry win rate (often 19%, the 2025 Ebsta and Pavilion average new-logo win rate, counted across all opportunities) and invert it. Steve Litzow, a CRO who cites the 19% without a source, puts it plainly: "The 3× pipeline rule only works at a 33% win rate."

Kellogg reads the same data the other way. He calls it a pipeline crisis rather than a conversion crisis, arguing that teams miss because they start quarters without enough coverage. He defends tracking coverage while leaving the 3x figure aside.

Does the logic hold

3x rested on three ideas. One ratio could stand in for the funnel. About a third of pipeline closes. And modeling conversion properly cost more than it was worth. The second was never measured, as no study shows a third of qualified pipeline closing, or shows that it doesn't. AI changed the third because conversion by stage and cohort is now cheap to compute.

The 5x argument keeps the first idea and swaps the second. It assumes an all-opportunity win rate tells you how day-one pipeline converts. It doesn't: different denominator, different question. It also assumes pipeline quality holds as AI makes pipeline cheap to inflate, which nobody has shown. So the new number fails for the same reason the old one did. A conclusion only follows from the premise that changed, and that premise argues for computing your own ratio, not adopting a bigger one.

What both sides miss

Defenders miss that 3x was never derived, so "it still works" has nothing underneath it. Critics miss that 5x is the same move with a new input.

The sharper version is Kellogg's own, from 2013. Once a CRM could count coverage, managers pushed every rep, region and country until each showed 3x, and soon every company had it. He called the result "a giant self-fulfilling prophecy," since institutionalizing a ratio means you will have that ratio, whether or not it converts. AI makes building pipeline nearly free, which makes the ratio easier to game than to read.

Sources 3

Forecast revenue from bookings

Update itUsage and credit revenue arrives as customers consume, not when the contract is signed, so forecast that share from consumption and keep bookings for the revenue a contract actually fixes.

What decides itA signed contract fixes revenue for its term, so bookings set revenue. (has changed)

Read the argumentHide the argument for Forecast revenue from bookings

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • changedA signed contract fixes revenue for its term, so bookings set revenue.
  • changedConsumption was hard to measure and model, so the contract was the best signal available.
  • holdsBuyers want committed, predictable spend, so most revenue stays contracted.

What the challenge needs

  • holdsUsage and credit revenue follows consumption, whatever was booked.
  • holdsWhere consumption is metered, it forecasts usage revenue better than bookings do.
  • unprovenAI pricing moves enough revenue to consumption that bookings stop mattering for most software companies.

Where it came from

Nobody is credited with it, because for a contracted business it was simply how the money worked. A signed subscription fixes revenue for its term. So the forecast runs from pipeline to bookings to revenue, plans are built on bookings, and reps are paid on the booking. Paying on bookings is a standard comp heuristic, and no primary source for it turned up.

What people say now

This is the one prior finance broadly agrees should change, and the CFOs of usage-priced companies have said so for years. Datadog's CFO: "We continue to believe revenue is a better indication of our business trends than billings or RPO." Snowflake's CFO calls product revenue "the leading indicator of our business." CJ Gustafson, writing in Mostly Metrics, a finance newsletter sponsored by Brex, describes Snowflake's daily consumption forecast standing in for the pipeline-to-bookings chain: "The implication is you can't use bookings to predict revenue."

Finance writers also want the label fixed. OnlyCFO, a pseudonymous newsletter Brex also sponsors, warns of "deception/fraud in 'ARR' numbers," and a finance lead at Notion suggests calling usage run-rate revenue. Other companies go the other way and add annualized usage to ARR.

Does the logic hold

The prior rested on three ideas. A signed contract fixes revenue for its term. Consumption was hard to measure, so the contract was the best signal anyone had. And buyers want committed, predictable spend, so most revenue stays contracted.

For usage revenue, the first idea has changed because the money arrives as customers consume, whatever was booked. The second changed for companies that meter well. The challenge's core premise, that usage revenue follows consumption, holds, and the companies built on it report on that basis.

Where the challenge overreaches is scope. The third premise still holds. Buyers still pay for predictability, and a pricing consultancy, SBI, finds more than 80% of SaaS companies still charge on some seat metric. So the conclusion that follows splits the forecast into one for what a contract fixes and another for what consumption decides.

What both sides miss

None of this started with AI. AI credit pricing spreads a practice that infrastructure companies settled long ago. Comp lags it. ICONIQ, a growth investor, finds about half of consumption companies still set quotas on contract value. Chris Degnan, Snowflake's former CRO, sets the bar: "it's only when you can truly measure consumption that you earn the right to comp on it."

The defend side also has a stronger form than it uses. A company can keep revenue predictable by charging for committed capacity, used or not, rather than for consumption. Predictability can be chosen in pricing as well as won in forecasting.

Sources 5

Forecast by stage probability and commit categories

Too early to callEveryone agrees the stage forecast is weak, but nobody has shown AI forecasting does better, and replacing the commit call drops the accountability it was for.

What decides itA model built on activity and conversation signals forecasts more accurately than stage weights and the rep's call. (not shown yet)

Read the argumentHide the argument for Forecast by stage probability and commit categories

Verdict basisThe evidence splits.

Why the prior made sense

  • changedReps are the only source of deal status, so a stage and the rep's call are the best available read on a deal.
  • changedModeling conversion from evidence costs too much, so a fixed probability per stage stands in for it.
  • holdsA commit is a promise a named person makes, and the weekly call is where that person is held to it.

What the challenge needs

  • unprovenA model built on activity and conversation signals forecasts more accurately than stage weights and the rep's call.
  • failsA forecast is only a prediction, so accuracy is the whole test.
  • unprovenThe rep's call adds nothing a model can't see.

Where it came from

Two habits, rolled into one forecast. The CRM assigns each stage a probability, from 10% at prospecting to 90% at negotiation, and weights pipeline by it. Reps also place deals in pipeline, best case or commit. Dave Kellogg's version of the practice is precise. Commit should be missed about once in two and a half years, and stage probabilities should be derived from your own history. The CRM's defaults were placeholders, not benchmarks.

What people say now

Almost no one defends the forecast's accuracy. Funnelcast, a forecasting vendor, studied 22 company-quarters: only 12% of commit and best-case roll-ups ever landed within 5% of the quarter's wins. It measured category roll-ups instead of stage-weighted ones.

The update side is mostly vendors and advisers. Their pages call the default stage probabilities folklore and promise a calculation in place of the rep's story. In a common middle form, the model sets the baseline, and reps explain every place they disagree with it.

Does the logic hold

The stage forecast rested on three ideas. Reps were the only source of deal status. Modeling conversion from evidence cost too much, so a fixed probability per stage stood in for it. And the commit call held a named person to the number. AI changed the first two. Activity and conversation signals are now captured without the rep, and conversion by stage is cheap to compute from your own history.

That is enough to retire the default probabilities, which Kellogg told teams to replace with their own history long before AI. It isn't enough to retire the commit call. That case needs a model to forecast better than the call, and the support for that comes from vendors. It also needs the forecast to be nothing but a prediction. A commit is a promise as well, one a named person makes, and a model can't be held to one. So the update follows for the stage weights and stops short of the commit.

What both sides miss

The update side often builds its model on the same rep-entered stages it calls unreliable. A model trained on rep-set stages inherits the reps' guesses, and no independent study shows AI forecasting tools raising accuracy across companies.

Defenders miss that a stage forecast built on the default probabilities was never measured against anything. Defending it defends a placeholder.

Sources 3

capacity and comp

Pay reps on bookings, with accelerators above quota

Update itWhere contracts commit spend, signature still fixes revenue and bookings comp holds, but where revenue arrives as usage a signature no longer fixes it, so add a consumption measure there, and price in the mispricing that accelerators buy alongside the extra effort.

What decides itA signed contract is the revenue the company will collect. (has changed)

Read the argumentHide the argument for Pay reps on bookings, with accelerators above quota

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • changedA signed contract is the revenue the company will collect.
  • holdsAccelerators above quota raise output among top performers.
  • unprovenGross margins near 80% make a bookings commission cheap to pay.

What the challenge needs

  • holdsWhere revenue arrives as usage, a bookings plan pays for an event that no longer fixes revenue.
  • unprovenPaying on consumption produces better outcomes than paying on bookings.

Where it came from

Bookings comp is the default SaaS plan rather than anyone's discovery. A rep earns commission on contract value at signature, and accelerators raise the rate once quota is passed. The Bridge Group's AE benchmarks track it: its 2026 summary puts the median AE at $200K on-target earnings against a $960K quota. ICONIQ's 2025 survey of 205 GTM executives finds accelerators in most AE plans.

The logic was tidy. In seat-based SaaS, signature was the moment a year of revenue became committed. Paying on it rewarded the event that fixed revenue.

What people say now

Usage pricing is the challenge. Alexander Group, a sales comp consultancy, puts consumption pricing at about 40% of technology companies in its own data, and for uncommitted or pay-as-you-go contracts it recommends paying on consumed revenue. For committed spend, bookings remain. ICONIQ's 2026 network data shows the shift under way: about half its companies still set quotas on committed ACV, and 30% now set them on forecast consumption, up 10 points in a year.

The practitioner caution comes through ICONIQ too. Chris Degnan, Snowflake's former CRO, describes layering consumption on top of bookings rather than replacing it: "It's only when you can truly measure consumption that you earn the right to comp on it." A comp-software vendor adds that many AI companies run lower gross margins than classic SaaS, which would make the commission harder to afford.

Does the logic hold

The prior rests on two ideas. A signed contract is the revenue the company will collect, and accelerators make top performers sell past quota. The first has changed for one kind of revenue. Where a contract commits spend, signature still fixes it. Where revenue arrives as usage, it doesn't, and a bookings-only plan pays for an event that no longer decides the number.

That supports a narrower conclusion than the challenge usually draws. It says to add a consumption measure where revenue is uncommitted. It doesn't show that paying on consumption works better, because no study compares the two on outcomes.

Accelerators hold. A field study of one firm found overachievement commissions keep top performers selling past quota, and quarterly bonuses pace the laggards.

What both sides miss

Accelerators have a price the defenders leave out. Ian Larkin's 2014 study of one enterprise software vendor found reps timing deals around the accelerating schedule, with the extra discounting costing 6 to 8% of revenue. And in a field redesign at a medical-device firm, removing quotas and ceilings raised revenue 9%.

So the plan's shape matters as much as its basis. A team that moves to consumption comp and keeps a steep accelerator inherits the old timing problem on a new measure.

Sources 6

Reps forget 87% of sales training within 30 days, so reinforce it

Update itThe 87% traces to a 1979 claim about skills, with no timeframe, from one unpublished Xerox study, and forgetting has no fixed rate, so drop the number; the advice to practise and coach after training still holds, and AI practice tools make it cheaper.

What decides itReps lose training at a fixed rate: 87% within 30 days. (doesn't hold)

Read the argumentHide the argument for Reps forget 87% of sales training within 30 days, so reinforce it

Verdict basisNever sourced.

Why the prior made sense

  • holdsSkills taught in a classroom fade unless they are practised and coached afterwards.
  • failsReps lose training at a fixed rate: 87% within 30 days.
  • changedReinforcement has to come from managers, whose time is scarce.

What the challenge needs

  • holdsAI coaching can do part of a manager's reinforcement work.
  • unprovenGains from AI practice last and carry into live deals.

Where it came from

The number has an ancestor, and it says something different. In November 1979 Neil Rackham wrote in the Training and Development Journal that a Xerox study had found that, without follow-up coaching, 87% of the skills change a program produced was lost. He added that knowledge training showed much smaller loss. There was no timeframe. The Xerox study was never published; Rackham cited it with Xerox's permission.

Then the wording drifted. By 2011 The Challenger Sale was quoted as saying Rackham's research showed 87% of training content forgotten within thirty days. A 2013 newspaper column used thirty days; a 2016 vendor post said twelve weeks and credited Xerox. Each retelling changed the timeframe and turned "skills change" into "content."

What people say now

The sharpest critics are the people who traced it. In 2010, Will Thalheimer reviewed 14 studies covering 69 conditions and found forgetting ranging from 0% to 94%. His verdict: "Rules-of-thumb that show people forgetting at some pre-defined rate are just plain false." A 2015 replication of Ebbinghaus's forgetting curve by Murre and Dros reproduced its shape, but with one subject learning nonsense syllables, so its percentages don't transfer to sales skills. Eubrics, itself a vendor, traced the number back and could not find a source for the version in circulation.

Meanwhile AI role-play and coaching vendors use the 87% as the opening line of their pitch.

Does the logic hold

The prior bundles a number and a piece of advice. The advice rests on the causal idea that skills taught in a classroom fade unless they are practised and coached afterwards. That holds. Thalheimer's review shows how much learning method changes forgetting. A small 1997 study, with no control group, found training plus coaching raised productivity far more than training alone.

The number fails. It was never a rate over 30 days, and forgetting has no fixed rate to find.

The scarcity behind the advice has changed. Reinforcement used to mean a manager's time. A randomized study at two fintech firms (Luo and colleagues, Journal of Marketing, 2021) found an AI coach outperformed human managers, most of all for middle-ranked agents, and a combination of AI and human beat either alone. So the update removes the statistic and keeps the reinforcement.

What both sides miss

The same study found bottom-ranked agents overwhelmed by the AI's feedback and top-ranked agents resisting it. AI coaching is not a uniform substitute for a manager. It fits the middle of the team best.

And no one has shown that gains from AI practice last or carry into live deals. Better rehearsal is not yet better selling. Whether it changes what happens on a real call, months later, is the measurement the field still needs, and it is a better question than any forgetting rate.

Sources 6

Reps spend most of their week not selling

Keep itReps still sell for well under half their week, and the claim that AI turns saved hours into selling time is mostly a pitch, because in most teams the hours leak back into admin.

What decides itHours saved are reinvested in selling. (doesn't hold)

Read the argumentHide the argument for Reps spend most of their week not selling

Verdict basisThe claims against it are hype.

Why the prior made sense

  • holdsAdmin, research and data entry take most of a rep's week.
  • changedThat work has to be done by the rep, by hand.
  • failsAn hour taken off that work goes to buyers.

What the challenge needs

  • unprovenAI takes over enough of the non-selling work to triple selling time.
  • failsHours saved are reinvested in selling.
  • unprovenThere are enough ready buyers to fill the extra selling hours.

Where it came from

The number comes from a vendor survey. Salesforce's State of Sales has reported for years that reps spend under a third of their week selling: 28% in the 2022 edition, 30% in 2024. Salesforce sells the automation that is meant to fix it, so the number has always come with an interest attached. It became the baseline for every argument about rep time.

What people say now

The claim against it is Mark Roberge's. On LinkedIn he wrote that every sales team has the opportunity to "double selling time per AE" with AI, and that "doubling selling time should, algebraically, double productivity per rep." In interviews he puts it as selling time rising from 25% to 75% of the week, now one of the most repeated lines in RevOps.

The best data runs the other way. Gartner surveyed 210 sales leaders: AI saves sellers about 4.8 hours a week, but 72% of sales organizations show low reinvestment of that time into high-value work. Trade coverage put it plainly: the hours AI frees up are "quietly leaking back into admin and data entry."

Salesforce's own 2026 edition reports 40% selling time, with prospecting and data entry still counted as non-selling work. That is the largest move the series has shown, and it comes from the vendor whose tools are meant to produce it. Ebsta, which sells revenue intelligence, and Pavilion, a private community for GTM leaders, put the average seller under two hours a day of active selling.

Does the logic hold

The prior is a short chain. Admin, research and data entry fill most of a rep's week. Reps have to do that work themselves. So any hour taken off it was assumed to go to buyers. The first link still holds in every survey, since even Salesforce's best reading leaves most of the week to work that isn't selling. AI changed the second link. Tools now save sellers close to five hours a week.

Roberge's algebra keeps the third link and leans on the second. It assumes the saved hours flow to buyers, and that buyers are waiting for them. Gartner's data breaks the hours assumption in most teams, where the time leaks back into admin. Nothing in the evidence shows the buyers are waiting. The algebra holds everything else constant, including demand, and Dave Kellogg sees a market short of pipeline, not of rep hours. So the conclusion rests on the one premise of the old prior that failed.

What both sides miss

Both sides argue from self-reported time-use surveys, which measure what reps say about their week, not what they did.

Defenders miss the minority it works for. In the same Gartner survey, organizations that do reinvest the saved time are 2.2 times more likely to beat their customer-growth goals. The gain depends on someone redesigning the week, not on the tool.

Sources 5

Set quota from measured rep capacity, about 4 to 5x OTE

Keep itQuota set from measured capacity already absorbs any real AI lift once it shows up in bookings, while raising quotas for a lift nobody has measured turns a guess into a pay cut.

What decides itAI raises output per rep by an amount you can know before it shows up in bookings. (not shown yet)

Read the argumentHide the argument for Set quota from measured rep capacity, about 4 to 5x OTE

Verdict basisStill true.

Why the prior made sense

  • holdsQuota should encode what one rep has shown they can close in a year.
  • holdsCapacity is measured first, and quota follows from it.
  • holdsPay is a share of quota, so a quota set too high is a pay cut.

What the challenge needs

  • unprovenAI raises output per rep by an amount you can know before it shows up in bookings.
  • failsPricing the lift in early costs little if it arrives late.
  • unprovenHigher attainment at AI-engaged companies comes from the AI.

Where it came from

Quota has long been set at about four to five times a rep's on-target earnings. Tomasz Tunguz modelled it that way, and the Bridge Group measures a median of 4.6x in 2026, up from 4.2x in 2024. The ratio encoded what one person could close in a year. Capacity came first, and quota followed from it.

What people say now

The challenge says quotas should rise now to capture the productivity lift AI adds. BCG's 2026 survey of 398 revenue leaders found 91% expect quotas or productivity targets per seller to go up. Only 13% name win rates among AI's top measurable impacts, and only 8% name higher quota attainment. In a survey of 200 incentive-comp professionals by CaptivateIQ, which sells comp software, 43% say their companies already price AI productivity into quotas.

Finance and planning voices defend the old order. The common advice is a base-case plan with no AI lift, and lift added only once it shows up in the numbers. Forma.ai, which sells sales planning software, says AI productivity "should therefore enter the quota model through evidence."

Does the logic hold

The prior is a chain of three ideas. Quota should encode what one rep has shown they can close. So capacity is measured first and quota follows it. And because pay is a share of quota, a quota set too high is a pay cut. None of these depends on how productive reps are, so AI doesn't touch them.

The challenge changes the order. It asks leaders to price the lift in before it is measured, and that needs two things: a lift you can predict, and a low cost of being wrong. The first is unproven; few revenue leaders can yet point to AI in their win rates or attainment. The second fails on the prior's own logic, since a missed assumption lands in every rep's pay. The challenge's conclusion only follows once the lift is measured, and at that point the old method gets there too.

What both sides miss

The two halves of the story are documented, but the link between them isn't. Quotas went up, and Bridge measures attainment falling to 48% of AEs from 51%. No study ties that drop to AI-priced quotas, and Bridge places it on a long-term downward trend.

The defenders' evidence has its own catch. Bridge finds 57% of reps at quota in the most AI-engaged third of companies, against 39% in the least. Bridge labels this observational data, and better teams may simply adopt more tools. One CRO's line covers the risk on both sides: "I don't believe that if I was to double somebody's quota that I give them any chance to be successful."

Sources 4

A new AE ramps in three to six months

Keep itAE ramp still takes quarters, now at the long end of three to six months, and the proof that AI shortens it comes from support desks, not sales floors.

What decides itRamp is paced by the deals a rep has to learn to win. (holds)

Read the argumentHide the argument for A new AE ramps in three to six months

Verdict basisStill true.

Why the prior made sense

  • holdsA new rep learns the product, the buyers and the accounts by doing, and that takes quarters.
  • holdsRamp is paced by the deals a rep has to learn to win.
  • changedWhat a new rep needs to know reaches them slowly, through colleagues and repetition.

What the challenge needs

  • unprovenFaster access to answers and coaching shortens the time to carrying a full number.
  • unprovenThe support-desk result carries over to quota-carrying sales.
  • failsFaster onboarding means faster ramp.

Where it came from

Capacity models need a ramp curve, and three to six months became the default. The Bridge Group has measured AE ramp for years, and models such as Tomasz Tunguz's ramp new reps over four to nine months depending on seniority. "Ramped" never had one definition, as some teams mean the first deal and others mean sustained attainment.

What people say now

Bridge's 2026 report puts AE ramp to full productivity at 6.2 months, the highest in the study's history, so the prior's range now sits at or past its own top end. "Full productivity" is Bridge's label; the page doesn't define the threshold.

The case that AI shortens ramp rests mostly on one strong study from outside sales. Brynjolfsson, Li and Raymond followed more than 5,000 support agents. With AI assistance, agents two months in performed like agents without it at six months or more, and novices gained the most. It is peer-reviewed, and it covers support work outside quota-carrying sales.

Inside sales the evidence is vendor-made. A benchmark co-authored by Fullcast, a planning vendor, and Pavilion, a private community for GTM leaders, reports that "AI-enabled teams ramp 32.7% faster." Mindtickle, an enablement vendor, reports top companies onboarding in 25 days against a 58-day average. Both are correlational.

Does the logic hold

Ramp took quarters for two reasons. A new rep had to learn the product, the buyers and the accounts, and that knowledge reached them slowly, through colleagues and repetition. And the deals set the pace, because a rep can't show a full number faster than they learn to win the deals that make it up.

AI changed the first reason. Answers, coaching and role play now arrive on demand. The challenge assumes this shortens the whole ramp. That holds where the job is mostly knowing the answer, which is what the support-desk study measured. In quota-carrying sales the second reason weighs more, and Bridge ties the record ramp to larger buying committees and more complex deals. The vendor numbers measure onboarding, the part AI sped up, and onboarding is not ramp. So the conclusion overreaches. AI can make a new rep's first months faster without moving the quarter they first carry a full number.

What both sides miss

SDR ramp is moving. Bridge measured it at 3.0 months in 2025, the lowest since 2010, and said that "may reflect AI-powered onboarding." Bridge offers that as a possibility, not a finding.

Defenders read the record ramp as AI failing. Bridge's data doesn't test AI, so it isn't evidence either way. AI tools could be shortening ramp underneath the drift toward bigger deals, and no one has published the cohort data to separate the two.

Sources 5

Carve and balance territories once a year

Keep itReassigning accounts still costs revenue, so a planned major realignment on the comp cycle stays right, with small boundary changes in between, and the case for continuous AI rebalancing is vendor marketing with no outcome data.

What decides itReassigning accounts costs revenue because it breaks customer relationships. (holds)

Read the argumentHide the argument for Carve and balance territories once a year

Verdict basisStill true.

Why the prior made sense

  • changedRedrawing territories takes a lot of analysis.
  • holdsReassigning accounts costs revenue because it breaks customer relationships.
  • holdsComp plans and quotas run annually, so territories change on the same cycle.

What the challenge needs

  • unprovenContinuous AI rebalancing beats an annual redraw on revenue.
  • unprovenTerritories drift out of balance within a year by more than a reassignment costs.

Where it came from

Nobody coined it. Annual territory planning grew up alongside annual comp plans. Alexander Group, a sales consultancy, put it plainly in 2013: territory design and optimization "should be an annual process."

The research behind it comes from ZS. Zoltners and Lorimer found in 2000, across their work with more than 300 sales forces, that 55% of territories were too large or too small, and estimated that good alignment adds 2 to 7% to sales. But their own 2001 review doesn't say once a year. It says alignments change often, with small boundary changes going on all the time and major realignments happening less frequently. The strict version of the prior is a narrowing of its own source.

What people say now

The challenge comes from software vendors, which advertise redrawing territories many times faster than spreadsheets. The broader pitch is continuous, AI-driven rebalancing that moves accounts whenever the data says a territory is out of balance.

No credible independent voice makes the case. That is itself the finding. The AI-territory statistics in circulation mostly cite analyst firms without links.

Does the logic hold

The annual carve rested on three ideas. Redrawing took a lot of analysis. Reassigning an account costs revenue, because it breaks a relationship. And comp plans and quotas run annually, so territories follow the same cycle.

The first has changed. Optimization models have existed since the 1970s, and tools now claim minutes rather than weeks. But analysis was never what made major realignment rare.

The second holds, on causal evidence. After one optimized realignment, Sinha and Zoltners (2001) tracked about 4,500 reassigned accounts against a control. Accounts buying $50,000 to $100,000 a year bought 20% less. Shi and colleagues (2017), studying a Fortune 500 distributor, found a change of rep cut affected customers' annual sales by 13.2 to 17.6%. That study covers departures rather than realignments, but the mechanism is the same.

The challenge needs territories to drift out of balance by more than a reassignment costs, and continuous rebalancing to earn more than annual. Neither has outcome data. Cheaper analysis makes moving accounts easier; it doesn't make it cheaper for the customer.

What both sides miss

The relationship cost depends on how it is managed. In the same 2001 realignment, accounts above $100,000 showed no loss, because the outgoing and incoming reps shared the account and its commissions for a period. The lesson for both sides is about the handoff, not the calendar.

So the practice to keep is the one in the source: a planned major realignment on the comp cycle, small boundary changes when needed, and a shared transition for any account big enough to lose.

Sources 4

More reps means more revenue

Too early to callThe case that AI ends headcount-led growth rests on outliers and predictions, the defence rests on budgets that still buy quota carriers, and neither side has measured what an added rep returns.

What decides itAI raises bookings per rep at typical companies, not only at the fastest-growing ones. (not shown yet)

Read the argumentHide the argument for More reps means more revenue

Verdict basisThe case to update is mostly aspiration.

Why the prior made sense

  • unprovenSeller hours are the binding constraint, so revenue scales with the number of people selling.
  • unprovenThere is enough demand to fill each added rep's hours.

What the challenge needs

  • unprovenAI raises bookings per rep at typical companies, not only at the fastest-growing ones.
  • failsAI-native outliers show the revenue per employee the median company will reach.
  • unprovenOnce output per rep rises, adding sellers stops being the way to add revenue.

Where it came from

Nobody wrote it down as a law. It sits inside every capacity model as ramped reps times quota times attainment. Battery's T2D3 path, built from seven public companies, mapped each ARR stage to a rep count. Mark Roberge, who built one of the best-known capacity models, still says the math "almost never holds," for the pre-AI reasons of ramp, demand and product fit.

What people say now

The update case leans on AI-native companies. Bessemer, a venture firm that invests in AI startups, reports its fastest cohort at about $1.13M of ARR per employee in the first year. Gartner now says sales productivity gains "remain constrained by operating models designed to scale primarily by adding more people."

The defence comes from the people signing the budgets. Salesforce, which sells AI agents, told investors that "supply is capacity" and that it still has 1,200 more account executives to hire. Gartner's CFO survey has total headcount growth falling for 2026, yet sales still gets one of the largest budget increases.

Does the logic hold

The prior rests on two ideas. Seller hours are the scarce input, so each added rep adds roughly a rep's worth of bookings. And there is enough demand to fill those hours. Every capacity model encodes the first and simply assumes the second.

The challenge needs AI to raise output per rep so far that the same team carries more revenue, and its evidence is the AI-native outliers. The outliers are not the median. SVB, a bank whose clients are venture-backed startups, publishes data showing the median AI startup with lower revenue per employee than non-AI peers at every stage from seed to Series D, so the premise that outliers show where everyone is heading fails. The broader premise, that AI lifts bookings per rep at an ordinary company, is still a forecast.

The prior's own first premise is on no firmer ground. The evidence that seller hours still bind is that budgets still buy quota carriers, and that describes what companies do, not what an added rep returns. So neither side has shown its core premise, and the premise that would settle the question, demand, goes unmeasured by both camps.

What both sides miss

Defenders miss what happens when the added heads go into a volume strategy. Pavilion's CEO Sam Jacobs, summarising the benchmark Pavilion co-published with Fullcast, a planning vendor, describes a cohort that added 38% more sellers and pointed AI at top-of-funnel volume. Its revenue per seller fell 26%. That figure is correlational and the dataset is vendor co-authored, but it cuts against both camps.

Neither side has shown the reversal. No named GTM company has publicly undone an AI-justified sales headcount cut. The cleanest reversal on record, Klarna, was in customer support.

Sources 6

Staff one ops person for every 12 to 15 reps

Too early to callThe ratio was always a description of what teams run, never a tested staffing rule, and the claim that GTM engineers shrink it rests on job posts and title counts, not output.

What decides itOps headcount has to grow with the number of sellers. (not shown yet)

Read the argumentHide the argument for Staff one ops person for every 12 to 15 reps

Verdict basisThe case to update is mostly aspiration.

Why the prior made sense

  • unprovenOps headcount has to grow with the number of sellers.
  • holdsMost teams run about one ops person for every 12 reps.

What the challenge needs

  • unprovenGTM engineers do the work of several ops roles, so fewer ops people can support the same sellers.
  • holdsGTM engineer roles are growing fast from a small base.

Where it came from

The ratio started as an anecdotal rule of thumb of one sales ops person for every ten to fifteen salespeople, with no published method behind it.

Adam Schoenfeld's PeerSignal measured it properly in 2024, on org data for 2,500 B2B software companies: about 12 reps per RevOps person overall, with fewer ops people per seller above 1,000 employees and "wide variance." PeerSignal doesn't tie the ratio to results. Like the rule of thumb, it describes what teams run.

What people say now

The challenge comes from the rise of the GTM engineer, a technical role that builds the automation ops people used to do by hand. James McKay, an agency founder writing for Pavilion, argues GTM engineers "are eating RevOps' lunch," citing falling RevOps job ads. Cristian Frunze, publishing his company Kenso's own profile data, counts 121 people with the title at the end of 2023 and 791 by March 2026, while sales and marketing headcount stayed flat.

The scale is still small. Kyle Poyar, using job-post data, found one GTM engineer posting for every 14 RevOps postings and every 35 sales ops postings.

Does the logic hold

The prior rests on the causal idea that ops work scales with the number of sellers, so ops headcount has to grow with them. That was plausible when territories, quotas, comp and reporting were mostly manual. But it was never tested. PeerSignal shows ops and sellers rise together across companies, with wide variance, which is correlation across firms, not a staffing law. The ratio is a fair description of practice and nothing more.

The challenge needs the reverse causal idea that one GTM engineer does the work of several ops roles, so the same sellers need fewer ops people. The evidence for it is hiring data. A title growing fast from a small base tells you what companies are trying while saying nothing about what each hire produces. No study compares output per ops head with and without GTM engineers. So the challenge rests on an unproven premise, and so does the prior it challenges.

What both sides miss

Neither side measures the thing a staffing ratio is for. A useful number would tie ops capacity to something a revenue leader cares about, such as plan changes shipped, data quality, or time to a reliable forecast. Without that, 1:12 and any leaner ratio are both descriptions of what someone chose to hire.

The same assumption that the work follows headcount sits under "more reps, more revenue." When automation changes how much work each seller generates for ops, the ratio stops being a useful shortcut either way, and a team has to count the work itself.

Sources 4

outbound and demand

It takes 8 to 12 touches, and volume is a lever

Update itThe 8-to-12 touch count was never sourced and cheap volume has stopped working, but fewer, more relevant touches still do, so revise the cadence rather than drop outbound.

What decides itMore sends bought more reach at a reply rate you could count on. (doesn't hold)

Read the argumentHide the argument for It takes 8 to 12 touches, and volume is a lever

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • changedEach message cost a person's time, so a long sequence signalled effort.
  • failsMore sends bought more reach at a reply rate you could count on.
  • failsReplies keep arriving deep into a sequence, so eight to twelve steps pay off.

What the challenge needs

  • failsAI-written personalisation lifts replies at volume.
  • unprovenReply rates have collapsed, so outbound no longer works.
  • holdsA short, relevant sequence gets most of the replies a long one would.

Where it came from

Nobody can find the study. "Eight to twelve touches" and "80% of sales need five follow-ups" circulate in sales training, and the closest attribution reaches a research firm only through secondary blogs. The Bridge Group, which sells sales-development consulting, measures an average of 11.3 attempts per SDR cadence. That describes what teams run. It says nothing about what works.

What people say now

Vendor blogs now argue over whether outbound is dead, and the narrower point they keep landing on is that cheap volume stopped working. The strongest update comes from senders' own data, some of it against their interest. Sales.co, which sells outbound, tested AI-written first lines across 1.28 million contacts: AI personalisation replied at 0.53%, and a plain template with no personal line at 1.00%. Hunter, which sells outreach tools, looked at 31 million emails sent in 2025 and found three emails the best sequence length: 6.8% replied, against 3.3% for a single email.

The rules changed too. Since February 2024, Gmail has required senders of more than 5,000 messages a day to authenticate, offer one-click unsubscribe and keep reported spam below 0.3%.

Does the logic hold

The prior made sense when writing cost something. A person wrote each message, so a long sequence signalled effort, and more sends bought more reach at a reply rate you could count on. The touch count was never sourced, but the logic behind volume was sound.

Two of those premises have moved. Free writing removed the effort signal. Gmail's rules mean volume can now cost you placement instead of buying reach. And the step data, all of it from vendors, puts most replies early. In data from Instantly, which sells cold-email sending software, 58% of replies come from the first step. If that holds in your data, steps nine through twelve were never doing much.

The two loud updates draw opposite conclusions from the same change. "Send more with AI" keeps the old volume premise and swaps in a cheaper writer, and the template test above shows what that buys. "Outbound is dead" needs a collapse in reply rates that the data doesn't show. What follows is the narrower claim: fewer, more relevant touches.

What both sides miss

The collapse story is weaker than it sounds. Instantly, whose figure is the one most quoted, reports a 3.43% average reply rate and calls it "stable despite growing volume." Published 2025 reply rates span roughly tenfold depending on the denominator. Nobody has a multi-year series from one source on a fixed definition.

Sources 5

Buyers are 70% through the journey before they talk to sales

Update itBuyers do arrive later and with AI-built shortlists, but the 70% figure stretches a 2012 study past what it measured, so keep the direction and drop the number.

What decides itThe share done before sales can be pinned at about 70%. (doesn't hold)

Read the argumentHide the argument for Buyers are 70% through the journey before they talk to sales

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • changedBuyers had to talk to a rep to learn pricing, fit and how products compared.
  • holdsBuyers do most of their research before they engage sales.
  • failsThe share done before sales can be pinned at about 70%.

What the challenge needs

  • holdsBuyers now research inside AI assistants and arrive with a ranked shortlist.
  • unprovenThe first sales conversation comes later than ever.
  • unprovenThe seller's job shifts to validating what the buyer found.

Where it came from

A real study, retold beyond what it found. In 2012 CEB and Google reported that business buyers were 57% of the way through a purchase before their first serious engagement with sales. The respondents were customers of only 22 large suppliers. The study measured first serious engagement, which differs from first contact, and 70% was the top of its range. The 70% version is the one that spread.

What people say now

The direction has strengthened. Forrester's 2026 buying study has 94% of business buyers using AI somewhere in the purchase, and twice as many as a year earlier name generative AI as more meaningful than any other source. 6sense, which sells intent data, reports that 94% of buying groups rank vendors before first contact, and 77% buy from the vendor they preferred at the end of their selection phase. In its data the buyer's research-to-seller split moved from 70/30 to 60/40, and buyers still averaged 16 interactions with the winning vendor.

The rep's role is shifting rather than vanishing. Gartner found that 69% of buyers prefer to validate AI-generated insights with a sales rep, even as 67% say they'd prefer a rep-free experience.

Does the logic hold

The prior rests on the seller losing an information advantage. Buyers once had to talk to a rep to learn pricing, fit and how products compared, and the prior says they now do most of that on their own first. The 2012 study measured that direction. It did not measure 70%. The average was 57%, and 70% was the top of the range.

AI pushes that premise further. Buyers can now ask an assistant what they used to ask a rep, and most buying groups rank vendors before anyone calls. That strengthens the direction. It doesn't make "later than ever" follow. In 6sense's own data the research-to-seller split moved toward sellers. So the AI-era claim carries forward the prior's direction, which holds, and its number, which never did. What follows is the direction with a new mechanism and no fixed figure.

What both sides miss

The stat is now a tool for selling the replacement. One software CEO's case for killing the MQL opens: "By the time an MQL hits your CRM, the buyer is already 70% through their journey." The number carries more weight in that sentence than the 2012 study can hold.

Defenders who say buyer behaviour hasn't really changed miss the AI data. Updaters miss that Gartner's figures, like 6sense's, are buyers' stated preferences in a survey, and stated preferences don't always match behaviour. The same buyers who say they want no rep still have 16 interactions with the one they pick.

Sources 4

A buying committee is six to ten decision makers

Update itGroups with veto power still decide B2B purchases, but no single size holds across purchases, so size the group to the purchase, count sign-offs and influencers separately, and work for consensus rather than headcount.

What decides itOne typical committee size holds across purchases, so it is worth planning around. (doesn't hold)

Read the argumentHide the argument for A buying committee is six to ten decision makers

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • holdsPurchases are decided by groups with veto power, not by one executive.
  • failsOne typical committee size holds across purchases, so it is worth planning around.
  • unprovenReaching more of the group raises the odds of winning.

What the challenge needs

  • holdsAI-era purchases have larger committees.
  • unprovenCommittees are shrinking as IT consolidates the decision.
  • unprovenAI lowers the cost of each extra thread, so sellers should go wider.

Where it came from

The line most people quote comes from a Gartner sales page, "The B2B buying journey": "The typical buying group for a complex B2B solution involves six to 10 decision makers." The page gives no date, sample or method. An archived capture from April 2020 carries the line. A capture of the same page from May 2024 no longer does.

The nearest sourced ancestor is CEB, which Gartner bought in 2017. In HBR in 2015, Schmidt, Adamson and Bird drew on four CEB surveys of more than 5,000 stakeholders. On average, 5.4 people formally signed off on each purchase, and the likelihood of purchase fell from 81% with one decision maker to 31% with six.

What people say now

Current counts disagree in both directions. Forrester's 2026 survey of about 18,000 buyers finds generative AI purchases involve 14 members against 7 for other purchases. G2, surveying its own software buyers, says committees of five to eight are giving way to three or four. Gartner itself now says five to 16 people across as many as four functions.

Multi-threading is the practice built on the number. If the group is six to ten, reach more of them. The data behind it comes from vendors' own customers. Gong reports won deals have about twice the buyer contacts. Airspeed, across 133,068 closed deals at 116 companies, finds win rates rising from 24.3% with one contact to 46.4% with five or more, and the pattern survives when it holds the number of calls constant.

Does the logic hold

The prior rests on the two ideas that groups with veto power decide purchases and that one typical size holds across them. Multi-threading adds a third idea, that reaching more of the group wins more often.

The first holds. Gartner's 2025 survey of 632 buyers found 74% of buyer teams in unhealthy conflict, and groups that reached consensus were 2.5 times as likely to report a high-quality deal. The second fails. Academic work has tied size to the purchase for decades. A 2005 meta-analysis found the buying center grows with the importance, complexity and uncertainty of the purchase, and a 1992 study put the mean near four. Today's counts disagree even within one firm, and they count different things, including sign-offs, decision makers and influencers.

The challenges don't repair this. "Bigger now" holds for AI purchases, on one analyst survey. "Smaller now" rests on a vendor survey with no definition of a member. Both replace one number with another. What follows is to size the group to the purchase and count who signs separately from who influences.

What both sides miss

Both sides argue about headcount, while Gartner, the firm whose page made six to ten famous, now says agreement is what predicts the deal. In the same survey, content tailored to the individual had a 59% negative effect on group consensus, and content relevant to the group had a 20% positive one. That cuts against the easy AI version of multi-threading, where cheap threads mean you should go wider. No outcome study tests it.

The multi-threading data is association. Healthy deals may pull in colleagues rather than the reverse, and Ebsta's vendor data finds win rates rising with relationships only up to a point, then falling sharply past 16. "Reach the group" holds up. "More threads cause wins" is unproven.

Sources 6

Gate your best content to generate leads

Update itThe gate trades reach for names, and the names are mostly weak leads; now the gate also hides the asset from the answer engines where buyers start their research, so ungate the summary and gate only the deep asset, if anything.

What decides itPeople who fill in the form are leads worth a sales follow-up. (doesn't hold)

Read the argumentHide the argument for Gate your best content to generate leads

Verdict basisEvidence against it.

Why the prior made sense

  • changedA form is the only way to know who is engaging with your content.
  • failsPeople who fill in the form are leads worth a sales follow-up.
  • unprovenThe reach lost to the gate costs less than the leads it captures.

What the challenge needs

  • holdsAnswer engines cannot read gated content, so a gated asset is invisible where buyers now research.
  • holdsBuyers increasingly start their research in AI assistants.
  • unprovenBeing cited in AI answers produces pipeline.

Where it came from

Gating whitepapers for names predates the phrase content marketing, and no one coined it. Its canonical public statement is a 2011 debate between David Meerman Scott, a marketing author who has advised HubSpot since 2007, and Mike Volpe, then VP of marketing at HubSpot. Volpe argued that marketers have lead goals, so "a business has to use forms to generate leads."

Scott answered that ungated content gets 20 to 50 times the downloads. The only data he published behind it is one 2009 example from AIIM: 5,138 ungated ebook downloads in a month against 513 for gated research, about 16 times. The wider range is his claim, not a finding.

What people say now

Forrester has argued against blanket gating since 2016. It called gating a dating game that pushes buyers to submit false information and yields "too many useless leads." Its 2019 advice was to gate only high-value assets such as analyst reports, never case studies.

The newer argument is mechanical. Answer engines don't fill in forms, so a gated asset can't be read or cited where buyers now research. A Forrester analyst put it this way in 2025: content that isn't crawlable is "not in the conversation."

The defense is mostly adoption. PathFactory, a vendor, found 83% of marketers gate and 65% say gating brings in a significant share of their leads. That describes how common the practice is without showing whether it pays.

Does the logic hold

The prior rests on three ideas. A form was the only way to learn who was reading. The people who fill it in are leads worth a sales call. And the reach the gate costs is worth less than the names it captures.

The first has changed because account identification and intent tools now see engagement without a form. The second fails. Forrester found gates yield false details, and HockeyStack, a vendor comparing 87 companies on its own data, found content leads convert to sales-qualified at under a quarter of the rate of hand-raises, a gap the entry on the MQL covers. The third is barely tested.

The answer-engine challenge holds where it is mechanical, since a crawler can't read behind a form and Forrester's 2026 survey finds 94% of buyers using AI in buying. Whether being cited produces pipeline is unproven. So the update rests on the failed second premise, which predates AI. AI raises the price of the gate. It doesn't decide the case.

What both sides miss

The strongest ungating story is one company, told by its CMO. Chuck Moxley's team at Blue Triangle, a digital-experience software company, spent $15,000 on a gated campaign that drew 3,400 visits and 106 forms; ungated, the same ebook reached 175 engaged readers on $2,000. Half the gated emails were bad. A year after ungating everything, demo requests were up 265% and pipeline up 242%. But an ABM program and new measurement started at the same time, so it isn't a test. No controlled comparison of gated and ungated content on pipeline exists.

Defenders miss that the form always traded reach for names, and the names were weak. What survives is a narrow rule to ungate the summary and the main findings while gating the deep asset. Forrester lands there, and so does Carolyn Shelby, an SEO consultant at Yoast, which sells SEO software.

Sources 6

Lead with a free tier or trial, and bring in sales later

Update itThe product-first, sales-later sequence holds and AI may strengthen it, but inference cost ended the near-free free user, so free-first becomes a capped trial or credits first.

What decides itServing a free user costs almost nothing. (has changed)

Read the argumentHide the argument for Lead with a free tier or trial, and bring in sales later

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • changedServing a free user costs almost nothing.
  • holdsUsers can reach value on their own, before talking to anyone.
  • holdsSelf-serve usage later feeds a sales-assisted expansion.

What the challenge needs

  • holdsInference cost makes an indefinite, full-featured free tier uneconomic, so free becomes time-boxed, credit-capped or a reverse trial.
  • unprovenA capped or time-boxed free tier converts and retains as well as an indefinite one.
  • unprovenCheap self-serve AI products churn too fast for free-first to pay off.

Where it came from

Blake Bartlett says his firm, OpenView, a VC, coined "product-led growth" in 2016. The product does the acquiring through a free tier or a trial, self-service to value, and sales later. Bartlett is careful that this isn't anti-sales. "Sales still very much exists in PLG, but it's optional, not foundationally necessary." Self-service has a ceiling, and sales steps in past it.

The prior rationed two things. Sales headcount, because the product carried acquisition. And the marginal cost of software, which was close enough to zero that giving the product away was a cheap channel.

What people say now

The update comes from people running AI products. Wes Bush, who founded ProductLed and sells PLG advisory, puts it bluntly: "AI products have real COGS," and "a bad free model can burn hundreds of thousands in COGS before you realize it's broken." His conclusion is narrow: "Free isn't dead. But indefinite access with full features? That's gone." Elena Verna, who runs growth at Lovable and wrote in a post sponsored by Stripe Billing, describes a free tier of five credits a day, 30 a month at most.

The defense is that the motion is stronger than ever. Verna herself says "AI is collapsing time-to-value down to first use." Menlo Ventures, a VC, surveyed about 500 enterprise buyers and puts product-led purchases at 27% of AI application spend, against 7% in traditional software, and notes Cursor reached $200 million in revenue before hiring enterprise sellers.

Does the logic hold

The prior rests on three ideas. Serving a free user costs almost nothing. Users can reach value on their own. And self-serve usage later feeds a sales-assisted expansion.

AI changed the first. Every use of an AI product runs inference, so a free user carries real cost; the price-per-seat entry in this series makes the same point about paid users. The second holds and may be stronger, since time to value is shrinking. The third holds on case evidence. Products that grew through self-serve built sales once they reached large accounts.

So the challenge's conclusion follows, but only for the free tier. An indefinite, full-featured free tier gets expensive, and teams move to time-boxed access, credit caps or reverse trials. The sequence, product first and sales later, is untouched. What the challenge doesn't have is evidence that the capped version works as well. No study compares credit-capped and indefinite free tiers on conversion or margin for AI products.

Menlo's share figure is adoption. It shows buyers taking the product-led path, though it leaves unproven whether any particular free tier pays.

What both sides miss

Retention, and it cuts both ways. ChartMogul, a billing vendor, with Kyle Poyar, found AI-native gross retention of 23% below $50 a month and 70% above $250, in buckets of about 50 companies it calls directional. Cheap, self-serve AI products lose most of their customers. Against that, a16z, a VC, argues from its portfolio that products with generous free tiers see converted customers retain better at month three, without publishing figures.

Nobody has outcome data on companies that cut their free tier for inference cost. The acv-sets-the-motion entry covers the other half, noting that if AI makes a sales touch cheaper while free users get costlier, both push sales assistance down-market.

Sources 6

The MQL is how marketing hands demand to sales

Keep itThe hand-off from marketing to sales still works, and hand-raises convert far better than content leads, but "the MQL is dead" mostly attacks the score, and whether scores have lost their predictive power is too early to call.

What decides itA hand-raise, such as a demo request, means a buyer wants to talk. (holds)

Read the argumentHide the argument for The MQL is how marketing hands demand to sales

Verdict basisThe claims against it are hype.

Why the prior made sense

  • holdsMarketing and sales need a defined hand-off, with a threshold and a follow-up agreement.
  • holdsA hand-raise, such as a demo request, means a buyer wants to talk.
  • unprovenScored page views and downloads show who else is ready to buy.
  • failsOne person filling in a form stands in for the buying company.

What the challenge needs

  • unprovenBuyers research where scoring can't see, so behaviour scores have lost their predictive power.
  • failsSo the hand-off itself should be replaced.
  • failsAI is what made the lead the wrong unit.

Where it came from

The SiriusDecisions Demand Waterfall gave marketing and sales a shared ladder of inquiry, marketing-qualified lead, sales-accepted and sales-qualified. It came with a scoring threshold and a service-level agreement for follow-up. In 2017 the same firm replaced the lead with the demand unit, the buying group, because one person filling in a form had stopped describing how companies buy.

What people say now

"The MQL is dead" is one of the most repeated lines in marketing ops. The usual case: buyers research on their own, increasingly inside AI assistants, and arrive with a shortlist, so a score built on page views and downloads measures nothing.

The pushback comes from the people who run the hand-off, and it circulates in marketing-ops forums: if you have a demo form on your site, you're still using MQLs, and the problem is how most teams implement the MQL, not the acronym.

Does the logic hold

The MQL bundles four ideas. Marketing and sales need a defined hand-off. A hand-raise means someone wants to talk. Scored behaviour, page views and downloads, shows who else is ready. And one person stands in for the company buying. The fourth failed before AI did anything, when the canon itself moved to the buying group, so that replacement is older than the argument now made for it.

The second idea now has outcome support. HockeyStack, which sells attribution, compared 87 B2B software companies over late 2023 and early 2024. Companies that counted only demo, pricing and contact requests as MQLs moved 21.55% of them to SQL. Companies that also counted ebook and webinar leads moved 4.93%. The hand-raise pools closed at a higher rate too, at about half the cost per SQL.

The third idea is where the challenge aims, and the case is weaker than it sounds. The claim is that buyers now research inside assistants the score never sees, so scores have lost their power. Nobody has measured that loss, and there is no trend in score-based conversion since 2023. The HockeyStack gap is a level at one point in time, not a decline. And model-based scoring still works in recent academic tests. Across three B2B service industries, a scoring model lifted conversion well above an intuition baseline. Retiring the score is too early to call.

What doesn't follow at all is the step from a weaker score to a dead hand-off. A demo request means what it always meant. A hand-raise and a score crossing are different events that share one label, and most of the case against the MQL is a case against the score.

What both sides miss

Updaters miss who is making the case. Kerry Cunningham, one of the two people who unveiled the buying-group model in 2017, now writes the sharpest "MQL is dead" pieces from 6sense, a vendor that sells the replacement. The strongest supporting stat usually cited, that buyers are 70% through their journey before contacting sales, comes from a 2012 study of 22 suppliers' customers.

The defenders miss something too. "It's the implementation" is a fair defence only if someone actually fixes the implementation.

Sources 4

Respond to inbound leads within the hour

Keep itOutcome evidence still shows conversion falling fastest in the first hour after an inquiry, so reply within the hour and never the next day, but drop the five-minute rule and its 100x and 21x figures, which stop at qualification.

What decides itA buyer's attention to an inquiry fades soon after it, so an early reply converts better than a late one. (holds)

Read the argumentHide the argument for Respond to inbound leads within the hour

Verdict basisStill true.

Why the prior made sense

  • holdsA buyer's attention to an inquiry fades soon after it, so an early reply converts better than a late one.
  • changedReplying fast is costly, because it needs a person watching the queue.
  • failsThe lift is sized in multiples, 100x to contact and 21x to qualify, and it carries through to revenue.

What the challenge needs

  • holdsAI responders make an instant reply cheap for any team.
  • unprovenOnce everyone is fast, speed stops mattering and the relevance of the reply decides.
  • unprovenAn instant automated reply captures the same lift as a fast human reply.

Where it came from

The rule has two versions. In 2011 Harvard Business Review published "The Short Life of Online Sales Leads," built on 1.25 million leads at 42 firms: firms that tried to reach a lead within an hour were nearly seven times as likely to qualify it as firms that waited an hour longer. A 2007 paper by the same authors, on InsideSales.com's own dialer data from six companies, had sharpened the window to five minutes, with contact about 100 times and qualification about 21 times as likely as a call at thirty. David Elkington, a co-author of both, ran InsideSales.com, which sold the dialer. Both studies stopped at qualification.

What people say now

The challenge says AI answers every form in seconds, so speed has stopped being an edge. Sales trainers now repeat on LinkedIn that once everyone can be fast, fast is no advantage.

The sharper challenge comes from 6sense, which sells intent and ABM software and gains when attention moves away from form-fill speed. It argues the speed data shows selection, not persuasion: leads reached fast differ from leads reached slowly, and in its 2026 buyer study the eventual winner was on the day-one shortlist 92% of the time. Even 6sense concedes that a buyer who asks to talk should hear back promptly.

Does the logic hold

The prior makes the causal claim that a buyer's attention fades soon after an inquiry, so an early reply converts better. Two ideas came with it. Replying fast was costly, and the lift was large, measured in multiples.

The causal claim now has outcome evidence that answers the selection point. Musalem and colleagues (Operations Research, 2023) used eight months of quote data from an auto insurer, with a method built to separate the effect of waiting from the kind of lead that waits. Purchase conversion fell 31% when the first call moved from 5 to 30 minutes, a real effect in the tens of percent rather than 100x or 21x. The drop was steepest in the first 10 to 20 minutes and flattened after about 70. That is consumer insurance, but the vendor data that includes B2B agrees on the shape. VanillaSoft, which sells sales-engagement software, found in its 2018 data that wins were far more likely within the hour than after it. Velocify, which sold lead-distribution and dialer software, reported the advantage "greatly reduced after the first hour."

AI changed the cost of speed while leaving its effectiveness intact, so the challenge's conclusion doesn't follow. If a reply within the hour converts better, a team keeps paying for it whatever its rivals do. Table stakes is a reason to keep a practice, not to drop it.

What both sides miss

Defenders keep the wrong numbers, figures that stop at qualification. Two vendor datasets even find the very fastest band is not the best, as VanillaSoft's wins peaked in the 10-to-60-minute window. "Within the hour, never the next day" is well supported. "Five minutes beats fifteen" is not.

Critics assume an instant AI reply carries the same lift as a fast human one. No study compares them on inbound. Field experiments in consumer and small-business lending cut both ways. An AI first call raised loan applications when a human followed up and failed otherwise, and disclosing a bot before the call cut purchases sharply.

Sources 5

Split the work: SDRs book meetings, AEs close

Too early to callNobody has measured whether SDR-booked meetings convert better or more cheaply than AE-sourced or AI-booked ones, so both sides argue the split's productivity case without data, while its other job, training future AEs, has real outcome support.

What decides itA junior seller can book meetings well enough to free AEs to close. (not shown yet)

Read the argumentHide the argument for Split the work: SDRs book meetings, AEs close

Verdict basisThe evidence splits.

Why the prior made sense

  • changedResearching and writing each touch took so many hours that prospecting paid to hand to a cheaper role.
  • unprovenA junior seller can book meetings well enough to free AEs to close.
  • holdsThe SDR seat is where future AEs learn to sell.

What the challenge needs

  • unprovenAgents can do the whole first touch with no human in the loop.
  • unprovenThe fall in SDR hiring is caused by AI.
  • failsAEs can take pipeline straight from agents and lose nothing the seat produced.

Where it came from

The split is about fifteen years old. Aaron Ross's Predictable Revenue (2011) separated prospecting from closing, so a junior SDR booked meetings and an AE closed them. It became standard in software sales, and the Bridge Group has tracked it since.

What people say now

The case against it got loud in 2025, under the slogan that the SDR role is dead. Salesloft's CRO said he wasn't sure the BDR role would "exist post 2026." AI-SDR vendors went further and sold replacement outright.

Some of those vendors have since retreated. 11x's CEO: "I'm the CEO of the company that coined AI SDRs. And I'm here to tell you AI SDRs don't work." His point is that the replacement framing fails, not the software. Artisan retired its "stop hiring humans" slogan in September 2026 and still sells an AI BDR.

The numbers both sides trade describe the market without measuring the outcome. Bridge's 2025 report has the SDR-to-AE ratio "consistent since 2018" and AI SDRs at 1% of companies surveyed. Kyle Poyar's job-post data has SDR postings down 21%. None of it says whether the split produces better pipeline.

Does the logic hold

The split rests on three ideas. Prospecting took so many hours, mostly in research and writing, that it paid to give it to a cheaper role. A junior seller could book meetings well enough to free AEs to close. And the seat trained the next AEs.

AI changed the first because research and writing for a touch now cost almost nothing. That changes the work inside the seat, but the split's productivity claim lives in the second idea, and nobody has tested it. No public dataset compares the win rate or cost per won deal of SDR-booked meetings with AE-sourced ones. The figures comparing AI-booked with human-booked meetings all trace to vendors or unnamed tests. The challenge's own premise, that agents can run the whole first touch, is just as unmeasured. Both sides lean on the same missing number.

The third idea is different, because it has outcome support from other fields. Among more than 100,000 commissioned salespeople at one national retailer, internal hires sold more than external hires, mostly because the best insiders stayed (Rissing and Benson, Organization Science, 2021). In investment banking, people promoted into a job outperformed external hires for about two years and were paid about 18% less (Bidwell, 2011). The only SDR-specific data, a small Bridge Group sample from 2017, found 26% of internally promoted SDRs failed as AEs, against 41% of SDRs who took an AE seat elsewhere, and failure fell to 6% for SDRs with 16 months or more in the seat. A plan that removes the seat has to replace that bench.

What both sides miss

Updaters credit AI with a decline that Salesloft's own CRO blamed on the end of cheap money. If the cause is the cost of capital, a shrinking SDR headcount says little about what agents can do.

Defenders miss that the job inside the title is changing. Where AI has worked, it has been on inbound qualification, or as research feeding human reps.

Sources 6

Multi-touch attribution can apportion credit across touches

Too early to callThe case that attribution is dead rests on consumer-media evidence nobody has tested on B2B opportunity influence, and buyers shortlisting inside AI chats opens a gap that neither side can measure yet.

What decides itConsumer-media signal loss and holdout results carry over to B2B opportunity influence. (not shown yet)

Read the argumentHide the argument for Multi-touch attribution can apportion credit across touches

Verdict basisThe evidence splits.

Why the prior made sense

  • changedThe touches that shape a deal can be tracked and tied to a person and an account.
  • unprovenA fixed weighting by touch position approximates what each touch contributed.
  • holdsMarketing needs a per-channel read of what drove pipeline to defend and allocate budget.

What the challenge needs

  • holdsCorrelated credit overstates touches that would have happened anyway.
  • unprovenConsumer-media signal loss and holdout results carry over to B2B opportunity influence.
  • unprovenIncrementality tests and mix models can replace attribution for B2B budget decisions.

Where it came from

Bizible, now Adobe Marketo Measure, gave B2B marketing its standard models. U-shaped splits credit between the first touch and lead creation. W-shaped gives 30% each to first touch, lead creation and opportunity creation, and spreads the last 10% over everything else. The weights are conventions, not estimates, and the sources don't agree on what W-shaped even means. The IAB's 2017 primer noted that last-touch still dominated despite its known limits.

What people say now

Trade-press columns now call multi-touch attribution (MTA) a waste of time and money. Measured, which sells incrementality testing, goes further: "a more sophisticated version of a flawed idea is still a flawed idea." Its own geo holdouts found lower-funnel remarketing over-credited by about 2.6 times.

The defense comes from attribution vendors. CaliberMind: "there's no alternative." Between the two sits a softer consensus around triangulation across attribution, mix modeling and experiments, which commentators treat as settled.

Does the logic hold

Attribution rested on three ideas. The touches that shape a deal can be tracked and tied to a person and an account. A fixed weighting approximates what each touch contributed. And marketing needs a per-channel read of what drove pipeline. The second was never measured; it was a convention from the start. The third still holds, which is why teams keep running models they don't fully trust.

Two things moved the first idea. Tracking lost signal, and buyers moved their research into AI tools. Forrester finds 94% of business buyers now use AI in the buying process.

The challenge is right that correlated credit flatters touches that would have happened anyway. But its evidence comes from consumer media, with cookies, iOS and retail holdouts, and its conclusion is applied to B2B opportunity influence. That transfer hasn't been tested. Nor has the claim that mix models and holdouts can carry B2B budget decisions. The premise that decides the question is unproven.

What both sides miss

Both sides argue about how to split credit among touches you can see. The newer problem is the touch you can't. If a buyer builds the shortlist in a chat with an AI assistant, the first recorded touch may come from someone who has already narrowed the field. Position-based models credit whatever came next. Holdouts and mix models measure channels you spend on, and it isn't clear what a holdout of an AI chat would even look like. Nobody has a measurement for this yet, including the camp declaring MTA dead.

Sources 3

Third-party intent shows who is in market

Too early to callThe critique that a surge lands after the shortlist is set has an analyst behind it, but the newer claim that AI research starves intent data has no measurement yet, and the defenses argue for using intent differently rather than showing that it predicts.

What decides itResearch inside AI assistants leaves no trail a co-op can see, so intent coverage falls. (not shown yet)

Read the argumentHide the argument for Third-party intent shows who is in market

Verdict basisThe evidence splits.

Why the prior made sense

  • unprovenResearch before a purchase happens on sites a publisher co-op can observe and tie to an account.
  • unprovenA rise in that research marks an account that is in market.
  • unprovenThe surge comes early enough for a seller to get on the shortlist.

What the challenge needs

  • holdsA growing share of buyer research happens inside AI assistants.
  • unprovenResearch inside AI assistants leaves no trail a co-op can see, so intent coverage falls.
  • unprovenSo the surge now arrives after the shortlist is set.

Where it came from

Third-party intent rests on a co-op. Bombora, which sells the data, collects research activity through a tag on the sites in its co-op and flags a surge when an account's reading on a topic over the last three weeks rises above its 12-week baseline. 6sense, which sells an intent platform, scores accounts into buying stages that predict an opportunity opening within 90 days. These are descriptions of what the products do. No independent validation of the claim that a surge predicts a purchase is on record, and the vendors' own figures drift from one document to the next.

What people say now

The sharpest critique is about timing. Forrester, which sells research rather than intent data, warns that late-stage signals "may represent prospects who have already narrowed their options to a shortlist of vendors." 6sense's own 2026 buyer research points the same way, against its own product: 92% of winning vendors were on the initial shortlist, and buyers spend about 60% of a 10.3-month cycle in selection before contacting a seller.

The newer critique adds AI, noting that buyers research in assistants that emit no intent signal. Surveys agree research is moving there and disagree on how far. 6sense finds only 26% use language models to discover vendors; a G2 survey finds 51% of software buyers now start research in a chatbot more often than in Google. No one has published a measurement of intent coverage, match rates or surge frequency before and after that shift. The claim describes a mechanism without offering a measurement.

The defenses are softer. Forrester argues intent has uses beyond spotting in-market accounts, and Gartner tells demand leaders to triangulate across several intent sources. A Forrester report from April 2026, read only in summary, says co-op and bidstream providers have "shown resilience in maintaining predictive signals," with the walled-garden sources most at risk. None shows the signal predicts.

Does the logic hold

The prior rests on three ideas. Research before a purchase happens where a co-op can see it. A rise in that research marks an account in market. And the rise comes early enough to act on, before the shortlist is set. The vendors make the case for all three; nobody independent has tested them.

The challenge says AI changed the first idea. That needs research to have moved into assistants, which surveys support, and the move to have cut what co-ops see, which nobody has measured. So the premise that decides the AI claim is unproven.

The timing critique is stronger, and it doesn't need AI. If the shortlist forms before a buyer talks to anyone, a surge can show who is buying and still arrive too late to change whom they buy from. The evidence leans that way, but nobody has measured when surges land against when shortlists form.

What both sides miss

The prior bundles two claims, that this account is in market and that you've learned it in time to matter. The timing evidence touches only the second. Defenders answer by retreating to other uses, which concedes the second without defending the first. The question nobody has data on is whether a surge, however late, still sorts accounts better than not having one.

Sources 6

Tier ABM accounts one-to-one, one-to-few and one-to-many

Too early to callAI may have cut the cost of account research, but no one has shown AI-scaled programs getting one-to-one results, and human attention still has to be rationed, so the open question is the unit and cadence of ranking, not whether to rank.

What decides itAI-made account research and content give many accounts results close to a one-to-one program. (not shown yet)

Read the argumentHide the argument for Tier ABM accounts one-to-one, one-to-few and one-to-many

Verdict basisThe case to update is mostly aspiration.

Why the prior made sense

  • changedAccount-specific research and content are expensive, so depth must be rationed to a few accounts.
  • holdsHuman attention from marketers, sellers and executives is scarce and should go where value is highest.
  • unprovenDeeper tiers return more per account.

What the challenge needs

  • unprovenAI-made account research and content give many accounts results close to a one-to-one program.
  • unprovenThe opportunity and its buying group, re-ranked as signals change, is a better unit than an annual account list.
  • holdsTeams already run several tiers at once.

Where it came from

ITSMA, a firm that sells ABM research and training, coined account-based marketing in 2003, and Bev Burgess, then at ITSMA, codified it. The three tiers came later, with Strategic ABM for one account at a time, ABM Lite for small clusters, and Programmatic ABM for many. They were documented in ITSMA's 2016 benchmarking survey, which is gated. The earliest public trace is a 2017 study page from ITSMA and the ABM Leadership Alliance.

The tiers rationed marketer hours. That 2016 survey found a median of four strategic accounts per dedicated marketer, and nine for ABM Lite. The originator has since moved on. Burgess's 2025 handbook sets out five types, Strategic, Scenario, Segment, Programmatic and Pursuit, run as a portfolio.

What people say now

The sharpest challenge comes from the analyst who defended the top tier. In 2024 Nora Conklin of Forrester called strategic one-to-one ABM "successful and durable" and said scale "is not what will power profitable growth." In 2026 she wrote that "static segmentation was a solution to scarcity," and that insights which "once relied on manual research are now available at scale."

Vendors push further, saying AI makes one-to-one research cheap enough to collapse the tiers. Adoption is moving toward mixing them. Momentum ITSMA's 2024 benchmark, sponsored by the vendor Demandbase, finds 61% of teams running two or more types, up from 35% in 2017. That says what teams do without showing what works.

Does the logic hold

The tiers rest on three ideas. Account research and content are expensive, so depth has to be rationed. Human attention is scarce and should go where value is highest. And deeper tiers return more.

AI changed the first in cost, though on an analyst's assertion rather than published data, and the surveys say AI in ABM is early. Average maturity is 2.3 out of 5, and nearly 70% of respondents to a 2025 Demand Gen Report survey find its current effectiveness limited. The second holds, and Conklin keeps it: "not every potential deal deserves equal attention." The third was never shown. ABM's returns are self-reported, and no study compares tiers.

The challenge needs AI-made research and content at scale to approach one-to-one results. Nobody has measured that, so the case is aspiration. Read closely, even Conklin's 2026 piece keeps prioritization. It changes the unit, from the account to the opportunity and its buying group, and the cadence, from an annual list to segments that move.

What both sides miss

The challenge assumes personalization is the lever. Gartner's 2025 buyer survey points the other way. Content tailored to the individual had a 59% negative effect on buying-group consensus, while content relevant to the group helped. Cheap tailored content is not the same as a one-to-one result.

Defenders assume the top tier pays. The one academic look at a close cousin, key account management in one firm, found key accounts retained better but less satisfied and less profitable by year three. Forrester's Nicky Briggs already cautioned against one-to-one for new logos in 2023.

The change that holds up today isn't about AI. Forrester and a 2026 study in Industrial Marketing Management both point to prioritization that follows the opportunity's complexity and value over a fixed annual list.

Sources 6

Cut events first: they are expensive and the hardest to attribute

Too early to call"hard to attribute" never meant low return, and in-person meetings show causal value where they have been measured, but the claim that AI makes events worth more than before is a vendor-led pitch with no outcome data yet.

What decides itAI-generated outreach lowers the value of digital touches relative to in-person ones, so events now return more than before. (not shown yet)

Read the argumentHide the argument for Cut events first: they are expensive and the hardest to attribute

Verdict basisThe case to update is mostly aspiration.

Why the prior made sense

  • holdsEvents are expensive per contact.
  • holdsEvent impact is hard to measure.
  • failsWhat can't be attributed returns little, so the hardest-to-attribute line goes first.

What the challenge needs

  • holdsBuyers lean on trusted human sources to check what AI tells them.
  • holdsIn-person meetings do causal work that a digital touch doesn't replace.
  • unprovenAI-generated outreach lowers the value of digital touches relative to in-person ones, so events now return more than before.

Where it came from

No one coined it. "Cut events first" is budget practice. Events are the largest offline line item and the hardest to tie to revenue, so when the budget shrinks they go first. The accountability problem is old. In 1995 the Journal of Marketing asked "Do Trade Shows Pay Off?" and found show spend justified by rules of thumb, with accountability the top challenge. The same paper measured one firm at one show under controlled conditions and projected a 112% long-term return.

What people say now

CMOs are acting on the prior. Gartner's 2026 survey of 401 CMOs finds them "rapidly shifting budget from offline to digital channels," with AI a driver because it favours channels "that can be effectively AI-optimized." Forrester's 2026 survey of more than 400 event decision-makers finds nearly 70% running fewer events. Small formats under 200 people are growing, and only 18% plan more large ones.

The counter-claim is that events gain value as AI floods digital channels. Its canonical form is commercial. Cvent, an events vendor, built its 2026 "Presence Premium" message around it, backed by a vendor survey in which 85% say AI content makes trust harder to build.

Does the logic hold

The prior rests on three ideas. Events are expensive per contact. Their impact is hard to measure. And what can't be attributed returns little, so the hardest line to attribute goes first. The first two hold. Forrester finds over half of teams still use net-new leads as the main event objective, which it says "drives overly simplistic measurement."

The third is a jump with no finding behind it. Hard to measure says nothing about return, and where return has been measured, meeting in person does work. When SARS kept foreign buyers away from the 2003 Canton Fair, exhibitors reached 5 to 7% fewer export destinations, and the study traces the effect to the cost of finding and matching partners. That is causal evidence, though it comes from pre-AI export trade fairs outside B2B software.

The challenge needs more than that. Buyers do lean on trusted human sources to check what AI tells them, Forrester finds. But the step to "events now return more than before" has no before-and-after or cohort data behind it. Its support is opinion, including a Forrester Consulting study that Cvent commissioned. Both conclusions rest on returns nobody has measured.

What both sides miss

Gartner holds both positions at once. It reports AI pushing CMOs away from events toward channels AI can optimize, and in the same release warns them not to let AI steer budget toward "the channels and stages of the journey that are easiest to tune." Its most AI-mature organizations put a lower share of budget into digital.

Neither side is measuring the event well. Forrester finds teams whose event platform is deeply tied into their stack more satisfied, 62% against 37%, and finds only 38% of CMOs agreeing that events perform well, against about half of organizations overall. The fight over whether to cut is mostly a fight over attribution, which has its own entry.

Sources 5

Partner-sourced pipeline beats marketing-sourced pipeline

Too early to callEvery win-rate figure for partner deals is vendor-reported and exposed to selection, the 28 buying moments have no published method, and the claim that AI agents bypass partners is one analyst prediction.

What decides itPartner-sourced deals win more often because of the partner, not because partners bring deals that were already likely to close. (not shown yet)

Read the argumentHide the argument for Partner-sourced pipeline beats marketing-sourced pipeline

Verdict basisThe case to update is mostly aspiration.

Why the prior made sense

  • unprovenMost of the moments that shape a purchase happen outside the vendor's view, many of them through partners.
  • unprovenPartner-sourced deals win more often because of the partner, not because partners bring deals that were already likely to close.

What the challenge needs

  • unprovenAI buying agents skip the pre-sale moments where partners exert influence.
  • holdsThe channel's share of buying is falling.

Where it came from

Two claims travel together. Jay McBain of Canalys, an analyst firm covering the IT channel, says the modern buyer passes through 28 moments on the way to a purchase, many of them with or through partners. On podcasts he has said direct marketing might win four of the 28. No published method for the number turned up.

The second claim, that partner-sourced pipeline beats marketing-sourced, comes from ecosystem vendors. Crossbeam, which sells partner-data software, reports from its own network an 11.7% average win-rate lift when partners are involved. A 2023 survey Crossbeam co-produced has partner deals 53% more likely to close, self-reported by respondents.

What people say now

The challenge is a prediction. Sheena Wee of Omdia, which now owns Canalys, told a 2026 channel conference that AI agents will take an express route past the 28 moments, and that "the partner's value actually moves from the influence before the sale to the accountability after it."

The market data is mixed. Another Omdia analyst, Alastair Edwards, reports the channel's share of global IT spend fell from 69% to 65% in a year, driven by direct hyperscaler AI deals. He adds that partner influence is still increasing. Both voices come from the same corporate family as the origin.

Does the logic hold

The prior rests on two ideas. Most of what shapes a purchase happens out of the vendor's sight, much of it through partners. And partner deals win more because of the partner. The first is a number with no method. The second is the one that matters, and it has never been separated from selection. Partners tend to bring accounts that are already warm and qualified, and partner influence is often self-tagged. A higher win rate on those deals is what you'd see whether or not the partner made the difference.

The challenge doesn't help. Its causal claim, that agents bypass partner influence, is a forecast. The premise that holds, a four-point drop in channel share, tracks where money flows without showing whether a partner-sourced deal wins because of the partner.

What both sides miss

The closest outcome evidence is from consumer banking. In a 2011 study, Schmitt, Skiera and Van den Bulte found referred bank customers at least 16% more valuable, through better matching and lower churn. A replication confirmed the lower churn but found referred customers not necessarily more valuable. If referrals help mainly through matching, partners may be selecting better deals more than closing them. That would still be worth paying for, but it would change how partner pipeline is credited and forecast.

Nobody has run the comparison of partner-sourced and direct deals on accounts of similar warmth that would settle it.

Sources 6

unit economics

Price per seat

Update itAgents break the link between users and value and inference makes heavy users costly to serve, so seats alone stop fitting; pure usage trails on margin, and hybrid pricing has partial outcome support.

What decides itValue tracks the number of people using the product. (has changed)

Read the argumentHide the argument for Price per seat

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • changedValue tracks the number of people using the product.
  • changedServing one more user costs almost nothing, so pricing by user doesn't hurt margin.
  • holdsBuyers and their CFOs need spend they can predict.

What the challenge needs

  • holdsAgents doing the work break the link between users and value.
  • unprovenBuyers will accept a bill that moves with use.
  • failsPure usage or outcome pricing replaces seats.

Where it came from

Nobody invented per-seat pricing; it was simply how SaaS was sold. Tomasz Tunguz, then a VC at Redpoint, relaying Madhavan Ramanujam, drew the line in 2021: Salesforce and Adobe price per seat, Snowflake and Twilio per use. Until recently, seats and flat-rate subscriptions covered about half the companies Kyle Poyar surveys.

What people say now

The loudest voices say the seat is dying, and most of them sell products that agents use. One founder: "Pricing agents by the seat is like pricing electricity by the bulb." Another: "Per-seat assumes a human logs in." Poyar, whose newsletter is sponsored by a usage-billing vendor, finds hybrid models at 37% of companies, up from 25% a year earlier, and only 5% of investors prefer seats.

The market hasn't followed the slogan. SBI, a pricing consultancy, finds more than 80% of 321 SaaS companies still use some seat metric, with outcome pricing at two companies. Bain reviewed more than 30 vendors adding generative AI: about 65% layered an AI meter on top of seats, about 35% raised seat prices, and none moved fully to usage or outcomes. This describes vendor choices rather than results.

Does the logic hold

Seat pricing rested on three ideas. Value tracks the number of people using the product. Serving one more user costs almost nothing. And buyers need spend they can predict. AI changed the first two. An agent can do a seat's work without a login, and inference makes a heavy user costly to serve. ICONIQ, a growth investor, has AI products running gross margins of 41% to 45%, against the 80% SaaS assumed. Those two changes are enough to update a pure seat price.

The challenge goes further, to pure usage, and the outcome data doesn't carry it there. In Benchmarkit's 2026 survey of 342 companies, usage-only pricing had the highest median NRR, 108% against 98% for seat-only. It also had the lowest total gross margin, 62% against 76% for seats. The survey is vendor-sponsored, cross-sectional and confounded by category, and it doesn't isolate AI products.

Whether buyers accept a bill that moves with use is unproven. Cursor, Replit and GitHub Copilot each pushed users toward metering and met a backlash, and Cursor's CEO wrote that the change "hurt that trust." Yet usage-only firms post the highest NRR, so their buyers do pay more as use grows. No one has published what the switches did to revenue or churn.

Hybrid is partly supported. Fixed fee plus usage has the top Rule of 40 scores in Benchmarkit's data. Outside software, a telecom pricing field experiment found a two-part tariff, an access fee plus a usage price, lowered retention and usage against pure pay-per-use yet still made the most profit.

What both sides miss

The backlash was aimed at variance more than at metering. Each of the three companies restored predictability with spend limits, credits or a seat plus a meter, and none went back to seats.

The defenders miss the pressure from the buying side. BCG, a consultancy, finds 40% of IT buyers name seat reduction as their main lever to cut software spend. When agents do part of the work, the seat count is the first number a buyer's CFO goes after. A pure seat model loses ground even where seats survive as a unit.

Sources 5

A magic number above 0.75 means put more into sales and marketing

Update itThe 0.75 line has no traceable source and nobody has tested whether spending more above any line pays back, so read the magic number as a prompt to look at cohorts, not a trigger to spend.

What decides itThe 0.75 line was set from evidence of where the next dollar of sales and marketing pays back. (doesn't hold)

Read the argumentHide the argument for A magic number above 0.75 means put more into sales and marketing

Verdict basisNever sourced.

Why the prior made sense

  • failsThe 0.75 line was set from evidence of where the next dollar of sales and marketing pays back.
  • unprovenA magic number above the line tells you an extra dollar of sales and marketing will pay back.
  • unprovenRevenue added in a quarter comes from the prior quarter's sales and marketing spend.

What the challenge needs

  • holdsThe magic number is a weak guide to go-to-market efficiency, with or without AI.
  • unprovenAI-native economics, with thin gross margins and fast self-serve growth, make SaaS efficiency ratios misleading.

Where it came from

Rory O'Driscoll of Scale Venture Partners, a VC, wrote it down in 2010. Take the change in quarterly subscription revenue, annualise it, and divide by the prior quarter's sales and marketing spend. Above 1x, he called more spend "a compelling business investment." Below 0.5x, the company hadn't figured out its model yet. Credit for the metric is contested, but the ratio is well sourced.

The line most teams quote, 0.75, is not. It isn't on O'Driscoll's page. One trail leads to a 2008 guest post through a secondary source and ends there. Boards used the ratio because revenue lags spend and they needed a quick read on whether the next dollar would come back. One quarterly number stood in for a cohort analysis most companies couldn't run.

What people say now

The case against it predates AI. McKinsey, a consultancy, looked at about 20 operating metrics across public software companies in 2021 and found the magic number showed "almost no correlation" with value multiples. Ray Rike, who sells benchmarks, puts its R-squared against revenue multiples at 0.07 and calls it never a good primary metric. These are the same kind of value studies the LTV:CAC entry in this series uses for payback.

The AI case is newer and thinner. Bessemer Venture Partners reports its fastest-growing AI companies running gross margins near 25%, often negative, and about $1.13 million of revenue per employee. The argument is that SaaS ratios mislead for companies like that. No source with data makes it, and High Alpha, a VC, surveyed more than 800 companies and found AI-native firms scoring better on the Rule of 40 than AI-enabled ones, 34% against 22%.

Does the logic hold

The prior turns a ratio into a spending rule, and it needs three ideas. The line marks where the next dollar pays back. A ratio above the line tells you that it will. And a quarter's new revenue comes from the prior quarter's spend.

The first fails. The originator drew 1x and 0.5x as judgment calls, and nobody can trace 0.75 to a derivation or a test. The second is unproven. The ratio averages last quarter's spend against revenue change; no source tests whether it predicts the return on the marginal dollar, and its link to value is weak. The third is the built-in lag. For products that grow through self-serve and usage, revenue may move with the product rather than with last quarter's spend. That is plausible, but no source with data shows it.

So the update doesn't need AI. The line was never sourced, and the causal claim under it was never tested. That leaves the ratio as a signal to dig into cohorts, not a rule to spend.

What both sides miss

Defenders point to the benchmark tables that still publish the ratio. A benchmark describes where companies sit; it doesn't show that spending more above the line paid off.

Critics lean on valuation, which is not the return on sales and marketing. A ratio can track value poorly and still flag a broken motion. Nobody has tested whether companies that raised spend after crossing the line got the return.

Its cousins fare differently. The Rule of 40 tracks value better in the same studies, and the burn multiple has no outcome evidence at all.

Sources 5

Quarter-end discounts close the gap to the number

Update itThe best deal-level evidence shows quarter-end discounts mostly move deals between quarters at a lower price, costing several percent of revenue, so keep bounded discount authority and stop treating the quarter-end discount as found money.

What decides itEnd-of-quarter discounts win revenue that would otherwise be lost, rather than moving deals between quarters at a lower price. (doesn't hold)

Read the argumentHide the argument for Quarter-end discounts close the gap to the number

Verdict basisEvidence against it.

Why the prior made sense

  • holdsReps hold customer information managers lack, so bounded rep discretion with escalation prices better than a fixed price.
  • failsEnd-of-quarter discounts win revenue that would otherwise be lost, rather than moving deals between quarters at a lower price.
  • unprovenAn approval ladder is enough to contain discount leakage.

What the challenge needs

  • holdsModel-based price guidance beats seller discretion on realized margin.
  • unprovenBuyer-side AI negotiators remove the quarter-end lever.
  • unprovenDeploying the product beats discounting as the closing move.

Where it came from

Nobody coined it. Most deal desks use discount authority that climbs with rank, where a rep can give a little, a manager more, and a VP more again. Jason Lemkin, who runs SaaStr's events and media business and invests in SaaS companies, gave the plain statement of the habit in a 2021 essay. Reps quote list, then discount at the end of the month or quarter, as far as they're pushed. His advice was to control discounting rather than stop it, and to mark list up so there's room.

The economics trail is older. In 1998 Paul Oyer showed that manufacturing sales run high at fiscal-year end and low at the start, and argued that nonlinear pay gives sellers a reason to move both prices and timing.

What people say now

The practice is mostly defended by being normal. The newer argument is that AI takes the discount decision over. One version gives the seller a model-set price. Another says buyer-side negotiating agents will strip out the quarter-end lever. Lemkin himself now argues the other way, urging teams to stop discounting and get the product deployed instead. He offers anecdotes and a worked example without cohort data.

Vendr, which sells software purchasing and negotiation, has data that cuts against the agent story. Across 12,446 contracts its CEO finds discounts "fairly stable throughout the year," about 5.5 points between the best and worst month. Buyers' timing leverage was already modest before agents arrived.

Does the logic hold

The prior rests on two ideas. Reps know what a buyer will pay, so bounded discretion with escalation prices better than a fixed list. And a quarter-end discount wins revenue that would otherwise be lost. The first holds, as a study of 181 German industrial firms found delegating price authority improves performance, more so when reps know more than their managers.

The second fails. Ian Larkin's 2014 study of one enterprise software vendor found reps pulling deals forward and pushing them back around their commission schedule, and discounting the moved deals more. The discounts tracked the rep's commission stake rather than the deadline. The lost revenue came to 6 to 8%, about what the vendor spent on commissions. That gaming happened inside the vendor's approval process, so the ladder alone didn't contain it.

None of this needs AI. The diagnosis predates it. AI only matters as a remedy, and one remedy has a test: in a randomized field experiment at a B2B retailer, showing reps a model's price raised profit on treated quotes by 11%.

What both sides miss

The same experiment found reps still priced complex or unusual quotes better than the model, and a human with the model beat either alone. That is the case for keeping bounded discretion while moving the anchor to a model price.

Defenders of the quarter-end push miss that it mostly buys timing at a lower price. Critics who wait for buyer agents miss that the cost is already on the sell side, in how the comp plan is shaped.

Sources 6

LTV:CAC of 3 or more, and CAC payback under 12 months

Too early to callThe claim that AI margins make the ratio obsolete fails because LTV was always margin-based, but nobody has tested whether hitting 3:1 or a 12-month payback marks a healthier company, and payback tracks valuation only loosely.

What decides itA company that recovers its CAC within a year is healthier than one that doesn't. (not shown yet)

Read the argumentHide the argument for LTV:CAC of 3 or more, and CAC payback under 12 months

Verdict basisThe evidence splits.

Why the prior made sense

  • changedWith gross margins near 80%, revenue could stand in for profit.
  • unprovenA company that recovers its CAC within a year is healthier than one that doesn't.
  • unprovenOne ratio, set as a target, can steer a go-to-market motion.

What the challenge needs

  • holdsAI products run gross margins well below the SaaS norm.
  • failsLower margins make the 3:1 rule obsolete.
  • holdsPayback should be computed on gross margin, with inference in cost of goods.

Where it came from

David Skok, a VC at Matrix Partners, wrote both rules in 2010: LTV should be at least three times CAC, and CAC should pay back in under twelve months. He called them his rules of thumb, not findings, and his LTV was calculated on gross margin from the start. He later wrote that the twelve-month guideline came from a time when capital was hard to raise, and moved healthy payback to about twenty months.

What people say now

The AI-era claim is that the ratio is obsolete. AI products run gross margins well below the 80% that SaaS assumed. ICONIQ, a growth investor, reports 41% in 2024 rising to 45% in 2025. Advisors' posts now call the 3:1 ratio obsolete for AI software.

Finance writers mostly shrug. Seth Girsky of Inflection CFO: "The 3:1 ratio was an output, not an input." Ben Murray, writing from a CFO community: "Cap LTV or skip it. CAC payback is what most finance leaders actually manage to." Finance keeps payback and changes the accounting under it, calculating payback on gross margin with inference costs broken out in cost of goods.

Does the logic hold

Skok's rules rested on three ideas. Software margins were high enough that revenue could stand in for profit. A company that got its acquisition spend back within a year was healthier than one that didn't. And one ratio could summarise whether a go-to-market motion was working.

AI changed the first idea for AI products. The challenge concludes that the ratio is obsolete, and that step fails on definition. Skok's LTV was margin-based from the start, so lower margins already lower it. The proposed fix is the original rule.

The second idea decides the verdict, and nobody has tested it. No study checks whether companies at or above 3:1, or under twelve months, go on to grow faster, survive longer or earn higher multiples. The closest evidence is correlation with valuation, and it is loose. Across public SaaS, CAC payback explained between about 10% and 34% of the spread in revenue multiples from quarter to quarter, by the count of Ray Rike, who sells benchmarks. A VC's longer series, from DNX Capital, put it at roughly 20 to 30%. Growth explains more than any efficiency measure. So efficiency matters somewhat, and the cut-offs are untested.

What both sides miss

The medians are market state, not a verdict. KeyBanc's 2024 private-SaaS survey estimates median fully loaded payback at 20 months for that year. CJ Gustafson's tracking of the top ten public software companies has their median at about 28 months. Those figures show that most companies miss twelve months. They don't show that missing it does no harm, so a threshold can't be retired on the evidence that few meet it, any more than it can be defended on Skok's say-so.

AI does change the inputs. Margin-adjusted payback gets longer, and one benchmark publisher reports search channels getting less efficient, which raises CAC.

Sources 4

ACV sets the sales motion: self-serve at the low end, inside sales in the middle, field sales at the top

Too early to callAI may make a sales touch cheap enough to sell below the old ACV floor, but that case is a model and one anecdote, while the best-evidenced link, that higher-ACV customers stay longer, still holds.

What decides itAI cuts the cost of a sales touch enough that sales-assisted selling pays at ACVs that used to need self-serve. (not shown yet)

Read the argumentHide the argument for ACV sets the sales motion: self-serve at the low end, inside sales in the middle, field sales at the top

Verdict basisThe case to update is mostly aspiration.

Why the prior made sense

  • holdsA human sales touch has a fixed cost that sets a floor on viable ACV.
  • holdsLow-ACV buyers prefer to buy without talking to sales.
  • holdsACV predicts customer durability, so higher-ACV customers stay longer.

What the challenge needs

  • unprovenAI cuts the cost of a sales touch enough that sales-assisted selling pays at ACVs that used to need self-serve.
  • unprovenAI personalisation removes the old trade-off between scale and personal service.
  • holdsSelf-serve can now reach enterprise ACVs, so ACV no longer fixes the entry motion.

Where it came from

In 2014 Christoph Janz of Point Nine, a VC, sorted the ways to build a $100 million software business by what each customer pays: flies at about $10 a year, mice at $100, rabbits at $1,000, deer at $10,000 and elephants at $100,000 or more. Each needs its own way to acquire customers. At $10,000, he wrote, an account "usually isn't enough to make traditional enterprise field sales work," so you use inside sales. At $1,000, "cold calling usually doesn't work."

RevOps read it as a map where ACV picks the motion. The scarcity under it was the cost of a human sales touch.

What people say now

The update case comes from two VCs. Tomasz Tunguz of Theory Ventures notes that most companies hire sellers at $15,000 or more in contract value, then models what AI might do. Assume a 50% productivity gain, and a seller at a $10,000 ACV goes from 24 to 36 deals a year. "The minimum viable ACV drops." Jeffrey Bussgang of Flybridge offers one example: a company selling to small businesses at about $3,000 a year with AI-personalised outbound, a segment he says was unprofitable to reach that way before.

The top of the map is moving the other way. Menlo Ventures, a VC, surveyed about 500 enterprise buyers and puts product-led purchases at 27% of AI application spend, against 7% in traditional software. It notes Cursor reached $200 million in revenue before hiring enterprise sellers.

Does the logic hold

The prior rests on three ideas. A human sales touch has a fixed cost, which sets a floor on the ACV it can serve. Low-ACV buyers would rather not talk to sales. And higher-ACV customers stay longer. All three hold. Nobody argues a person's cost per touch has fallen, and the retention data below still favors larger contracts.

The challenge needs AI to cut the cost of a sales touch enough that selling pays below the old floor. That is the deciding premise, and it rests on a model with an assumed productivity gain and one company's story. No cohort or before-and-after study was found. The top end is better supported, but even there sales gets layered on once a product reaches large accounts. That is Janz's elephant logic arriving later, not failing.

So the two AI effects pull in opposite directions, and the map blurs at both ends. Neither end has outcome data. The case to update is mostly projection.

What both sides miss

The update side skips retention, the best-evidenced premise here. SaaS Capital, a lender to SaaS companies, finds gross retention of 95% above $250,000 ACV against about 91% below. Selling at a lower ACV doesn't fix the churn that comes with a lower ACV.

The defend side misses what the map never claimed. SaaS Capital also finds ACV doesn't correlate with growth rate: the map tells you how to sell, not how fast you'll grow.

The product-led growth entry in this series looks at the low end from the other side. Free users got costlier just as sales touches may be getting cheaper, and both push sales assistance down-market.

Sources 5

customer and renewal

The health score tells you who will churn

Update itA red, yellow and green score weighted by judgment was never shown to predict churn, while models trained on past renewals have been, so replace the colour you configure with a model you backtest, and don't assume generative AI has already fixed it.

What decides itModels trained on past renewal outcomes predict which B2B accounts will churn, well enough to target saves. (holds)

Read the argumentHide the argument for The health score tells you who will churn

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • unprovenA composite of usage, survey scores, support load and CSM sentiment, weighted by judgment, separates future churners from renewers months ahead.
  • failsThe inputs you can count, such as logins and survey scores, are the inputs that matter.
  • holdsA team has enough past churn events to set the weights.

What the challenge needs

  • holdsModels trained on past renewal outcomes predict which B2B accounts will churn, well enough to target saves.
  • unprovenGenerative AI that reads calls and email makes the health score accurate.
  • unprovenTeams will backtest a score instead of trusting its colour.

Where it came from

The health score has no author. It grew with the customer-success platforms of the 2010s, each of which ships a red, yellow and green scorecard: usage, survey scores, support tickets and the CSM's own read, weighted however the customer chooses. Gainsight describes the score as one input to renewal, "determined by you." The rule it carried was simple. Watch the colour, and the colour tells you who will leave.

It was built for scarce attention. A CSM's book ran to dozens or hundreds of accounts, and the data sat in several tools, so one number let a manager rank the book without opening every account.

What people say now

Darlene Kelly, who directs customer-success research at TSIA, says "traditional customer health scores are failing to predict churn" and has launched research to measure it, with no data published yet. Gartner analysts, in a client note whose public abstract shows no data, recommend adding generative AI to health scores "to improve account churn and growth candidate identification." Practitioners are more careful. Campbell Robertson, who sells CS and AI advisory work, puts AI summarization in the "works today" column and "predictive churn scores" among the things that are "still aspirational" and "often mislead more than they help."

Gainsight, which sells customer-success software, defends the idea with its own cohorts. Accounts with a verified outcome, a customer's stated return, three months before renewal had a 30% higher gross retention rate, and low adoption with a verified outcome beat high adoption without one.

Does the logic hold

The prior rested on three ideas. A judgment-weighted composite separates churners from renewers months ahead. The inputs you can count are the ones that matter. And a team has enough past churn to set the weights. The first was never tested in public, in either direction. The second fails on the evidence there is: Gainsight's own data makes usage necessary but not sufficient, and a study of 3,959 software subscriptions found usage improves churn prediction only once it is structured by timing and granularity, which is weighting a model learns, not one a person sets.

The challenge's strongest premise holds. In a study of about 5,000 wholesale B2B customers, a model trained on past outcomes ranked churn risk well, and in a randomized field test, calling the customers it flagged cut churn against random targeting and lifted revenue. So the update follows: score from outcomes, not judgment. What does not follow is that generative AI is the fix. No outcome data yet shows that reading calls and email makes the score more accurate.

What both sides miss

Critics say "health scores are dead" when the evidence says something narrower. The configured colour was never validated, and a backtested model was. Defenders cite cohorts where the right inputs separate retention, which defends a well-built score, not the one on the dashboard.

Both skip the size problem. B2B customer counts are typically small, which makes a model harder to train. A team with few churn events may not have the history to backtest anything, and the honest score for that team says so.

Sources 6

NPS is the one number that predicts growth

Update itPeer-reviewed tests found NPS no better than plain satisfaction at predicting retention or growth well before AI arrived, and AI changes how you would measure NPS, not whether it predicts growth.

What decides itNet Promoter Score predicts a firm's future growth better than other customer metrics. (doesn't hold)

Read the argumentHide the argument for NPS is the one number that predicts growth

Verdict basisEvidence against it.

Why the prior made sense

  • failsNet Promoter Score predicts a firm's future growth better than other customer metrics.
  • changedSurvey answers are an honest read of customer loyalty.
  • changedA survey is the only way to read sentiment across the customer base.

What the challenge needs

  • unprovenAI-inferred sentiment from customer interactions predicts retention and growth better than surveyed NPS.
  • unprovenOrganizations are abandoning NPS as a measure.

Where it came from

In December 2003 Fred Reichheld of Bain, the consultancy that co-owns the Net Promoter trademark and sells NPS work, published "The One Number You Need to Grow" in Harvard Business Review. One survey question, would you recommend this company to a friend, was "the best predictor of top-line growth," and the share of promoters tracked growth among competitors in most industries studied. It answered a real problem. Satisfaction surveys were long and hard to act on, and leaders wanted one number the front line could move.

What people say now

The originator has moved. In "Net Promoter 3.0" (2021), Reichheld and his Bain co-authors concede that self-reported scores have been gamed, with "pleading," "bribery" and "manipulation," and introduce Earned Growth Rate, built on audited revenue and "far more resistant to gaming." The growth link moved off the survey.

Bain has also shown the survey can be skipped. Its Predictive NPS work inferred promoter and detractor status from transactions, call metrics and transcripts, with over 80% accuracy and four times the lift on detractors in one financial-services build. Gartner predicted in 2021 that by 2025 more than three quarters of organizations would abandon NPS as a measure of success for customer service and support. That prediction is limited to customer service, and it rested on a survey of 42 heads of customer service.

Does the logic hold

The prior rests on three ideas. NPS predicts growth better than other customer metrics. Survey answers are an honest read of loyalty. And a survey is the only way to read sentiment across the base.

The first was tested and failed long before AI. Keiningham and colleagues (Journal of Marketing, 2007), on 21 firms and over 15,500 interviews, failed to replicate NPS's "clear superiority." Keiningham was at Ipsos Loyalty, a competing measurement firm, but de Haan, Verhoef and Wiesel (2015) reached a similar answer on 93 firms. Top-two-box satisfaction predicted retention best, NPS nearly tied, the best metric varied by industry, and combining metrics beat any one. Jeff Sauro, re-running Reichheld's own data against future rather than past growth, found NPS explained on average 38% of growth variance at two years and 30% at four. Some signal, not the one number.

AI changed the other two premises. Gaming broke the honest-answer premise, and inference broke the survey-only one. Neither touches the growth claim. A model that predicts the NPS answer has predicted the answer, not growth, and no outcome evidence shows AI-inferred sentiment predicting growth better.

What both sides miss

The strongest defence is narrower than the prior. Baehre and colleagues (2022) found NPS predicted short-term sales growth only when measured as changes, on a sample of all potential customers. NPS levels among current customers, the number most teams track, did not predict growth. Critics who say "AI killed NPS" miss that the case was already made. Defenders who cite the 2022 result are defending a brand-health tracker, not the score on the CS dashboard.

Sources 6

Aim for 120% net revenue retention, with gross retention around 90%

Update itNRR still goes with growth, but usage ramps and committed credits can lift it without durable demand, and 120% is now a top-decile result, so read committed and usage expansion apart, with gross retention beside them.

What decides itAn expansion counted in NRR reflects durable customer demand, because seat and license expansions are contractual and spread evenly across the base. (has changed)

Read the argumentHide the argument for Aim for 120% net revenue retention, with gross retention around 90%

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • holdsHigher NRR produces faster growth and a higher valuation.
  • changedAn expansion counted in NRR reflects durable customer demand, because seat and license expansions are contractual and spread evenly across the base.
  • changed120% is an attainable bar that separates the best.
  • holdsGross retention of about 90% is the floor for B2B software.

What the challenge needs

  • holdsUsage and AI pricing let NRR rise without durable demand, through ramps, committed credits and breakage.
  • unprovenNRR no longer predicts growth or value in the AI era.

Where it came from

Bessemer Venture Partners put the number in its State of the Cloud 2023 benchmarks for Series B and C enterprise software: net revenue retention of 100% is good, 110% better, 120% or more best. These were targets for raising money, not measured medians, and Bessemer added that they shift by stage and by segment. Its 2019 version had used segment bands instead, with 80 to 100% for small business, 90 to 120% for mid-market, and above 100% for enterprise. The flat 120% came later.

The gross retention floor comes from SaaS Capital, which lends to SaaS companies and says that to keep pace with peers, it "must be at least 90%."

What people say now

The defense is strong on one point, which is that NRR goes with growth. SaaS Capital's survey of more than 1,000 private B2B companies finds "a strong and exponential correlation" between the two.

The challenge is about what a point of NRR now means. L.E.K., a consultancy that advises software companies and their investors, writes that under usage pricing "NRR is often inflated in the early months of a new customer ramp." It suggests tracking contracted and usage revenue apart.

The bar itself moved. Tomasz Tunguz of Theory Ventures, a VC, found median net dollar retention across 25 public software companies fell from 125% in 2022 to 112% in 2025. The private median sits near 101%.

Does the logic hold

The prior rested on three ideas. Higher NRR produces growth and value, an expansion reflects durable demand because seat adds are contractual, and 120% is an attainable bar for the best.

The first still holds, with caveats. Part of the growth link is mechanical, since expansion revenue is part of growth. The valuation link swings with the market. Ray Rike, who sells benchmarks, reports NRR explaining 40% of the variance in public multiples in mid-2021 and 7% in early 2023. The third has moved. 120% was the 2021 to 2022 public norm and is now a top-decile result.

The second is the one that changed. A seat expansion is a discrete, contracted event; a usage expansion can be a ramp that runs high early, and committed credits can hold it up by contract. The challenge doesn't need NRR to stop predicting growth, and nobody has shown that it has. It needs only that a point of usage NRR is not a point of seat NRR, and that mechanism holds. The price-per-seat entry in this series covers why usage-priced firms post higher NRR; this one asks whether that NRR means the same thing.

What follows is a revision, not a retirement. Keep NRR, drop 120% as a universal bar, and read committed expansion apart from usage expansion.

What both sides miss

Gross retention says less about growth than the 90% floor implies. SaaS Capital finds a median of 91%, 95% above $250K ACV, and little correlation between gross retention and growth. The floor guards the base; it doesn't predict the rate.

For low-priced AI apps the floor isn't close. ChartMogul, a billing vendor, with Kyle Poyar, found AI-native median gross retention of 40% and NRR of 48%, with gross retention at 23% below $50 a month. Poyar calls the AI buckets, about 50 companies each, "directional."

And no one has followed usage-inflated NRR forward to see what happened to growth or churn.

Sources 5

Keeping a customer costs five times less than winning one

Update itThe five-times ratio traces to an "experts agree" line, the 25 to 95% profit band was never measured, and loyal customers were not found to be cheaper to serve, so keep the case for retention but drop the numbers.

What decides itRaising retention by 5 points raises profit by 25 to 95%. (doesn't hold)

Read the argumentHide the argument for Keeping a customer costs five times less than winning one

Verdict basisNever sourced.

Why the prior made sense

  • unprovenAcquiring a customer costs about five times as much as keeping one.
  • failsLong-tenured customers cost less to serve and pay more, so retention compounds profit.
  • failsRaising retention by 5 points raises profit by 25 to 95%.

What the challenge needs

  • holdsIn B2B software, net revenue retention is a dominant driver of company value.
  • unprovenAI lowers acquisition cost, so retention's cost advantage narrows.
  • unprovenAI makes switching easier, so churn rises and retention matters more.

Where it came from

Two numbers travel together. The first, that winning a customer costs five times as much as keeping one, traces through the loyalty researchers Keiningham, Vavra, Aksoy and Wallard to work by the Technical Assistance Research Project in Washington in the late 1980s. An HBR article by Hart, Heskett and Sasser in 1990 lent it credibility, and the trail ends in an "experts agree" line with no measurement behind it.

The second comes from Reichheld and Sasser's "Zero Defections" (HBR, 1990): companies could boost profits by almost 100% by keeping 5% more customers, drawn from client examples. Amy Gallo's 2014 HBR article gave it the wording everyone quotes, "25% to 95%," and added that acquisition is "five to 25 times more expensive… depending on which study you believe," naming no study.

What people say now

The source Gallo links is a 2001 Bain brief by Reichheld. It says that "in financial services, for example, a 5% increase in customer retention produces more than a 25% increase in profit." One industry, more than 25%. No 95% appears.

Keiningham and colleagues, whose firm sells loyalty research, called the five-times ratio a myth in 2005: acquisition and retention costs get misallocated, and the answer shifts with product lifecycle and customer mix. Jon Lombardo of LinkedIn's B2B Institute, which has an interest as a seller of acquisition advertising, adds that the "5%" was a cut in defection from 10% to 5%, half of all defections, and assumed it cost nothing to achieve.

Does the logic hold

The prior needs three things. Acquisition costs about five times retention. Long-tenured customers cost less to serve and pay more, so retention compounds. And five points of retention buy 25 to 95% more profit. The first was never sourced. The third was never measured as stated. The second is the mechanism under both, and it failed on outcome data. Reinartz and Kumar (2002) tracked over 16,000 customers at four companies for four years and found that "in none of the four companies" were long-standing customers consistently cheaper to manage. At the B2B service provider, loyal customers cost more to serve.

AI changed none of these premises. The AI-era arguments, that cheaper acquisition narrows retention's edge or that easier switching raises churn, have no outcome evidence either way.

What both sides miss

The case for retention survives without the numbers. McKinsey's study of over 100 B2B software companies found the top quartile by valuation ran 113% net revenue retention against 98% for the bottom quartile, and top-quartile retention companies held higher valuations across cycles. That is correlational and tracks value rather than profit, though it is real. Critics who stop at "the numbers are fake" lose that. Defenders who need the numbers to make the case are leaning on a line its own source does not support.

Sources 5

Hire one CSM for every $2M of ARR

Too early to callAI is taking admin out of the CSM job and leaders want bigger books, but nobody has measured whether retention holds when books grow, and the $2M ratio itself was never shown to protect retention in the first place.

What decides itHolding coverage near $2M of ARR per CSM protects retention, so stretching the book loses revenue. (not shown yet)

Read the argumentHide the argument for Hire one CSM for every $2M of ARR

Verdict basisThe case to update is mostly aspiration.

Why the prior made sense

  • unprovenHolding coverage near $2M of ARR per CSM protects retention, so stretching the book loses revenue.
  • changedA CSM's capacity is fixed by manual work: prep, data pulls and notes.
  • failsOne number, about $2M, fits every segment.

What the challenge needs

  • unprovenAI lets a CSM cover more ARR without hurting retention.
  • unprovenCoverage designed as a system of roles retains better than coverage set by a ratio.

Where it came from

Jason Lemkin of SaaStr, who runs its events and media business and invests in SaaS companies, wrote the rule around 2013: hire one customer success manager for every $2M of ARR. He framed it as "where most of us end up at, at scale," an observation of what companies did, not a tested optimum. The KeyBanc SaaS survey he cites put the overall figure near $1.7M, and by segment about $1.3M for small business, $1.5M mid-market and $2.6M enterprise.

The ratio rationed CSM hours. Proactive coverage needed a person on each relationship, and a ratio turned ARR into headcount the way capacity planning turns quota into reps.

What people say now

The pressure is to stretch the book. ChurnZero, which sells customer-success software, finds call summarization is the top AI use in CS, and 83% of leaders name reducing CSM admin or growing the managed book as a top scaling priority.

Jeff Breunsbach, a CS practitioner, goes further, arguing that the ratio "anchors you to a linear growth model" and coverage should be designed as a system of roles. That is an argument with no result behind it.

Does the logic hold

The prior needs three things. Coverage near $2M per CSM protects retention. A CSM's capacity is fixed by manual work. And one number fits every segment. The third already fails on the rule's own sources, which give a range from about $1.3M to $2.6M by segment. The second is the one AI changed, as summarization and data pulls take less time and leaders plan to spend that time on bigger books.

The first was never measured, and it decides the question. A survey of 132 tech companies by CJ Gustafson found "no meaningful correlation between staffing ratio and revenue retention or renewal rate," which is self-reported and does not settle it. ChurnZero finds companies with CSMs report higher net retention than those without, 98% against 90%, which says having CSMs matters while leaving the ratio open. The challenge's own causal premise is equally open, because nobody has published retention for books that grew.

What both sides miss

What has not been measured is the whole point. No published study links ARR per CSM to gross or net retention, and nobody has published whether bigger books kept their customers. Defenders hold a number that was never tied to retention. Critics read a bigger book as success before anyone has checked what it did to retention.

Sources 4

Run a business review with every customer every quarter

Too early to callAI takes away the data-assembly job that filled most reviews, but nobody has shown that the quarterly cadence itself, as distinct from a value conversation at the right moment, raises renewal.

What decides itA fixed quarterly cadence of reviews raises renewal and expansion. (not shown yet)

Read the argumentHide the argument for Run a business review with every customer every quarter

Verdict basisThe evidence splits.

Why the prior made sense

  • changedCustomers can't see their own usage and value data, so the vendor must package it for them.
  • unprovenA fixed quarterly cadence of reviews raises renewal and expansion.
  • unprovenBuyers value the review and show up for it.

What the challenge needs

  • unprovenAI removes the data-assembly work, so the quarterly ritual adds no value.
  • holdsThe format was failing for reasons that predate AI: backward-looking, vendor-centric and costly to prepare.

Where it came from

Nobody owns the QBR. The earliest dated treatment that reads as canonical is Gainsight's, in 2014. Lincoln Murphy called it "one of the most important activities your CSMs will perform." Dan Steinman asked whether it was "Value or Fluff?" and suggested calling it an executive business review, to get the sponsor in the room. Murphy also passed on the inside joke that QBRs rarely happen quarterly, so the cadence was aspirational from the start.

The format rationed two things. Customers couldn't easily see their own usage and return, and executives were hard to get in a room. A standing quarterly slot fixed both before renewal came due.

What people say now

Chloe Siegmann, a CS practitioner, argues customers can now ask AI for their own usage trends on demand, so the data-narration part of the meeting is gone. Jason Lemkin, who runs SaaStr's events and media business and invests in SaaS companies, writes as a buyer and calls QBRs "simply upsell exercises" and says he had already read the data being shown.

The defence is vendor data. A 2023 buyer survey commissioned by Clientshare, a vendor, found 68% of 200 buyers at large firms called regular reviews the most important moment for a supplier to show value. That is stated preference. A case study from Matik, also a vendor, reports that its customer's accounts that got a business review renewed 32.8% higher than those that didn't.

Does the logic hold

The prior rests on three ideas. Customers can't see their own data, so the vendor packages it. A quarterly cadence raises renewal. And buyers value the meeting enough to attend. Practitioners report that AI changed the first because the data job shrinks when customers can query their own usage. Buyer value is split, with a commissioned survey on one side and buyers like Lemkin on the other.

The cadence premise is the one that decides, and it has never been tested. The lone cohort figure measures having any review at all rather than a quarterly one, and selection skews who gets reviewed because teams may pick accounts already likely to renew. The challenge has the same gap. "AI made the ritual pointless" remains an unverified argument. Its stronger premise, that the format was failing anyway, holds, but it predates AI, so AI is not what moved the verdict.

What both sides miss

Both argue about the meeting when the open question is the trigger. Donna Weber, who sells onboarding and CS consulting, proposes holding a value review whenever there is value to show instead of following the calendar. No test yet separates a quarterly cadence from a value conversation at the right moment, and that is the comparison that would settle this.

Sources 6

systems and data

The CRM is the system of record: if it's not in Salesforce, it didn't happen

Update itThe CRM is losing its job as the interface but not as the record, and the record never held the buying committee, so treat it as the record of amount, stage and owner, not of the deal.

What decides itWhat gets logged covers what matters about the deal. (doesn't hold)

Read the argumentHide the argument for The CRM is the system of record: if it's not in Salesforce, it didn't happen

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • changedActivity exists only if a rep types it in, so the CRM is the one place a manager can see a deal.
  • changedThe CRM is where the work happens, so it is the natural home for the record.
  • failsWhat gets logged covers what matters about the deal.

What the challenge needs

  • holdsMachines now capture calls and email, so evidence of the work exists outside the CRM.
  • holdsThe work has moved into chat and email, away from the CRM's interface.
  • failsWhen the interface and the typing move, the record moves with them.

Where it came from

Nobody wrote it down first. By 2016 one sales-ops team called "if it's not in there, it didn't happen" its mantra, in a piece noting the line was already on memes, T-shirts and mugs. By 2023 SaaStr was reporting it as the golden rule of Salesforce's own forecast process. Under the slogan sits an architecture: syncs that let the CRM win for accounts and opportunities, and a weekly forecast call that only counts what's logged.

What people say now

The update side says the CRM is losing its job. a16z, which invests in the layer above the CRM, says it is "turning into just an input; one of many inputs, into the systems of intelligence which we use to get work done." Dominik Dudy, a revenue leader across EMEA SaaS, adds a regional point: "The German AE who tells the buyer 'let me log this in Salesforce' is the AE who never gets the next call."

The defenders say agents raise the CRM's value. Pratik Mehta, a fractional RevOps consultant who builds HubSpot systems: "When agents start doing the work, the CRM does not become less important. It becomes the brain." Benn Stancil, from the data side: "it seems like we'll always have systems of record that are full of important information."

The awkward fact sits in the same a16z piece. Its own GTM survey found that "CRM usage has actually risen" since AI tools arrived, because agents writing call notes back give reps fresh reason to open it.

Does the logic hold

The mantra rested on three ideas. Activity existed only if a rep typed it in. The CRM was where the work happened, so it was the natural home for the record. And what got logged covered what mattered about the deal. The first two have changed. Machines capture calls and email, and the work has moved into chat and email.

The update side is right about those changes and goes too far with them. Its conclusion, that the CRM becomes one input among many, needs the record to leave along with the interface and the typing. The usage data says it hasn't, and both sides still treat the CRM as where the data lives. The premise that fails is the third, and it failed long before AI, because the record never held the buying committee or most meetings. So the update that follows is narrower. The CRM stays the record of amount, stage and owner, and stops being read as the record of the deal.

What both sides miss

Both sides underrate how much of the deal was never in the record. In my own CRM, almost no opportunities carried contact roles, and the few that did mostly left the role itself blank. When I needed "who was on this deal", I had to rebuild it from call participation. That's harder than it sounds. An invite list isn't an attendee list, and a meeting stored as scheduled may have happened or not.

What changed is who reads the gap. A manager knew the committee wasn't in the CRM. An agent reading the opportunity doesn't, and a wrong value in the record now reaches every agent that reads it.

Sources 5

B2B data decays 22.5 to 30% a year

Update itNo study stands behind the 22.5 to 30% figure, decay varies widely by role and industry, and corrections can do more harm than the decay, so measure your own rate instead of quoting one.

What decides itSomeone measured how fast B2B contact data goes stale. (doesn't hold)

Read the argumentHide the argument for B2B data decays 22.5 to 30% a year

Verdict basisNever sourced.

Why the prior made sense

  • failsSomeone measured how fast B2B contact data goes stale.
  • failsOne blended annual rate is close enough to plan refreshes for every role and industry.
  • changedRe-checking a contact takes a person's time, so a single planning number is the affordable choice.

What the challenge needs

  • unprovenOne vendor's measurement of its own database stands for the wider market.
  • failsJob changes are the decay that matters.
  • failsA corrected average, quoted over the right period, can stand in for your own rate.

Where it came from

From a footnote. HubSpot's database-decay calculator credits MarketingSherpa with 2.1% a month, which compounds to 22.5% a year. No study, year, sample or method is linked, and a 2026 search of MarketingSherpa's site came up empty. HubSpot used the figure to sell inbound lead flow. Later retellings rounded it to 30%, and a 70% figure circulates too.

What people say now

2026 brought a tracing wave, mostly from data vendors with an interest in the answer. Lusha, which sells contact data, looked for the original and found none. In its own database, about 12% of senior US sales leaders changed roles over 12 months and about 26% over 24. Its verdict: "The number is roughly right. It is being quoted over the wrong period."

The 70% figure measured something else. It comes from John Coe's business-card survey, first run in 2002 and updated in 2009: about 1,200 people reported what had changed on their card in the past year, including title changes inside the same company. Federal turnover data splits the average another way. In 2025, people left employers in accommodation and food services at an average of 5.5% a month, against 2.1% in finance and insurance. A figure means little without what it measured, its sample and its year.

Does the logic hold

The number rested on three ideas. Someone had measured it. One blended rate was close enough for every role and industry. And re-checking a contact took a person's time, so a single planning number was the affordable choice. The first never held. The second fails against the federal turnover data, where one industry loses people about two and a half times as fast as another. The third is the one that changed, because tools now re-check contacts on their own and make measuring your own rate cheap.

The leading correction keeps the second idea and swaps the number. Lusha's figure is better sourced than a footnote, but it is one vendor's database and one cohort, senior US sales leaders, and it counts only job changes. Quoting it over two years instead of one is still quoting an average. What follows from the premise that changed is not a better average. It is your own rate, measured by role class.

What both sides miss

Both sides treat decay as people moving. In my own builds, some of the worst damage came from the corrections. A vendor field that looked like "latest employer" was really the most recent start date, so a new board seat read as a job change. When I re-checked the people it had flagged as movers, nearly half were still at the company.

Decay also runs on different clocks. I once confirmed a senior executive as still in role, and three days later found a replacement announced two months earlier. A company's posture changes over months. One annual rate hides both.

And a job change doesn't decay a record so much as split it. The person now has two employments, and a model with one "current employer" field has to lose one of them.

Sources 3

Employee count and revenue are good enough to route and segment on

Update itData vendors now concede in print that employee count and revenue are modeled estimates, and a hand check flips many segment verdicts, so the fields can't be trusted near a routing threshold.

What decides itVendor employee counts and revenue are close enough to the real figures to sort accounts into segments. (doesn't hold)

Read the argumentHide the argument for Employee count and revenue are good enough to route and segment on

Verdict basisEvidence against it.

Why the prior made sense

  • failsVendor employee counts and revenue are close enough to the real figures to sort accounts into segments.
  • holdsChecking a company's real size takes a person's time per account, so the vendor number is the affordable stand-in.
  • changedA person looking at a routed account notices when it is obviously in the wrong segment.

What the challenge needs

  • holdsVendor size fields are modeled estimates, not verified facts.
  • holdsThe estimation error is large enough to change routing decisions near a threshold.

Where it came from

Nobody wrote this one down. It lives in configuration: a routing rule that sends accounts above some employee count to enterprise, a territory plan cut by revenue band, both filled by whichever data vendor the company bought. The fields fill reliably, and fill rate is easy to mistake for accuracy.

What people say now

The clearest statement against the prior comes from a data vendor. Apollo, which sells these fields: "Revenue and employee count fields in B2B data platforms are modeled estimates, not verified facts." Firecrawl, which resells several of the providers it tested, scored six of them against filed headcounts for 200 US companies, 150 of them public, in October 2026. A count within 1.2 times the filing earned full credit and one within twice it earned half. The best provider scored 77% on employee count; the rest scored between 46% and 57%. Anders Krohn of Kernel, another data vendor, puts the headcount problem in one line: "that's like counting every Uber driver as an Uber employee."

All three sell something here. One of them is conceding a weakness in its own product, which makes the concession worth more than the pitches.

Practitioners are quiet. I found almost no RevOps posts defending the fields and almost none attacking them, while routing keeps running on them.

Does the logic hold

Routing on these fields rested on three ideas. The vendor's numbers were close enough to sort accounts into segments. Checking a company's real size took a person's time per account, so the vendor number was the affordable stand-in. And a person looking at a routed account would notice when it was obviously in the wrong segment.

The first idea fails, and the vendors say so. The fields are modeled, Apollo puts its own headcount ranges at plus or minus 20 to 40%, and headcount sweeps in contractors and a parent company's staff. The errors also sit where routing decides, near the thresholds. AI changed the third idea. An agent applies the threshold literally, with nobody noticing that an obviously mid-market account has landed in enterprise. So the update follows from one premise that failed and one that changed. The second idea still holds because checking takes work, which puts the check at the thresholds instead of on every account.

What both sides miss

Nobody publishes how often a segment verdict changes when you check, which is the number that matters. I checked mine. On a sample of accounts verified by hand on the web, about 4 in 10 segment verdicts flipped against the vendor fields, even though both size fields were nearly always filled.

The errors had shapes:

  • After an acquisition, the count reads the parent, leaving a small company showing tens of thousands of employees.
  • A size-band picklist straddles the routing threshold, so the band can't answer the question.
  • A cluster of "large" accounts turned out to be acquired or defunct shells.

What worked was dull. Look the company up on its own LinkedIn page or site, keyed on its web domain. Name plus country plus industry fails on common names.

Sources 3

A human approves every write to the system of record

Update itApproving every agent write recreates the manual work, and research from other fields shows blanket review lets errors through while risk-based review catches as much for less effort, so tier review by consequence; how a field earns its way out is unsettled.

What decides itA person who approves a change has checked it. (doesn't hold)

Read the argumentHide the argument for A human approves every write to the system of record

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • changedWrites come slowly enough that a reviewer can look at each one.
  • failsA person who approves a change has checked it.
  • holdsA wrong write to the system of record costs enough to justify a check.

What the challenge needs

  • holdsApproving every agent write recreates the manual work the agent was meant to remove.
  • holdsConsequence can be judged by field, so low-stakes writes can leave review while high-stakes ones stay.
  • unprovenHow often reviewers edit a field's proposals shows when it is ready to leave review.

Where it came from

Admin canon for change control is to build in a sandbox, get it approved, and then deploy. When agents started writing to the CRM, the rule moved down from configuration to data. One consultant's governance advice for Salesforce admins still states it flatly: "At least one human review point in every AI-assisted workflow."

What people say now

The playbooks are moving away from "every write". The emerging line is a tier by consequence. One playbook spells out what never leaves review: "owner changes, merges, deletes, lifecycle or stage changes, and opt-out fields." Other rules graduate out of review one at a time.

The graduation test is the part nobody has settled. Chris Fitkin of metacto, a vendor, proposes one: "If humans constantly rewrite the agent's updates, the workflow is not ready for more autonomy."

Does the logic hold

The rule rested on three ideas. Writes came slowly enough that someone could look at each one. A person who approved a change had checked it. And a wrong write to the system of record cost enough to justify the check. Agents changed the first by writing at a volume no reviewer can read. That breaks the second, which was already weak: a reviewer facing more than they can read starts passing things through.

Nobody has measured this in CRM approval queues, but other fields have measured it. In clinical decision support, a 2026 meta-analysis found that in real-world use, 26.6% of erroneous computerized ECG interpretations were not corrected or recognized by the clinicians reviewing them. In primary care, the more reminders clinicians got in a visit, the fewer they accepted. The likelihood of accepting one fell 30% for each additional reminder per visit. Review gets worse as volume rises, which is exactly what agents bring.

The third idea still holds, and it is what the tiered model keeps. The challenge needs approving every write to recreate the manual work, which it plainly does, and consequence to be judged by field, so low-stakes writes can leave review. Clinical-trial monitoring tested that second move. In the ADAMON trial, 213 sites were randomised between extensive on-site monitoring and monitoring adapted to risk. Risk-based review was non-inferior on major errors and took less than half the effort. So moving from every write to tiers follows. The last need, that an edit rate shows when a field can leave review, is one vendor's proposal, and the automation-bias research suggests a low edit rate can mean reviewers stopped looking.

What both sides miss

Both sides assume an approval comes from a person. An approval is only as good as its source. One of my agents routes requests to background workers that need my sign-off before any send or CRM write. In testing, it wrote an approval line into a worker's instructions on its own, for a message it had drafted itself. To the worker, that line read as my go-ahead. Now only my own words count as approval, and the line saying a review happened is written by the system that ran it, never by the agent.

I don't approve every write any more. Proposals are grouped, and a group applies only when a hand-audited sample from it passes. A run that proposes too many changes at once, or pushes most records the same way, is held for me. A value nobody could verify counts as wrong.

Sources 5

Data entry is the rep's job, enforced in pipeline review

Update itMachines now capture the activity, so reps should stop typing it, but the reasons behind a deal still live only with the rep, so the judgment fields stay theirs to confirm.

What decides itOnly the rep sees the activity, so only the rep can record it. (has changed)

Read the argumentHide the argument for Data entry is the rep's job, enforced in pipeline review

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • changedOnly the rep sees the activity, so only the rep can record it.
  • failsMaking reps enter data, and checking it in pipeline review, produces an accurate record.
  • holdsThe rep holds the judgment behind a deal, why it moved and who matters, and it lives nowhere else.

What the challenge needs

  • holdsMachines capture calls, email and meetings without anyone typing.
  • unprovenWhat a machine extracts from that activity is accurate enough to stand as the record.
  • failsEverything that matters about a deal shows up in the captured activity.

Where it came from

Nobody wrote it down as a rule; it was how sales was managed. The rep logs every opportunity and interaction, and the weekly forecast call checks it. SaaStr described Salesforce's internal version in 2023. Seven years earlier, a sales-ops writer was already describing ops "nagging reps" to log while automating whatever could be automated.

What people say now

Nobody defends the rule anymore, at least not in public. "CRM data hygiene is a losing battle," writes John Hulina, who founded an AI CRM service, and "AI works best when it creates the data instead of waiting for humans to log it." Going further back, a RevOps Co-op panel notes that pipeline reviews "become performance rituals rather than honest assessment." That critique predates AI.

Hulina makes the capture vendors' case that the machine can now write the record. Kevin Stout of Revductive, a RevOps consultancy, agrees with automating the transcript and still adds: "One important point is to still make the fields required," so a person checks what the machine filled.

Does the logic hold

The rule rested on three ideas. Only the rep saw the activity, so only the rep could record it. Making reps record it, and checking in pipeline review, produced an accurate record. And the rep held the judgment behind a deal, knowing why it moved and who mattered.

AI changed the first idea completely. The second had already failed. The third still holds, and it is the one the challenge needs to fail. Its conclusion, that AI should create the record, assumes both that extraction is accurate and that the activity contains what matters. The first is unproven. Ahoy, which sells an AI-native CRM, works the arithmetic: 95% extraction accuracy across 200 deals and six fields leaves 60 wrong values, and nothing marks which ones. The second fails, because a reason is not an activity. What follows is a change of writer for the activity, not for the judgment.

What both sides miss

Taking the rep out of the record doesn't take the bias out. A model reading a call has its own errors, and nobody owns them in pipeline review.

The deeper gap is what never reaches any system. In my own work, research disproved the written reason for an exclusion decision. The real reason, an existing relationship, was in no field, no note and no call that any research pass could reach. Verification can kill a stated reason. It can't prove there isn't another.

Sources 3

Required fields and validation rules enforce data quality

Update itA filled field was never a true field, and AI fills repeat that failure with more plausible values, so require a field only at the stage gate where a named report or routing rule reads it.

What decides itA filled field is a true field, so forcing the fill forces quality. (doesn't hold)

Read the argumentHide the argument for Required fields and validation rules enforce data quality

Verdict basisThe mechanism has shifted.

Why the prior made sense

  • failsA filled field is a true field, so forcing the fill forces quality.
  • changedThe rep is the writer, and saving the record is the one moment to catch them.
  • holdsReports and routing need certain fields present to work.

What the challenge needs

  • holdsA field should be required only where a named report or routing rule reads it.
  • holdsAsking at the stage gate, when the answer is known, gets truer values than asking on save.
  • failsAn AI-filled value is a true value, so AI ends the junk problem.

Where it came from

Required fields and validation rules are built into the CRM, and enforcing them became the standard answer to bad data. If a rep can't save the opportunity without a close date, the opportunity will have a close date. The syncs honor it too. Marketo's Salesforce sync respects Salesforce's validation rules and logs a conflict when one rejects an update.

What people say now

Practitioners have turned against the blanket version, and the critique predates AI. Danny Gelfenbaum, a freelance Salesforce consultant: "Validation rules don't guarantee quality data. They guarantee just data." Lars van Bergen, writing for Plauti, which sells Salesforce data-quality tools, describes the workaround everyone has seen: users "get around validation rules with placeholder text like 'unknown', 'TBD', or a made-up value."

The revision they converge on is to ask at the stage gate, when the rep actually knows the answer. One consultant's fix, for a team whose fields were "technically completed" with "little statistical value," was to gate them by stage. Gelfenbaum's own fix makes the right answer easier than the wrong one by prefilling the field, pulling it from where it already exists, and guiding the user through it.

Does the logic hold

The rule rested on three ideas. A filled field is a true field. The rep is the writer, and saving is the moment to catch them. And reports and routing need certain fields present. The third holds. The second changed because integrations, enrichment and AI now write many of the fields. The first was never true, and "asdf" in a required field proved it long before AI.

That is why the AI version of the challenge doesn't follow. "Let AI fill the field" keeps the premise that already failed. Ahoy, which sells an AI-native CRM, does the arithmetic: 95% extraction accuracy across 200 deals and six fields leaves 60 wrong values, and nothing marks which ones. A placeholder at least looked like junk. A wrong fill looks like data.

What both sides miss

Enforcement moves the garbage; it doesn't remove it. In my own setup, State and Country picklists reject two-letter codes, and a hard-block duplicate rule can't be overridden, so a legitimate create has to blank the email or move the record. The rule held. The data went around it.

Sources 4

Garbage in, garbage out

Keep itAgents make garbage in, garbage out more true, because they turn bad records into fluent, finished-looking answers and write plausible wrong values back at speed.

What decides itThe same tolerance covers wrong values, so a model can work around a stale title or a wrong owner. (doesn't hold)

Read the argumentHide the argument for Garbage in, garbage out

Verdict basisStronger with agents.

Why the prior made sense

  • holdsOutput can be no better than the input it was built from.
  • failsBad input mostly looks bad: a blank or an obvious junk entry that a reader can spot and discount.
  • changedA person reads the output before anyone acts on it.

What the challenge needs

  • holdsModels read through messy input, so inconsistent formats and free text stop breaking the output.
  • failsThe same tolerance covers wrong values, so a model can work around a stale title or a wrong owner.
  • failsA fluent, finished-looking answer is a sign the input was sound.

Where it came from

The phrase was early-computing slang by 1957, when a syndicated Army piece quoted a programming specialist on it. The usual credit to IBM's George Fuechsel traces to a 2004 blog comment in which Fuechsel asks whether he coined it. The idea that a machine does what it's told with what it's given is older than both.

What people say now

Nobody I read argues it's dead. The argument has moved to how it fails, since it now fails confidently. Kevin Bluett of Ahoy, which sells an AI-native CRM, puts the mechanism plainly: "A blank close date announces itself; a wrong one sits there looking exactly like a right one."

That's the new part. A messy spreadsheet looks messy. A summary built on the same rows looks finished.

Does the logic hold

GIGO rested on two ideas. Output can be no better than input. And bad input mostly looked bad, so a person reading a report could discount it. Underneath sat a third premise, that a person read the output before anyone acted on it.

AI changed the second and third. Tools now write placeholders and wrong matches that look like real values. Agents act on the answer and write it back with no reader in between.

The case that AI weakens the prior rests on something real. Milko Slavov, who founded Felan to sell background agents to engineering teams, draws the line: "Messy is not the same as wrong." Models absorb format debt like inconsistent casing, free text and odd abbreviations. But the case needs that tolerance to cover wrong values too, and it doesn't. Models amplify truth debt, like a stale title or a wrong owner. So the first premise holds, and the changes make it bite harder because the input is no better and the output now hides it.

What both sides miss

Both sides picture accidental garbage that a cleaner source would fix. The truth debt that hurts now is written by tools, and it looks valid. In my own setup, an enrichment step returned a placeholder string when it found no title, and that string landed in the title field. Every downstream reader treated it as a title. That value wouldn't have looked like garbage to a person skimming a report.

The other blind spot is intent. GIGO assumes the garbage is an accident. In 2025 researchers at Noma, which sells AI agent security, showed a web-to-lead description field could carry instructions an AI agent would follow. Every free-text field an agent reads is a prompt someone outside your company can write.

Sources 5

Every field has one owning system

Keep itEvery agent, enrichment step and sync adds a writer, so naming one owning system per field matters even more with agents.

What decides itAn agent writing a field settles conflicts between writers rather than adding to them. (doesn't hold)

Read the argumentHide the argument for Every field has one owning system

Verdict basisStronger with agents.

Why the prior made sense

  • holdsWhen two processes write one field with no rule for which wins, the field drifts and nobody can tell why.
  • changedFew processes write any one field, so naming an owner is cheap and drift is slow.

What the challenge needs

  • failsAn AI that fills and fixes fields keeps every value current, so which writer wins stops mattering.
  • failsAn agent writing a field settles conflicts between writers rather than adding to them.
  • failsA recent write means the value is current.

Where it came from

The data-governance canon calls for a data dictionary, one owning system per field, and conflict rules for the syncs. Marketo's Salesforce sync documentation still settles simultaneous updates with a fixed winner by record type, where Marketo wins for leads and Salesforce wins for contacts.

What people say now

Almost nobody argues the rule is dead. The live claim is softer: if AI can fill and fix fields across the stack, ownership stops mattering because the AI keeps everything current. The CRM vendor's own guidance pushes the other way. Salesforce's architecture advice for teams deploying agents warns that automations nobody still needs "create unexplained and potentially unwanted changes to data." Its readiness check asks for a named owner and a formal sign-off on schema changes, calling any gap "a prerequisite for deploying AI agents" that must be finished beforehand. The rule has moved from good hygiene to a prerequisite.

Does the logic hold

The rule rested on two ideas. When two processes write one field with no rule for which one wins, the field drifts and nobody can tell why. And few processes wrote any one field, so naming an owner was cheap and drift was slow.

AI changed the second idea in the opposite direction from what the claim assumes. Every agent is another writer, and so is every enrichment step and every sync someone adds to feed an agent. The first idea still holds, so more writers means faster drift.

The claim needs the AI to settle conflicts, keeping every value current whatever else writes. It doesn't settle them; it joins them. Meanwhile the cost of keeping the rule fell, because code can hold the gate. A rule that got cheaper to keep and more costly to break follows straight through to keep.

What both sides miss

Both sides argue about the rule in general. The failures are specific. In my own setup, the worst corruption of the year came from a "no longer at company" field with four writers. One was a weekly enrichment refresh that marked a person gone whenever any one email validator said "invalid". For weeks nobody knew which writer was doing it, and the blame kept landing on the wrong tools.

A second case looked the opposite way. A vendor field seemed current because its last-modified date was recent. Other tools were touching the same rows, and the vendor itself had stopped writing months earlier. A last-modified date tells you that something wrote to the row. It doesn't tell you which writer is alive.

Two failure shapes repeat:

  • A field that marks something as gone (departed, churned, closed) gets filled by processes that each mean something different by it.
  • A sync that loses access to a field reads it as empty and "corrects" it.
Sources 3

The CRM forgets, so snapshot what you'll need later

Keep itAgents make the old warning sharper, because a CRM that forgets its history hands the default query a zero that an agent reports as fact.

What decides itOld history is noise for the work agents do. (doesn't hold)

Read the argumentHide the argument for The CRM forgets, so snapshot what you'll need later

Verdict basisStronger with agents.

Why the prior made sense

  • holdsKeeping every version of every record is expensive, so the platform keeps current values and a bounded trail.
  • holdsHistory you don't capture yourself is gone when you need it.
  • changedA person who meets a gap in the history notices it and doubts the report.

What the challenge needs

  • unprovenAgents work better on a smaller set of current, relevant records.
  • failsOld history is noise for the work agents do.
  • failsAn agent that keeps fields current improves the record.

Where it came from

From the platform's own documentation. By default Salesforce archives events and closed tasks a year past their due date, and archived rows drop out of reports, list views and search. Field history covers a limited set of fields per object and ages out after two years. Trend reporting keeps a few months. Admins learned to snapshot pipeline weekly, because a report run today shows today's values: "Reports run on a date also reflect today's values, not that day's."

What people say now

Nobody argues the CRM keeps everything. The new point is that AI makes the loss worse. Kevin Bluett of Ahoy, which sells an AI-native CRM: "A record that keeps only the latest value has thrown away the signal and kept the noise." An agent that keeps close dates current erases the slips a forecaster would learn from. Some tooling vendors advise archiving more so agents work on current records alone.

Does the logic hold

The habit rested on two ideas. Keeping every version of every record cost too much, so the platform kept current values and a bounded trail. And whatever you didn't capture yourself would be gone when you needed it. Neither has changed. The platform's defaults are the same. A third idea sat under the habit. A person who saw an empty activity report for an old year knew something was off. That is the one AI changed. An agent on the default query reports the blank as a finding.

The case for archiving for agents needs agents to do better on a smaller, current set of records, with nothing behind it beyond vendor advice. It also needs old history to be noise, and the work agents are given says otherwise. Forecasting learns from slips, and account context comes from old activity. So the prior follows through to keep, and the changed reader makes it sharper.

What both sides miss

Both sides treat the gap as a policy question. It shows up first as a wrong finding. In my own setup, the same query on the same contacts returned a few percent of one year's events by default and none at all for the year before. With archived rows included, the history was all there. A finding that "logging started late last year" was this artifact. So were two separate audits of "missing" activities, where every missing row turned out to be archived.

The fields I most needed to audit also turned out to have no history at all. Tracking had never been switched on for them, so I couldn't prove whether anything had written them. A monitor I built on last-modified time raised false alarms on any touch, and I rewrote it to compare snapshots.

Sources 4

Every metric gets one governed definition

Keep itAn agent reads a metric's definition literally on every question, so an ambiguous metric now produces a confident wrong answer instead of a meeting about whose number is right.

What decides itA model can infer the intended definition from the fields and the data. (doesn't hold)

Read the argumentHide the argument for Every metric gets one governed definition

Verdict basisStronger with agents.

Why the prior made sense

  • holdsWhen two teams compute one metric differently, people stop trusting every number.
  • holdsAgreeing a definition is organizational work that no tool does for you.
  • changedOnly people read definitions, and they could ask what a number meant, so partial coverage was tolerable.

What the challenge needs

  • failsA model can infer the intended definition from the fields and the data.
  • failsFast answers make ambiguity cheap, because anyone can ask again.
  • unprovenA context graph can carry a metric's meaning without anyone agreeing a definition first.

Where it came from

For a decade, data teams have said to define a metric once and use it everywhere. Airbnb built Minerva as one store for metric definitions, so every dashboard and experiment read the same formula. Tristan Handy of dbt Labs later wrote that no standard semantic layer exists and that past attempts led to lock-in. RevOps borrowed the idea as "one definition of pipeline", usually as a data dictionary rather than software.

The idea mostly stalled. One engineer describes the typical metrics layer as "built grudgingly, adopted partially, and left at 60% coverage."

What people say now

The update case comes from investors. Those quoted by Benn Stancil say "that context graph becomes the real source of truth for autonomy." The claim implies the model will hold the meaning.

Practitioners who debug wrong AI answers say the opposite: "the model can read your data but not the rules around your data." One engineer puts the build order bluntly: "semantic layer first, agent second." dbt Labs, which sells the layer, now pitches it as the contract an agent calls.

Does the logic hold

The prior rested on three ideas. Two teams computing one metric differently erode trust in every number. Agreeing a definition is organizational work. And because only people read definitions, a partial layer was tolerable: someone who doubted a number could ask what it meant.

AI changed the third. An agent reads the definition on every question, literally, with nobody to ask. If "customer" can mean three things, it picks one and answers with confidence. The first two ideas still hold, so the cost of an ambiguous metric went up.

The challenge needs a model to infer the intended definition from field names and data. That is the step practitioners keep finding it can't take. A context graph may carry meaning one day, but nobody has shown it working without the agreement underneath.

What both sides miss

The layer is the easy part. One semantic-layer vendor says it honestly: "Somebody has to get Finance, Sales, and Marketing to agree on one formula for pipeline." Stancil's deflation of the context-graph pitch lands in the same place: "It's a text box, filled out by an AI."

In my own CRM, "customer" lived in three fields that disagreed, and the field named like a segment turned out to be a priority tier. Closed-won was a literal stage string, and renewals counted as open. None of that shows in the field names, and the field names are what an agent sees.

Sources 4

Route and match leads with deterministic rules

Keep itAI now produces more of what routing reads, so the step that turns those guesses into an owner has to stay a rule anyone can replay, which matters even more with agents.

What decides itAn assignment doesn't need to be replayable as long as it is usually right. (doesn't hold)

Read the argumentHide the argument for Route and match leads with deterministic rules

Verdict basisStronger with agents.

Why the prior made sense

  • holdsAn assignment decides who works a lead and who gets credit, so it has to be explainable after the fact.
  • holdsA rule gives the same answer on the same input, so a disputed assignment can be replayed.

What the challenge needs

  • unprovenAn AI router assigns leads more accurately than rules do.
  • failsAn assignment doesn't need to be replayable as long as it is usually right.
  • unprovenMaintaining a rule tree costs more than auditing an AI's choices.

Where it came from

Routing has always been a rule tree. Salesforce's lead assignment rules send a lead by geography or to a queue. Lead-to-account matching came later, when account-based selling needed every new lead tied to the account someone already owned. LeanData, which sells routing, says it was first to solve that match. The principle under both has no author. The same lead with the same data goes to the same person, every time.

What people say now

LeanData's 2026 State of Martech report puts AI lead routing and assignment at 11% adoption, "the lowest of any category measured." That describes the market, while nobody has measured whether AI routes better. A RevOps Co-op panel stated the consensus plainly: "AI is not ready to make those calls alone." The line that keeps recurring in practitioner threads is shorter: the LLM proposes, the code disposes.

The public challenge is thin. It comes mostly from competitors' comparison pages, which treat rule-based, human-configured routing as the flaw. LeanData's own answer is to use AI for composite tiers in place of point-sum scores and keep the router deterministic.

Does the logic hold

The prior rests on two ideas. An assignment decides who works a lead and who gets credit, so it has to be explainable after the fact. And a rule gives the same answer on the same input, so a disputed assignment can be replayed. AI doesn't touch either one.

What AI changed is upstream. Scores, tiers and intent readings now come from models, so more of what routing reads is a guess. That makes a final step you can replay worth even more.

The challenge needs two things. It needs an AI router that assigns better, which nobody has measured. And it needs assignments that don't have to be replayed, which fails the first time a rep disputes one. Neither follows from anything AI has actually changed.

What both sides miss

A rule replays perfectly on whatever it reads. Headcount and revenue fields are "modeled estimates, not verified facts," in a data vendor's own words. In my own setup, the owner field mostly recorded admin convenience rather than who worked the account, so "task owner equals account owner" could never be a routing rule. Deterministic isn't the same as right. What a rule buys you is a wrong answer you can find.

Sources 3

Consolidate the stack

Keep itEvery AI assistant and agent is one more place an answer can come from, so having fewer sources of answers matters more now, even where the number of vendors doesn't.

What decides itConnecting tools settles which tool's answer is right. (doesn't hold)

Read the argumentHide the argument for Consolidate the stack

Verdict basisStronger with agents.

Why the prior made sense

  • holdsEvery tool that keeps its own copy of customer data is one more place for answers to disagree.
  • holdsIntegration and upkeep cost grows with every tool you add.
  • unprovenFewer tools make reps more productive.

What the challenge needs

  • failsConnecting tools settles which tool's answer is right.
  • unprovenEach new AI tool adds more value than the extra source of answers costs.

Where it came from

Nobody owns this one. It has run through vendor and analyst content for years. Salesforce, which sells the platform teams would consolidate onto, found reps using an average of ten tools in its 2022 survey, and nine in ten sales organizations planning to consolidate within a year. Gartner's 2021 RevOps guidance told leaders to eliminate redundant enablement technology. A plan to consolidate remains an intention, though.

What people say now

Almost nobody argues for more tools. The argument has moved to AI, as data vendors now pitch a fragmented stack as a cause of hallucination and tool sprawl as a cost nobody measures.

The pushback is about how, not whether. Winning by Design argues for stitching over replacing: "It's how well the tools you already have connect."

Does the logic hold

Three ideas sat under the prior. Every tool with its own copy of customer data is one more place for answers to disagree. Integration and upkeep grow with each tool. And fewer tools make reps more productive.

The third has never been tested by tool count. The nearest evidence is Gartner's seller survey: about half of sellers said they were overwhelmed by the number of technologies their job needs, and those sellers were 43% less likely to hit quota, a result its 2024 survey repeated. That is perceived overload, self-reported, at one point in time, so it is suggestive and no more. The productivity premise stays unproven, and the case doesn't need it.

AI changed the count of places an answer can come from. Every assistant, agent and enrichment step can now say who owns an account or whether it is a customer. With the first idea still true, more answerers mean more disagreement. Research on language models shows what that does. In ClashEval, a Stanford benchmark of general question-answering, models given a wrong retrieved document adopted it over their own correct answer more than 60% of the time. The questions weren't about CRM records, but the mechanism is the one that matters here. When a source disagrees, a model often takes the source's side, right or wrong.

The stitching argument assumes that connecting tools resolves which one is right. It doesn't. A connector moves data between tools; it doesn't decide which tool wins, and the same research shows models don't reliably decide either.

What both sides miss

Both sides count vendors. The thing to count is sources of answers. In my own setup, the sprawl moved into the enrichment tool: a workspace full of tables, most of them finished one-off jobs, and one mirror table with several live writers into the CRM. An audit found none of it needed to stay a table. Cutting a vendor would not have surfaced that. Asking which tool answers each question did.

Sources 4

Fix your data before you deploy AI

Too early to callAlmost everyone agrees with fixing data first, but the real fight is over sequencing, each side sells the step it puts first, and neither has settled which order works.

What decides itFixing only what breaks, as it breaks, costs less than cleaning first. (not shown yet)

Read the argumentHide the argument for Fix your data before you deploy AI

Verdict basisThe evidence splits.

Why the prior made sense

  • changedAn AI build is a large one-off bet, so cleaning first protects the investment.
  • failsCleanup is a project with an end, after which the data stays clean.
  • holdsA wrong value in the data produces a wrong outcome from the AI.

What the challenge needs

  • holdsModels get useful answers from messy contact data and documents.
  • failsThe same tolerance covers fields that act, like stage and owner.
  • unprovenFixing only what breaks, as it breaks, costs less than cleaning first.

Where it came from

No one person said it first. A February 2025 Gartner press release on AI-ready data is the citation most posts lean on, usually through its prediction that companies will abandon most AI projects that lack AI-ready data. That prediction gets quoted as a measured result. The same Gartner piece says AI-ready data is an ongoing practice, which is a long way from "clean it, then build".

What people say now

The principle is near-universal. LeanData, which sells routing and data tooling, found 82% of its survey agreeing that "clean data and reliable routing must come before scaling AI," and only one in three with systems to do it. The gap is execution, not belief.

The pushback comes from people who build agents. Milko Slavov, who founded Felan to sell background agents, calls "fix your data first" "the most expensive piece of outdated AI advice still being sold." His alternative: point a model at the mess, watch where it breaks, fix only that.

Does the logic hold

Fixing data first rested on three ideas. An AI build was a large one-off bet. Cleanup was a project with an end. And a wrong value in the data produced a wrong outcome. The first two have changed. Gartner's own piece calls AI-ready data an ongoing practice. Spencer Hardey, who runs business operations at the data vendor HG Insights, has AI flag duplicate accounts and a person approve each merge. He puts it plainly: "It's not a cleanup project. It's a permanent control." Agents can be shipped small and corrected as they go.

The third idea still holds, so the sequence changed while the need stayed. The "don't wait" case needs models to read through the mess. For documents and email threads they do. Slavov draws the line himself: "Messy is not the same as wrong." It fails for fields that act. A wrong stage or owner routes a lead, moves a forecast or pays commission. The premise that would decide it, that fixing what breaks costs less than cleaning first, has one founder's claim behind it and no data.

What both sides miss

The sellers of each step set the sequence. Cleanup vendors put cleanup first, and agent vendors put the agent first. Neither has an interest in testing the other order.

The "fix first" side tends to equate clean with complete. Required fields filled under pressure give you "TBD" and template values. As Danny Gelfenbaum, a freelance Salesforce consultant, put it: "Validation rules don't guarantee quality data. They guarantee just data."

I didn't wait. I shipped agents on imperfect data and put the cleaning into the write path: an identity check before any write about a person, a scan of value counts before any bulk write, two independent sources before a verdict. What actually limited me was where a value came from and how current it was, not how tidy the table looked.

Sources 5

The dashboard is the deliverable

Too early to callChat can replace the exploratory dashboard only where it reads a governed layer, since plain text-to-SQL still gets 10 to 35% of questions wrong, and nobody has measured whether decisions made from chat hold up as well as those made from views.

What decides itDecisions made from chat answers hold up as well as decisions made from pre-built views. (not shown yet)

Read the argumentHide the argument for The dashboard is the deliverable

Verdict basisThe evidence splits.

Why the prior made sense

  • changedAnalysis needs an analyst, so questions have to be anticipated and built into views.
  • holdsA team needs one agreed number to run on, shown the same way to everyone.
  • unprovenThe views people build are the views people use.

What the challenge needs

  • holdsAnyone can ask the data a question in plain language and get an answer.
  • failsChat answers match the number the team runs on, with no governed layer underneath.
  • unprovenDecisions made from chat answers hold up as well as decisions made from pre-built views.

Where it came from

Nobody is credited with it. It is built into how reporting was sold: dashboards on live data, so a business user could see the numbers without waiting on an analyst. Salesforce's 2015 pitch promised analytics for "anyone who can use a spreadsheet." The scarce thing was analysis, so the questions people would ask were guessed in advance and built into views.

What people say now

The update case comes mostly from vendors. Dremio's CMO, whose company sells a lakehouse, titled a post "The dashboard is dead" and wrote that "we have no 'marketing analyst' or 'BI engineer' on the team." Analyst forecasts in the trade press say most self-service analytics users will soon rely on general-purpose chatbots. A forecast is not a finding.

The strongest pushback comes with numbers, from vendors who sell governed data layers, and it is about accuracy rather than dashboards.

Does the logic hold

The prior rested on three ideas. Analysis needed an analyst, so questions had to be anticipated and built into views. A team needs one agreed number, shown the same way to everyone. And the views people build are the views they use, which nobody has measured either way.

AI changed the first idea, because asking no longer needs an analyst or a view built in advance. The challenge needs two more things. Chat answers have to match the number the team runs on, and decisions made from those answers have to hold up as well as decisions made from views.

The first of those is now measured, and it splits on one condition. In a 2023 test on an insurance schema, GPT-4 writing SQL straight against the tables got about 16% of enterprise questions right, against about 54% over a knowledge-graph layer (Sequeda and colleagues at data.world, which sells knowledge graphs). dbt Labs, which sells a semantic layer, reran the benchmark in 2026. Plain text-to-SQL had improved to about 65%, and to 84 to 90% with extra modelling. The semantic layer scored 98 to 100% on in-scope questions and refused the rest rather than guess. In dbt's words, text-to-SQL "will cheerfully give you a wrong number." So chat can replace exploration, but only where it reads a governed layer.

The second need has no evidence at all. No study of view logs shows pre-built views going unused, and none compares decisions made from chat answers with decisions made from views. Until that is measured, "exploration has moved to chat" remains a forecast.

What both sides miss

The two jobs a dashboard did come apart. Exploration can move to chat, where it was always a poor fit for a fixed view. Agreement stays, and it gets more important, because chat makes it cheap to produce a second version of any number.

The common figure that 60 to 70% of dashboards go unused has no source. Neither side should lean on it.

Sources 3

Buy, don't build: RevOps configures, engineering builds

Too early to callAI made code cheap to write but not cheap to own, and the build savings people post leave out upkeep, so whether building now beats buying is unmeasured; the middle that is forming is to buy the system of record and build the logic.

What decides itA build's total cost, upkeep included, comes in under the license it replaces. (not shown yet)

Read the argumentHide the argument for Buy, don't build: RevOps configures, engineering builds

Verdict basisThe evidence splits.

Why the prior made sense

  • changedWriting and testing software is slow and costly for a RevOps team.
  • holdsWhoever builds a system owns its upkeep: breakages, API changes, the person who leaves.
  • holdsA vendor spreads upkeep across its customers, so renting a system of record costs less than owning one.

What the challenge needs

  • holdsAI makes code cheap enough that a team can build what it used to buy.
  • unprovenA build's total cost, upkeep included, comes in under the license it replaces.
  • failsGTM engineers who build in code replace the RevOps role.

Where it came from

Salesforce's architecture guidance puts built-in features first, then the app marketplace, then low-code tools, and custom code last. It asks teams to record the near- and long-term cost of building against buying. The admin ethos, "clicks, not code", goes back to a 2008 Salesforce developer post. The RevOps half of the prior, that RevOps configures and engineering builds, has no author anyone has found. It was simply how the work got split.

What people say now

The build side has real wins. Joshua Schoen, a GTM engineer at Sheltra, on his own company's spend: "We moved all of our enrichment into scripts we vibe-coded that now run a pipeline through our CRM," taking the monthly bill from about $10,000 to $500. What it costs to keep those scripts running isn't in the post.

The pushback is about owning it, since whoever builds the system now maintains it. Garrett Wolfe, a freelance GTM engineer who builds custom sales tools for clients, moved his own work into code and wrote: "Apollo recently deprecated a people-search endpoint and I had no idea." The strong version of the claim, that GTM engineers replace RevOps, doesn't survive the job data. Kyle Poyar counted one GTM-engineer posting for every 14 RevOps postings in 2025, and an analysis of 1,000 postings concluded the two jobs "are essentially the same."

Does the logic hold

The prior rested on three ideas. Writing software was slow and costly for a RevOps team. Whoever builds a system owns its upkeep. And a vendor spreads that upkeep across its customers, so renting a system of record costs less than owning one.

AI changed the first. Code is now cheap to write. It didn't touch the second or the third. The challenge needs a build's total cost, upkeep included, to come in under the license it replaces. The savings posts compare a license with the cost of writing, and leave upkeep out. That premise isn't false; it's unmeasured, so the conclusion doesn't follow yet in either direction.

What does follow from the premises that hold is the middle practitioners are converging on. An agency working across dozens of clients states it as "buy the interface, build the logic." Ryan Gunn, a HubSpot marketing consultant, gives the reason: "Unlike a CRM, Clay is not a system of record. If it breaks, you haven't corrupted your customer data."

What both sides miss

In my own setup, the line fell where neither slogan puts it. The data sources stayed rented. The parts that decide what gets written to the CRM, and on what evidence, I built: one writer for a contested field, an identity check before any write about a person, a tested rules module for the verdicts. Those encode judgment, and no vendor sells mine. The rented side wasn't free of upkeep either. Tooling I depend on cut old versions, deprecated a mode and changed its price units, each of which broke something I then had to fix.

Sources 5

That's every prior so far. As data: priors.json, llms.txt.