For three years the metric for AI was adoption — seats deployed, pilots launched, tools rolled out. In 2026 that metric collapsed under its own bills. Only 14% of CFOs report meaningful value from AI today; a Gartner study of 350 executives at billion-dollar firms found workforce-reduction rates were equal whether ROI was high or low — no link between cutting people and getting returns.[1][2] Meanwhile the bills arrived: developers reported GitHub Copilot charges jumping from $29 to $750 a month under token pricing, and Gartner projects AI coding costs could exceed developer salaries by 2028.[3] The buyer's question changed from how much AI are you using to what did each token return — an emerging outcome scoreboard the industry is still learning to name. The per-seat software model was built for the first question. It is now being graded on the second.
The metric that measured AI for three years was adoption: how many seats, how many pilots, how many workflows touched. It was easy to grow and easy to report, and it is now dead — killed by the bills it generated and the returns it did not. In 2026 the buyer stopped asking how much AI are you using and started asking what did it return, per unit of spend.
The evidence is a wall of surveys pointing one way. An RGP survey of 200 CFOs (December 2025) found only 14% report meaningful value from AI today, even as 66% expect it within two years. Zuora and the Harris Poll (April 2026): just 28% of finance teams see a measurable financial impact. MIT's widely-cited 2025 study found 95% of enterprise generative-AI pilots delivered no measurable P&L impact.[1] Adoption was near-universal; return was not.
Gartner sharpened it into a causal finding. In a study of 350 executives at billion-dollar-plus firms, ~80% of those piloting AI reported workforce reductions — but the reduction rates were nearly equal for firms with high ROI and firms with low returns. There is no link, its analyst said, between laying people off and getting ROI from AI.[2] The layoffs that were sold as the return did not correlate with the return. Meanwhile the token meter kept running: developers reported Copilot bills jumping from $29 to $750 a month, Uber reportedly burned its entire 2026 AI budget in four months, and Gartner projects AI coding costs could surpass developer salaries by 2028.[3]
The honest complication: revenue-per-token is a lens, not yet a standard. No accounting body has canonized it; the field is converging instead on cost-per-outcome and value-per-token, and the closest named metric is Tomasz Tunguz's gross-profit-per-token.[4] And the pricing shift is not one-directional — Salesforce actually retreated from usage pricing back toward seats, with Benioff openly mocking the SaaSpocalypse narrative, arguing customers want predictability. The clean casualty is HubSpot, which fell ~19% the day it moved to outcome-based pricing for its AI agents.[5] The scoreboard is real and the per-seat model is genuinely at risk — but the winning metric is still being written, and at least one giant is betting the seat survives.
66% expect returns within two years; 14% see them now.[1] Adoption was easy to grow and easy to report. The gap between the two numbers is the space where the metric changed from how much you use to what it returned.
How the metric for AI shifted from adoption to accountability — and put the per-seat model on the clock.
a16z argues per-seat is no longer the atomic unit of software: if AI handles a share of support, companies need fewer seats, forcing a rethink of pricing toward outcomes. The Zendesk example becomes the template.[4]
The ThesisMIT: 95% of enterprise GenAI pilots deliver no measurable P&L impact. RGP's CFO survey: only 14% report meaningful value today. Adoption is near-universal; return is not. The metric starts to crack.[1]
The GapA Gartner study of 350 executives finds workforce-reduction rates nearly equal for high-ROI and low-ROI firms — no link between AI layoffs and returns. The layoffs sold as the ROI did not correlate with the ROI.[2]
No LinkHubSpot moves to outcome-based pricing for its AI agents — pay when the agent resolves a ticket — and the stock falls ~19% in a day. The market is not sure the new metric is safe to price on yet.[5]
Gartner projects AI coding costs could exceed developer salaries by 2028, with monthly bills escalating from $20–$100 toward $2,000–$5,000 per developer — and no direct relation between token consumption and productivity.[3]
The MeterThere seems to be no link between laying people off and getting ROI from AI investments. — Helen Poitevin, Distinguished VP Analyst, Gartner, May 2026
| Dimension | Evidence |
|---|---|
| Customer (D1) Origin · 86 | The lever is the buyer changing the question from adoption to return-per-spend. Only 14% of CFOs report meaningful value; 95% of pilots showed no P&L impact.[1] D1 is the origin because the scoreboard did not come from a vendor or a regulator — it came from the customer deciding that seats-deployed was no longer an answer, and demanding to know what each token returned. Every downstream pressure follows from that one shift in what the buyer will accept as proof.The New Question |
| Revenue (D2) L1 · 84 | The per-seat pricing model, built for the adoption era, is being tested against outcomes. a16z argued per-seat is no longer the atomic unit; HubSpot moved to outcome pricing and fell ~19%; Salesforce moved through per-conversation and back toward seats.[4][5][6] D2 amplifies from D1 because when the buyer's question becomes return-per-spend, the pricing model has to answer it — and the industry is visibly split on how, which is what keeps this at-risk rather than resolved.The Model Under Pressure |
| Operational (D6) L1 · 78 | The operational reality forcing the question is the token meter. Copilot bills reported jumping $29 to $750; Uber reportedly exhausted its 2026 AI budget in four months; Gartner projects coding costs exceeding developer salaries by 2028, with no link between tokens burned and output.[3] D6 amplifies alongside D2 because the bill is what made adoption untenable as a metric — you cannot celebrate usage growth when usage is the cost line that is running away from the return.The Token Bill |
| Quality (D5) L2 · 74 | The standard of proof rose from deployment to measurable value. Gartner's finding that layoffs did not correlate with ROI, and the 95% no-P&L result, redefined what counts as evidence that AI worked.[1][2] D5 is where adoption is exposed as a proxy that stopped predicting the thing it was supposed to proxy: a deployed seat used to imply value, and by 2026 it implied only a bill. The quality bar migrated from what you installed to what you can measure it returned. |
| Employee (D3) L2 · 68 | The AI-attributed layoffs were the adoption era's headline return, and the scoreboard demoted them. Gartner: reduction rates were equal across ROI tiers.[2] D3 sits here because the workforce cuts were the most visible claim that AI paid off — and once the buyer demanded return-per-spend, a layoff with no measurable ROI behind it became a cost story, not a value story. The people were cut on the promise of a return the data could not find. |
| Regulatory (D4) 62 | D4 is the longest-lag dimension: whether outcome accounting becomes a governed standard. The FinOps Foundation formalized token economics and value-per-token at FinOps X 2026, and cost-per-verified-outcome was floated as a candidate metric — explicitly a first draft.[7] D4 is slow because a metric only becomes a scoreboard once there is agreement on how to compute it; until an FinOps-style standard settles, revenue-per-token stays a lens each buyer applies its own way, which is exactly why this case is at-risk and not diagnostic.Watch — Standard |
The cascade originates in D1 — Customer — because the lever is the buyer changing the question: from adoption to return-per-unit-of-spend.[1][2] From D1 it amplifies into D2 (the pricing models under pressure — per-seat giving way to usage and outcome pricing) and D6 (the operational reality of exploding token bills) together, then D5 (the proof standard — measurable value, not deployment) and D3 (the AI layoffs that no longer count as return). D4 (whether outcome accounting becomes a governed standard) is the longest-lag dimension. Cross-references: [UC-251] is the same reckoning at the capital-markets layer — the buyer demanding return is the investor demanding return, one tier up; [UC-114] is the per-seat model's earlier warning; [UC-246] is the human-ROI counterpoint. Honest hedge, stated: revenue-per-token is an emerging lens not an adopted KPI, and Salesforce is a live counter-example, retreating toward seats even as HubSpot is punished for leaving them.
-- UC-252: The Revenue-Per-Token Scoreboard: 6D At-Risk Cascade
-- Adoption died as the KPI (cluster: UC-251/253/254/255; per-seat: UC-114)
FORAGE revenue_per_token_scoreboard
WHERE adoption_metric_dead = true
AND outcome_accountability_rising = true
AND token_bills_exploding = true
ACROSS D1, D2, D6, D5, D3, D4
DEPTH 3
SURFACE revenue_per_token_scoreboard
DIVE INTO metric_shift
WHEN buyer_asks_return_per_spend = true
AND per_seat_model_at_risk = true
TRACE adoption_to_accountability_cascade
EMIT revenue_per_token_signal
DRIFT revenue_per_token_scoreboard
METHODOLOGY 82
PERFORMANCE 43
FETCH revenue_per_token_scoreboard
THRESHOLD 1000
ON EXECUTE CHIRP high 'Only 14% of CFOs report meaningful AI value and Gartner found no link between AI layoffs and ROI while token bills explode - the buyer's question shifted from how much AI you use to what each token returned; but revenue-per-token is an emerging lens not a standard and Salesforce is retreating toward seats even as HubSpot is punished for leaving them'
SURFACE analysis AS json
Runtime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905193
Seats deployed and pilots launched go up whether or not value comes out. That is exactly why the metric died: 95% of pilots showed no P&L impact, yet adoption looked healthy the whole time.[1] A KPI that can rise while nothing is returned is a vanity number waiting to be replaced.
Gartner's finding is the quiet devastating one: firms that cut staff for AI got the same returns whether the ROI was high or low.[2] The headcount reduction that was supposed to *be* the payoff turned out to be uncorrelated with the payoff. The proof was never the layoff.
Per-seat pricing was predictable; token pricing is not. Copilot bills jumping from $29 to $750, coding costs projected to exceed salaries by 2028 — with no link between tokens burned and productivity gained.[3] The scoreboard exists because the bill made the question unavoidable.
The honest limit: revenue-per-token is a lens, not a standard, and the shift is contested. HubSpot was punished for going outcome-based; Salesforce retreated toward seats and mocked the death-of-SaaS narrative.[5][6] The per-seat model is at risk — but its replacement has not been ratified, and one giant is betting the seat survives.
Seven sources: the CFO-ROI survey wall (RGP, Zuora/Harris, MIT), Gartner's no-link-to-ROI study, the token-cost blowups (Copilot billing, Gartner's coding-cost forecast), the pricing-model thesis (a16z), and — for the honest counter — Salesforce retreating toward seats while HubSpot is punished for leaving them.
Stop counting seats and pilots. Start pricing what each token returned. The buyer already has.