The short answer
Outcome-based pricing did not end acceptance disputes. It moved them into the rate card: if "what counts as an outcome" was never negotiated before you signed, you handed the right to define your results across the table yourself. "Outcome" is a three-layer word sold as one — the resolution the contract defines, the resolution the customer privately feels, and a business gain that is real, measurable, and traceably attributed are three different things. In products actually being sold right now, "the customer stopped replying" is enough to book a billable resolution. So before you sign an outcome-priced contract, work all five acceptance questions to the end — settle them up front, don't litigate them after the bill.
The problem this piece solves: before you sign, nothing obliges the vendor to pin down what counts as an outcome — the five questions below are that acceptance checklist. Scope, stated before anything else: this piece is for buyers about to sign — or already inside — an outcome-priced AI contract. It is not a case against outcome pricing itself. If the five questions will not close, fall back to choosing your pricing model by attribution and autonomy, and seats or a flat fee may well be the right call. Every market number below is a guest's estimate or prediction, not a statistic. And the billing definitions in question have a change history — on the day you sign, re-verify the official terms against that day's version.
The clause she signed, read in full only when the bill came
The quarterly bill landed, and the head of customer support pulled the contract to the billing terms and read them three times. She had signed the thing herself; this was the first time she had read it word for word.
The boss signed six months earlier. The math looked easy: a support team of some twenty people, a few million a year in payroll; swap in an AI agent that charges 0.99 dollars per "resolution," nothing when it fails to resolve. Nobody at the signing read the terms closely — on paper the vendor took the risk first. Who proofreads that?
Fourteen thousand "resolutions." She sampled a few dozen, and a huge share looked identical: the customer received the AI's reply and never spoke again.
Maybe the problem was solved. Maybe they never opened the reply. Maybe they just couldn't be bothered to argue with a machine. She went to the vendor to reconcile, and the vendor turned to the line that settled it: a customer's confirmation of resolution, or no further request for help, each counts as one billable outcome.
She wanted to dispute. The contract has no appeals channel, and a manual review of a 99-cent charge costs more than the amount in dispute. The quieter cut came after that: from next quarter on, the quality reports she hands her boss have to be filled in on the vendor's terms. Her own scorecard is gone.
I have hit this fight more than once doing AI implementations — different versions, same skeleton. She did not lose an argument. She signed away the right to define "resolved" the day she signed.
"Outcome" is a three-layer word sold as one
This model keeps getting praised as the most honest business of the AI era: no seats, no monthly fees — pay only when it works.
But "outcome" is a big word with three layers packed inside. The resolution as the contract defines it is one layer. Whether the customer feels resolved is another. Whether the business actually benefited, with the gain traceable to whoever produced it, is a third. The three are not the same thing. Selling them as if they were — that is the business model.
This is not a thought experiment. On October 7, 2026, I checked Intercom's official pricing page: Fin charges 0.99 dollars per outcome, and three situations each count as a billable one:
| Billing situation | The definition |
|---|---|
| Customer confirms resolution | The user says the problem is solved |
| No further request after the reply | The help docs are blunter: users who leave without asking again are treated as presumed resolved |
| A configured procedure runs to completion | Including a handoff to a human — that also counts |
One charge per session.
You thought you were buying "pay when it works." You actually bought a cash register that can file "nobody replied" under "done" — because whoever defines the outcome is the same party collecting for it.
For the vendor it is a business model: the cost of acceptance gets pushed outward and becomes waiting inside the customer's silence. For the buyer it is a transfer of managerial control. Acceptance has always been part of management — how it gets written into an outsourced contract, who has the final word on performance targets — argued across tables for decades. Now there is a counterparty that doesn't argue at all. The clause argued for it.
Outcome pricing did not eliminate acceptance disputes. It printed them onto the price list: every future "does this count" fight already has a price sticker on it.
The playbook is public: how vendors grab the pen
The remarkable part: the textbook that teaches vendors how to take this position is public.
In July 2025, Lenny's podcast featured Madhavan Ramanujam — a former managing partner at Simon-Kucher who set pricing for hundreds of companies during his years there; by the time of the recording he had left the firm and moved into full-time investing.
He offered a two-by-two: one axis is autonomy — can the AI finish the work without anyone watching; the other is attribution — can the result be credited to it. The top right corner, high on both, is outcome pricing. He calls it the golden quadrant.
His warning to vendors was blunt: charge on outcomes before you can prove attribution and you will fail. Pick the model that matches where you stand today, then build the road — engineer attribution into the product, take humans out of the loop.
His positive example was coding tools: seat pricing to hybrid, edging toward outcome pricing. His reasoning was plain — it is roughly like hiring an AI engineer plus an AI tester; the agent writes the code and checks its own work, so the results are both attributable and autonomous. Customer support has no such luxury: no compiler can check whether a "resolution" happened; it exists only in somebody's telling. That comparison is mine, not his.
Keep listening and it is all moves to claim the position. One of his favorite moves is the value audit: the vendor concedes a discount; in exchange, the customer's internal team runs a value assessment every six months. The elegance of the mechanism: the assessment is done by the customer's own people, the business case is the customer's own, and from then on every renewal negotiation has someone inside the customer making the vendor's argument.
Then there is the ROI model. His position: align on inputs before the pilot even starts — how long the process takes today, how many people it pulls in. Once the customer has accepted the inputs, disputing the outputs gets hard. Produce an ROI from scratch after the pilot and, in his telling, that battle is already lost.
Buyers should read that page in reverse: every move on it is a fight over who signs off on "what counts as an outcome." The value audit is co-produced; the input assumptions are jointly confirmed. Whoever is in the room, their definitions become the contract's. If you are not at the table, you have handed the pen across it.
Definitions move: what you signed is one version of many
How big is this play right now? The same guest's estimates: about 5 percent of companies truly charge on outcomes; the best of them capture 25 to 50 percent of the value they create — traditional software does well to capture 10 to 20. Within three years, he predicts, that 5 percent becomes 25. His prediction, not a statistic.
Does the pricing world itself admit the attribution problem? In late 2025, the annual predictions from Ibbaka, a pricing consultancy, said it outright: when many actions, many people, and many agents jointly produce one result, there is no industry consensus on how to attribute and price it — and pricing experts rarely admit that publicly, because no solution is in sight.
Harder still: the definition itself moves. Fin is treated as outcome pricing's flagship, and in roughly fifteen months its billing definition changed three times:
| When | The definition | Source |
|---|---|---|
| July 2025 | Charged only for fully self-served resolution; no charge once a human stepped in | The guest's account on that episode |
| March 2026 | Handoff to a human after a configured procedure became billable | Intercom's own change note |
| October 7, 2026 | The three situations above; one charge per session | Official pricing page, checked for this piece |
One more disclosure while we are here: Fin sponsored that episode — the host says so on air. It changes nothing about the mechanism; it should put an extra string on your ear when you listen to the interview.
The version of the terms you signed is one version of many.
Before you sign: settle the five acceptance questions
So if you are actually signing an outcome-priced contract, negotiate all five questions before the pen moves — settle them up front, don't litigate them after the bill. One thing first: do not hand the whole table to legal. Whether the input assumptions in the terms hold is known only to the people who run the process. The support lead from the opening story should have been in that signing meeting.
Question one: the acceptance event
What action counts as "the outcome happened"? Does customer silence count? It goes into the contract body — not waved off as "standard industry practice." Ask it straight across the table: your pricing page lists three billable situations, so what about the fourth — the customer switches channel, or switches to a different person? Which bucket does that fall into?
Question two: input assumptions
An outcome only stands on premises — data quality, customer cooperation, a process that stays unchanged. Who guarantees them? And when a premise changes, does the definition of the outcome move with it? Force it into a sentence you can sign: we maintain the knowledge base, and both sides confirm the maintenance standard. When the knowledge base causes a wrong answer, is that billable?
Question three: exceptions and appeals
What does not count as an outcome? How does review work? Who adjudicates? A billing clause with no appeals channel hands one party the rulebook, the score, and the whistle — all three. The minimum, not the ceiling: a right to sampled review — fifty disputed "resolutions" a quarter, each side scoring them independently; divergence above ten percent reopens the definition.
Question four: attribution scope
Over the same stretch you are also reworking your processes and plugging in other vendors' tools. What entitles the vendor to book the entire lift? And the mirror question: how do we carve out the part it did not cause? It is not an elegant sentence. It is more elegant than the same fight after the bill.
Question five: review cadence
How often do we reconcile? Against which version of the terms? When the terms change, is there a duty to notify you first? That flagship vendor from earlier widened its billing definitions a little over half a year ago — the version you sign is not the final one. This is just the start.
All five on one card — run through it before the signing meeting:
| The five questions | What to ask | The sentence to get into the contract |
|---|---|---|
| 1. Acceptance event | What action counts as "the outcome happened"? Does customer silence count? | Into the contract body. And the fourth situation — the customer switches channel or person — which bucket? |
| 2. Input assumptions | Data, cooperation, process stability — who guarantees the premises? | When the knowledge base causes a wrong answer, is that billable? |
| 3. Exceptions and appeals | What does not count as an outcome? How reviewed? Who decides? | Fifty disputed "resolutions" sampled a quarter, each side scoring independently; divergence above ten percent reopens the terms |
| 4. Attribution scope | Changes you made over the same period — the lift they added: whose is it? How do we carve out what it did not cause? | Not an elegant sentence. More elegant than the bill. |
| 5. Review cadence | How often do we reconcile? Against which version? Notice before terms change? | The version you signed is not the final one |
The five questions are not a game of winning them all. Whichever one will not close is where the real cost of this deal lives — the clause the vendor refuses to put in writing is usually the one it plans to earn on through ambiguity.
When the five questions do not apply
Under what conditions does this set of five fall short? Said up front.
First: failing to reach agreement is not failure. Fall back to the guest's own two-by-two and choose by attribution and autonomy — pick the model that matches where you stand today. Seats and flat fees can be perfectly reasonable. Outcome pricing is a direction, not an entry requirement.
Second: every figure in this piece carries a timestamp — the interview is July 2025, the change note March 2026, the pricing page checked October 7, 2026. On the day you sign, re-verify the official terms against that day's version. That is the fifth question, applied to itself.
Third: the opening argument is a composite reconstructed from cases I have handled — field experience, not any single company's billing data, so don't try to match it to a real invoice. The guest's 5 percent, the 25 to 50 percent, and the 25-percent-in-three-years are all estimates and predictions. Treat them as statistics and the failure is mine, not his.
Already signed? The five questions still work
Back to the support lead. The contract is already signed; the five questions still apply — the fifth can start today: from next quarter on, which version of the terms we reconcile against each quarter, and what channel disputes run through. Get both into a supplementary agreement, in writing.
Outcome-based pricing is not a bad business. What is bad is leaving acceptance to be heard for the first time on the day the bill arrives.
I spent eighteen years in HR taking apart roles and responsibilities, and now I run a one-person AI organization — and the conviction I trust most came out of that first half: to know who owns a thing, read whose name the responsibility is written under. Acceptance has always been part of management. The five questions are not a tongue-twister for legal; they are how the responsibility for "what counts as an outcome" gets signed back under your own name. Next time a vendor pushes the contract across the table, pull up one extra chair — for the person who actually runs the process — work the five questions to the end, and only then pick up the pen.
Related reading
- AI Replaces Actions, Not Organizations: AI can take over the action; it cannot take over the accountability — acceptance, in the end, lands on a named person.
- The New Labor Contract: another template for writing accountability into a contract — this time, into a procurement contract.