Earn the decision, then price it

1. Two problems, one boundary

In May, InsurTech's AI Margin Reckoning set out the cost side. InsurTech vendors sell inference-heavy workflows on per-policy, per-seat, per-claim or flat platform fees, while the cost of running the AI moves with use. Unlimited consumption on a fixed price leaves the vendor carrying the risk. That piece named outcome-linked pricing as an emerging answer but did not test it.

Earlier this month, Claims AI stops at the first Euro set out the operating side. The most advanced claims operations in Europe let AI take claims in and prepare the file. People still take the decisions that move consequential money. In Adacta's 2026 survey of 110 European insurance professionals, 51% automate 10% or less of claims-processing decisions, and only 8% say more than half happen without human review.

Read side by side, the two describe one problem. The pricing answer to the first only works where the second has been solved.

2. What outcome pricing needs

Here, an outcome price means the vendor is paid for a result its system produced on its own. Two things have to be true when the invoice goes out. The result is visible, against a definition both sides agreed in advance. And no person made the call.

Customer service comes closest. Intercom charges from $0.99 per Fin outcome. A resolution counts when the customer confirms the answer helped or leaves without asking for more. If the customer asks for a person, nothing is billed. If they come back to the same conversation later, the charge is reversed. Sierra works on the same logic: when its agent hands over to a human, the customer pays nothing for that interaction.

Even there, the billed event is a proxy. A customer who leaves without asking for more may simply have given up. Both sides accept the proxy because it shows within minutes and the agent owns it end to end. The buyer can check its value against handling time, which every contact centre already tracks.

The market uses the word more loosely. Air Street's State of AI Report 2026 puts pricing by task outcome one level lower, at "co-work", where professionals review the AI's work rather than redo it (slide 112). In claims, that level is "AI acts after approval". The approval is where the insurer's accountability sits, so what the vendor delivers there is a reviewed task. It can be labelled an outcome price, but the attribution problem described below comes with it.

Autonomous outcome pricing, the strict sense used in this article, starts where the system acts alone. Air Street ties it to a system that can judge the correctness of its own output. A model's confidence score does not meet that test. It is the model's view of its own work.

3. Outcome pricing is available where it is worth least

Map the claims chain against those two tests and a pattern shows. The steps where AI already acts alone produce outcomes you can see the same day. The steps that decide how much money leaves the insurer still sit with a person, and their outcome shows up months later, if at all.

Autonomy levels from Claims AI stops at the first Euro. Visibility and billing columns are Retep analysis.

Most of what a claim costs is the indemnity paid, not the expense of handling it, and a wrong settlement decision costs more than slow handling. The steps that set the settlement are the ones where the AI only recommends. A vendor can charge for outcomes in the cheap half of the process. In the expensive half, it is still selling a tool to a claims handler.

The process step is only a first cut. Claim type cuts across it: a small claim with a coverage dispute belongs in the expensive half, and a large but simple property claim can be settled on rules. Move four below deals with that.

4. The claims-mix trap

A claims AI vendor carries its highest cost per claim in the steps where it has the least pricing power. A windscreen claim needs a photo check and a policy lookup. A disputed liability claim brings a long file of medical reports, correspondence and conflicting statements, and the model reads and reasons over all of it, often more than once.

Token prices understate the gap. Artificial Analysis measures what a completed task costs, counting input, cached, reasoning and answer tokens, and finds frontier models carry the highest cost per task (State of AI Report 2026, slide 110).

A flat fee per claim hides this as long as the book stays mixed. The illustration below uses the same assumptions as the Margin Reckoning model: 75% gross margin, with AI inference at 30% of the cost of revenue.

Illustrative figures, not market data. Assumes a €20 flat fee and €3.50 of other cost of service per claim, held constant per claim in every scenario. Blended inference is €1.50 per claim, or 7.5% of revenue.

The simple claims carry the margin, and they are the easiest claims to take away. They sit in the steps where the AI acts alone, which is where a carrier can most cheaply rebuild the workflow or move it to a cheaper tool, as Your sales demo is now your competitor's spec argued.

Each simple claim that leaves shifts the vendor's book towards the claims it cannot bill by outcome.

That is the claims-mix trap. It is not inevitable: a vendor with measurably better accuracy, or data a carrier cannot reproduce, can keep the simple claims. The bridge below shows the downside case, not a forecast.

The bridge combines three separate effects, one per step. The carrier takes 40 of every 60 simple claims in-house. Procurement then uses the rebuild cost to push the fee from €20 to €17. To lift accuracy on hard claims, the vendor moves them to more capable models, which doubles their inference cost to €18.

At that point every complex claim loses money: €17 of revenue against €21.50 of cost, a gross margin of −26.5%. The blended 57% hides it. Nor does the result rest on the last step. If complex-claim cost never rises, the first two steps alone take margin from 75% to 66%.

Revenue from the customer roughly halves without the customer churning, and none of it needs a change in token prices.

The vendor is squeezed from both ends. The claims that carry its margin are the easiest for a carrier to replicate. The claims it keeps cost the most to process, and it has no outcome to bill on them. This is the exposure the Margin Reckoning described, concentrated where pricing power is weakest.

The insurer has its own version of the problem. Business cases now give inference its own budget line, while the benefit is still counted in hours freed rather than cost removed. The insurer pays a known monthly cost for a benefit it can only estimate, and negotiates accordingly.

5. Cheaper models do not fix this

The obvious objection is deflation. Epoch AI estimates that the price of a given level of AI performance falls about 47% a quarter, roughly 13 times a year (State of AI Report 2026, slide 109). On that curve, a flat price per claim should get more profitable every year.

It does only if the vendor keeps the saving. Cost per task falls more slowly than cost per token, because the hardest claims get the most capable model and the most reasoning, as the bridge assumes. The buyer sees the same curve, and cheaper inference lowers the rebuild ceiling at every renewal.

Under a flat price, the buyer collects most of the deflation. Under an outcome price, the vendor could keep a share of it, but only for outcomes it is allowed to own.

6. The buyer cannot sign what it cannot audit

Insurers are not turning down outcome pricing out of habit. For the expensive half of the claims process, they have nothing to audit it against.

An outcome price on liability or fraud needs a baseline: today's leakage on that line, measured the way the vendor will be measured. Few insurers have one. Operators rarely publish straight-through rate, decision accuracy and leakage together, and a "95%+ AI confidence" figure is none of them. The obvious claims proxy, a claim settled and not reopened within an agreed window, says nothing about whether the amount was right.

Self-assessment does not fill the gap. In a simulated banking assignment in the State of AI report, Claude Opus 5 passed all 13 of its own checks while working from a wrong financial assumption (slide 101).

Attribution is the second gap, and it is a governance problem more than a legal one. Regulation does not forbid automated claims decisions. GDPR Article 22 restricts decisions based solely on automated processing that have legal or similarly significant effects, but allows them where they are necessary for a contract, with safeguards that include the right to human intervention. EIOPA's Opinion on AI governance and risk management (August 2025) covers explainability and human oversight, in proportion to the risk.

What both ask is that the insurer can stand behind each decision it is accountable for. When a handler approves the AI's liability recommendation, the insurer stands behind the handler. Paying the vendor for an autonomous outcome is hard to reconcile with a governance model that attributes the decision to the handler. The vendor can be paid for supporting the decision, but not for making it.

A fee tied to euros not paid out raises a different problem. It gives the vendor a financial interest in lower settlements, and an insurer's claims governance will challenge that before procurement gets to price it.

Buyers will cap vendor pricing at the benefit they can evidence. In the half of the process where the evidence is thinnest, that cap is low, so the buyer prices the AI as a tool and the vendor carries the inference variance.

7. Four pricing moves

The pricing model has to follow the decision rights, step by step and claim type by claim type. That gives a vendor four moves.

1. Price the steps the system owns

Where the AI already acts alone, charge per completed unit against an acceptance test agreed in advance: a file that is complete and correctly routed, or a bounded claim settled and not reopened within an agreed window. Copy the customer service reversal rule. If the claim comes back, the fee is credited. That is a task price that behaves like an outcome price, and the buyer can audit it.

2. Stop absorbing variance where the system recommends

For complex claims, a flat fee per claim leaves the vendor holding the inference risk. Move these steps to a platform fee plus consumption bands by claim complexity. This is the first response from the Margin Reckoning, applied where the exposure actually sits.

3. Sell the measurement, then price the next step up

Build straight-through rate, override rate, independently validated decision accuracy and sampled leakage audits into the product and the contract. Override rate alone proves little: a handler can overrule a correct recommendation or wave through a wrong one. Write in an autonomy clause: when independently validated accuracy on a claim type holds above an agreed level, over an agreed sample and period, that claim type moves to delegated decisions and to a price per decision. Price the decision, not the euros saved.

The State of AI report describes the loop that makes this work: failed tasks become repeatable tests, and corrections and retries show whether a failure came from missing context, a tool or the model (slide 114). Its predictions for the next twelve months include an agent that halves its failure rate after a month of customer work, without a model upgrade (slide 238). Whoever owns that loop decides how fast the boundary moves.

That loop is also the moat. In the terms of AI Moat or AI Leak?, it is loop ownership: each delegated decision produces the evidence for the next one, and a competitor starting cold has none. It only compounds if the contract allows it. No insurer will grant unrestricted rights to its claims data, and a processor generally cannot reuse client data for its own purposes. The right worth negotiating is narrower: use of aggregated, anonymised performance data across customers, surviving termination. Without it, the vendor earns decision authority one customer at a time and starts again with every new logo.

4. Segment by claim type, not by product

Outcome pricing already works in narrow lines: pet, windscreen and simple household claims, where the decision space is small and monitored. Price those per settled claim now. Keep complex lines on platform plus consumption. Move a line across when the evidence from move 3 supports it.

8. What a board or an acquirer should ask

For a PE board, or anyone pricing a claims AI business, five questions show how much of the revenue can survive normalised inference costs.

The answers sort ARR into two kinds. Revenue priced on units the product controls keeps its margin as inference costs move, and it can be repriced as decision authority grows. Revenue behind a handler's decision carries the mix shift and the inference variance described above. An acquirer should not pay the same multiple for both.

In the illustration, one customer goes from 75% to 57% gross margin and loses about half its revenue, without churning. On the Margin Reckoning's arithmetic, nine points of gross margin took a Rule of 40 score from 50 to the low 40s. If the lost gross margin falls straight through to EBITDA, as that model assumed, eighteen points would take the same company into the low 30s.

Where the price can follow

Outcome pricing will reach the expensive half of claims when the measurement does. The InsurTech companies that pull ahead will earn ownership of claims decisions one claim type at a time, and the economics that come with them. The rest will keep selling seats to claims handlers while their book drifts towards the claims that cost most to run.