The Outcome-Based Pricing Playbook

The opening scene

A B2B vendor signs a Band 4 outcome pricing contract in Q4 2025. The deal is clean on paper.

The vendor charges per “successful outcome.” The customer’s procurement team approves it because outcome pricing reads as fair — pay only when value is delivered. Sales books the win. The PM moves on.

By Q2 2026 every renewal cycle has a legal review attached. The contract says “outcomes.” It never defines, operationally, what one is. Customer A counts an outcome when their support ticket closes. Customer B counts an outcome when their CSAT score rises above a threshold. Customer C counts an outcome only if their internal analyst signs off in a quarterly review. Three customers, three counting methods, three different invoices the vendor’s finance team can’t reconcile.

By June 2026 the math is in. The pricing model didn’t fail. The dollar amount was right. The unit economics worked. What failed was eight or nine missing words — the operational definition of “one outcome” that should have lived in the contract from day one. Those eight or nine words turned a $14M ARR product line into a $2M ARR product line plus a $400K legal-services line item. The vendor’s CFO is now in every renewal conversation. The PM who signed the original contract is no longer the PM on this product.

This is the standard failure mode of outcome-based pricing in 2026. Not the price. The definition.

The frame

Most teams reaching for outcome pricing think the hard work is in the number. Is it $0.99 per outcome or $1.50? Is the margin sustainable at scale? Can the customer’s procurement team approve it? Those are the easy questions. The hard question is the one almost no team writes down: what, precisely, counts as one outcome?

Intercom Fin’s $0.99 per outcome works because of what comes after the dollar figure: “billed when conversation is resolved or has no follow-up within a defined period.” Zendesk’s automated resolution model works because the count is “after 72 hours of inactivity, after the AI agent provided a generative reply based on the user’s last message.” Both definitions are concrete enough to hold up in a legal dispute. Both are defensible. Both can be audited from system logs.

A team that writes “$0.99 per resolved support ticket” without defining what “resolved” means has shipped a slogan, not a pricing model. A team that copies the Zendesk number ($1.50 per automated resolution) without copying the 72-hour inactivity rule has imported the price and dropped the structural component that makes the price work.

The playbook is the discipline of writing the eight or nine words.

The 60-second answer

The outcome-based pricing playbook is eight operational layers. Every outcome contract has to lock all eight before it ships. Skip one and the contract fails on first contact with a customer who has a different read. Only 5–17% of AI vendors successfully ship outcome pricing — the rest master 3–4 layers and roll back at week 12.

• L1 · Outcome Definition — the 8–9 words. Discrete, measurable, attributable, auditable.
• L2 · Attribution Infrastructure — trace logging, decision-point tags, attribution-as-code. Solves the multi-touch problem.
• L3 · OMA Contract Structure — outcome-measurement agreement: definitions, methodology, tiers, dispute path, refunds.
• L4 · Sales Motion — reps trained on outcome modelling; comp on outcomes-driven ARR, not event volume.
• L5 · Finance Reconciliation — ASC 606-compliant variable-revenue recognition, customer-level reporting, forecast variance.
• L6 · Customer Communication — pricing calculator, real-time dashboard, sample bills. Kills the black-box rejection.
• L7 · Dispute Resolution — auditable logs, auto-credit on disputed counts, named arbitrator.
• L8 · Continuous Calibration — quarterly per-outcome rate reset as resolution rates climb. Preserves unit economics.

Most teams ship with three or four of these. The 5–17% ship with all eight.

The operational definition test

Before writing the eight or nine words, the PM has to be able to answer four questions about the proposed outcome. If the answer to any of them is “no” or “it depends,” there is no outcome pricing. There is a slogan that will collapse in renewal.

Discrete

The outcome is a single, countable event. Not a state. Not a feeling. Not a period of time. “A resolved conversation” is discrete. “A satisfied customer” is not. The PM must be able to point at a row in a database and say “that one — that’s an outcome.”

Measurable

The outcome can be detected automatically by a system, with no human in the loop required to make the judgement call. If a person has to read a transcript and decide whether something was “really resolved,” the model collapses at scale. The detection has to be deterministic from logged data.

Attributable

The outcome can be tied causally to the AI’s action, not to the user’s effort, not to the human agent’s intervention, not to the natural flow of the workflow. This is the hardest property and the one that requires explicit window definition. Without an attribution window, a customer can ship a deflection campaign in March, see the ticket close in June, and dispute that the AI did anything.

Auditable

The evidence that the outcome occurred is logged, timestamped, retrievable, and survives long enough to defend a dispute six months later. If the system overwrites the log after 30 days and the customer disputes a count in month four, there is no audit trail. The dispute is unresolvable.

A useful PM exercise: write the proposed outcome on a notecard. Hold it against the four properties. If you can’t answer “yes” to all four with concrete evidence, the outcome is not yet defined. Most teams stop at “discrete” and “measurable” because those are easy. The real work is in “attributable” and “auditable” — and that work takes weeks, not days.

The verified anchors

Three real models, all in production, all surviving contact with the renewal cycle. Each one is worth studying because each one solves a different version of the operational definition problem.

Zendesk: the 72-hour rule

Zendesk’s official definition of an Automated Resolution (AR) is the gold standard for compactness. The count fires when the AI agent provides a generative reply based on the end-user’s last message, and the conversation has 72 hours of inactivity afterwards. That’s the eight-clause definition that has held the model together since launch.

Look at what those words do. “Generative reply” rules out canned responses. “Based on the end-user’s last message” rules out unsolicited outreach. “72 hours of inactivity” creates the discrete trigger event — the audit log can detect it, the customer can verify it, and the dispute path is bounded. If the customer comes back on hour 70, no AR is counted. If they come back on hour 73, the count holds. There is no judgement call.

Zendesk’s per-AR pricing sits at $1.50 committed and $2.00 pay-as-you-go (the structure detailed in L2-T14). The price is not the moat. The 72-hour rule is the moat. A competitor can match the dollar figure tomorrow. Replicating the operational definition — and the audit infrastructure that defends it — takes a year.

Intercom Fin: the resolution-or-no-follow-up rule

Intercom Fin’s $0.99 per outcome is billed when the conversation is resolved or has no follow-up within a defined period. The definition is structurally similar to Zendesk’s but tuned for Intercom’s product surface.

The result, per the Growth Unhinged analysis and the GTM Now interview with Intercom’s president: tens of millions in ARR, 80%+ of support volume now flowing through Fin, and a 67%+ resolution rate. The model isn’t working because $0.99 is the right price. The model is working because the definition is concrete enough that customers can predict their bill, and Intercom can defend every count.

The PM lesson: when the customer can independently verify the count from their own system logs, the dispute rate collapses. When they can’t, every invoice becomes a negotiation.

Metronome’s 2026 catalog: hybrid is winning, pure outcome is rare

Metronome’s April 2026 report cataloguing 50+ AI pricing models is the most useful market-wide look at what’s actually shipping. The directional finding: hybrid pricing models — typically a base subscription plus a usage or outcome component — are now the dominant pattern. Pure outcome pricing remains rare.

The reason is the one this post keeps returning to. Pure outcome pricing requires the full operational infrastructure: the unit definition, the trigger, the attribution window, the dispute path, the audit trail. Hybrid pricing lets vendors collect a predictable base while they’re still figuring out the outcome side. Most vendors don’t have the infrastructure to ship pure outcome on day one — so they ship hybrid, build the infrastructure on revenue from the base, and migrate.

This is not a failure mode. It’s a sequencing decision. The PM who tries to ship pure outcome pricing without the infrastructure is the PM who creates the legal-services line item. The PM who ships hybrid first and graduates to outcome as the infrastructure matures is the PM whose model holds.

The attribution window question

The single most under-specified component of every outcome contract is the attribution window. How long after the AI’s action does the outcome have to materialise for the AI to get credit? There is no industry standard yet. The PM has to specify per workflow.

The closest analogue is ad platforms. Meta uses a 7-day click attribution window. Google uses a 30-day click attribution window. These are not arbitrary. They were calibrated over a decade of A/B testing on what made advertisers feel the attribution was fair while keeping the platforms’ incentives clean.

For AI outcomes, the calibration is per use case. A customer-support AI agent is operating in seconds and minutes — a 72-hour inactivity window is reasonable. A sales-development AI is operating in weeks — a 14-day or 30-day window is more honest. A medical-coding AI processing claims is operating in months — a 90-day window may be required to capture the actual revenue cycle outcome.

The trap is to copy the customer-support window into the medical-coding contract. The PM who does that has built a contract where the AI never gets credit for outcomes that take longer than 72 hours to materialise — and 80% of medical-coding outcomes take longer than 72 hours to materialise. The model collapses on the first invoice.

The fix is one sentence in the playbook: the attribution window is set by the natural cycle of the workflow, not by convenience or by precedent from a different domain.

The CSAT trap

A pattern visible in pricing decks across 2025: vendors tying outcome pricing directly to a customer-satisfaction score. The pitch sounds clean — “we only get paid when the customer says they’re satisfied.” Use the standard top-two-box CSAT methodology (4-5 on a 1-5 scale), bill on every “satisfied” response.

The model has a structural problem. CSAT is a measurement tool, not a billing trigger. The moment you tie revenue to a survey response, the vendor’s optimisation function shifts. Instead of optimising for the actual outcome, the vendor optimises for the survey itself. Survey timing gets manipulated. Survey wording gets gamed. Customers who are likely to give a low score get filtered out of the survey panel. Within two quarters the CSAT-tied outcome pricing produces inflated CSAT scores and degraded actual outcomes. The customer notices. The contract gets renegotiated. The pricing model is dead.

Smart deployments separate the two. CSAT remains the product feedback mechanism — used internally to detect quality problems, used externally to report on customer health. Billing runs on a separate, defensible operational definition: the 72-hour inactivity rule, the conversation-resolved trigger, the workflow-completion event. The two systems inform each other but they don’t share a billing wire.

The principle: the billing trigger has to be invariant to vendor incentive distortion. If the vendor can game the trigger to inflate their revenue without delivering more value, the trigger is wrong.

The cost-revenue lag — the most underestimated trap

The vendor incurs the inference cost in January. The customer captures the outcome in May. The vendor invoices in May. Cash hits the vendor’s account in June. The vendor has been cash-flow underwater on this customer for five months.

This is the most underestimated trap in outcome pricing and the one that breaks the smaller vendors first. A startup with 18 months of runway cannot afford to be cash-flow underwater on its largest contracts for five months. The model is fine in the unit economics — over a 12-month period the customer is profitable. The model is fatal in the cash flow — the vendor runs out of cash before the cumulative outcomes catch up to the cumulative inference costs.

The four cost layers from L1-T08 (training cost, inference cost, evaluation cost, drift-management cost) hit the P&L on their own schedule. The outcome revenue hits on its own schedule. When those two schedules diverge by a quarter or more, the vendor needs financial mechanics in the contract to bridge the gap.

The fix lives in three contract levers:

• Advance payments. A negotiated upfront commitment paid at signing or quarterly, drawn down as outcomes materialise. The customer is buying capacity, not paying retroactively.
• Floor minimums. A monthly or quarterly minimum payment that holds whether or not outcomes are achieved. Below the floor, the vendor is protected. Above the floor, the customer pays per outcome on top.
• Hybrid components. A base subscription that smooths the inference-cost curve, plus an outcome layer that captures upside. This is why Metronome’s catalogue shows hybrid as dominant. It’s not a pricing-model failure of imagination — it’s a cash-flow defence.

The PM who ships pure outcome pricing without one of these three levers is the PM who is on a finance call in month four explaining why working capital is negative.

The eight-layer outcome-measurement agreement

The reusable contract structure. Write it once. Lock it as the standard. Refuse to sign an outcome contract without all eight layers filled in. Use the lowercase term outcome-measurement agreement (OMA) internally — the acronym isn’t a standard industry term yet, but the structure is reusable across every outcome deal.

Component 1 is the field most teams skip. It is the field that determines whether the contract holds. Component 2 separates “Conversation status set to resolved AND no inbound message for 72 hours” (a trigger) from “Customer is happy” (not a trigger). Component 3 is set by the natural cycle of the workflow, not by precedent from another domain. Component 4 is the difference between “Disputes are resolved by mutual agreement” (not a dispute path) and “Disputes raised within 30 days of invoice are reviewed by a named third-party auditor under the terms of Schedule 3” (a dispute path). Component 5 is the system that lets the customer independently verify counts.

The playbook is to build this template once with legal and finance, lock it as the standard, and refuse to ship outcome pricing in any form without all five filled in. The PM who lets sales close before the operational definition is finalised is the PM who reopens every contract in legal six months later.

The XPO Logistics anchor

Cross-link to L1-T10 (the value model). XPO Logistics’ bottom-line attribution framework maps AI output directly to income-statement lines — $29M per single efficiency point in their operations. That is the structural cousin of outcome pricing. Internal attribution discipline rather than external billing, but the operational rigour is identical.

XPO doesn’t say “AI improved efficiency.” XPO says “AI moved this specific KPI by this specific amount, attributable to this specific intervention, measured over this specific window, audited by this specific log.” That sentence is the operational definition pattern of outcome pricing, applied internally instead of in a customer contract.

The PM lesson: the team that runs XPO’s framework internally — naming the unit, the trigger, the window, the dispute path, the audit trail for their own internal AI investments — is the team whose external outcome contracts hold. The discipline is the same. The audit infrastructure is the same. The eight-or-nine-word definitions are the same. The customer contract is just the internal framework with a price tag attached.

If you can’t run the framework on yourself, you can’t run it on customers. The internal version is the dress rehearsal.

The four traps

If the operational definition test is the diagnosis, here are the four traps to break specifically — the patterns most PMs ship by default and which will, quietly and expensively, undo their outcome pricing.

Cross-links to other posts

• L1-T08 (Cost in Every PRD) — the four cost layers that determine whether the unit economics work at the price you’re proposing. Inference cost is the dominant variable on most outcome models.
• L1-T09 (Why Per-Seat Pricing Dies) — the four-question pricing test that maps a customer to a band. Outcome pricing is Band 4 in that spectrum, and Band 4 is the right answer for a narrow set of customers, not a default.
• L1-T10 (Value Model) — Measurement × Adoption = ROI. Outcome pricing only works when both measurement and adoption are real. The XPO framework is the internal version.
• L2-T14 (AI Pricing Models) — the 5-band spectrum. This post is the operational depth on Band 4.
• L2-T17 (Stakeholder Translation) — how to present the outcome metric at the board layer. The eight-word definition has to survive the executive translation, not just the legal review.
• L3-T07 (Golden Quadrant) — outcome pricing at scale. The strategic move once the operational infrastructure is in place.