What counts as a resolution? Outcome pricing for AI agents

On 18 May 2026, Zendesk changed what its customers pay for. Not the price — the event. Under the new automated resolution tiers, an AI conversation now lands in one of three buckets: assisted escalation (free), contained resolution (free), or verified resolution (billable). Anyone who signed a per-resolution deal before that date signed against a different definition. That is the whole problem with outcome pricing, and it is why the question of what counts as a resolution in AI agent billing is a contract question, not a product one.

Outcome pricing is sold as buyer-friendly, and in principle it is. But the principle is worth nothing if the vendor writes the definition, runs the meter, holds the log and adjudicates disputes. Here is what the two vendors that publish a per-resolution price charge, how each defines the billable event, and the clause language to insist on.

What counts as a resolution in AI agent billing: the two published definitions

Zendesk and Intercom both publish a per-resolution price. Their definitions of the billable event do not match.

Zendesk charges $1.50 per automated resolution. What you actually pay depends on how you buy it: its documentation states that "committed usage provides a better per-resolution cost than overage billing (or pay-as-you-go)". Critically, "only conversations that were not escalated are evaluated for automated resolutions", and the pricing page promises you "pay only for customer requests that were successfully resolved by the AI agent, without any escalation to a human agent".

Intercom charges $0.99 per outcome for Fin, under a three-part definition: an outcome is counted when a customer confirms their issue is resolved, or they don't ask for more help after Fin responds, or Fin completes a workflow (a Procedure), including handoffs. Intercom is explicit that "you're only charged once per conversation, even if multiple questions are answered", and a minimum monthly commitment applies when running Fin on an existing helpdesk.

Read those side by side and the gap is obvious. Under Intercom's second limb, silence is a win. Under Zendesk's model an escalated conversation is not billable; under Intercom's third limb, a completed Procedure "including handoffs" is. Same word, two commercial realities. Comparing $1.50 with $0.99 is meaningless until you normalise the definitions — the first thing procurement should do.

Who decides an AI agent outcome happened — and why it is currently the vendor

Sierra's outcome-based pricing post (10 December 2024) sets out the principle and skips the question. It promises "clear, agreed-upon criteria for each outcome upfront" and notes that "if a case needs to be escalated, in most cases, there's no charge" — but never says who validates an outcome, what evidence is retained, or what happens when the parties disagree. It carries no figures, prices or sources.

Zendesk does answer it, and the answer should give any CFO pause. Under the tier model, "when the conversation ends, a verification process is performed by a large language model (LLM) that evaluates the text of the conversation to confirm that the customer's request was satisfactorily resolved". A model owned by the vendor decides whether the vendor gets paid.

That is not automatically wrong — an automated judge is more consistent than a sampling audit, and excluding escalations, test conversations and sandbox traffic from the count is genuinely buyer-favourable. But it is an unusual arrangement to accept without a right of inspection. You would not let a BPO supplier mark its own SLA report with no audit clause. MIT Project NANDA's State of AI in Business 2025 (preliminary findings; interviews with 52 organisations and survey responses from 153 senior leaders, January–June 2025) found that the buyers who actually get returns "act like BPO clients, not SaaS customers… and hold vendors accountable to business metrics". An unauditable meter is the opposite of that.

The measurement window is channel-dependent, not a single number

Zendesk's own blog describes a resolution as "often defined by a 72-hour quiet period", as if it were one rule. Its tier documentation gives three different windows: email counts 72 hours after the last email, messaging defaults to two hours after the last message (configurable up to 72), and voice ends immediately upon hangup.

That two-hour default matters commercially. A short window converts customer fatigue into billable success: the customer who gives up and returns tomorrow on another channel generates a second billable event, not a reopened first one. Fix the window per channel in the contract, and if the vendor will not net off a same-issue recontact within, say, seven days, ask why.

The clauses to insist on: outcome SLAs, decision logs and dispute credits

Mayer Brown's analysis of agentic AI contracting (17 February 2026) identifies six areas where the SaaS template fails. Four decide whether your outcome price protects you.

• Delegation of authority, stated explicitly. The contract should set out what the provider can do using AI agents "(e.g., offer an on-site service call)" and what it cannot "(e.g., offer refunds or accept liability for claims)". An outcome the agent had no authority to produce should not be billable.
• Performance warranties, not disclaimers. The SaaS default is "THE SERVICE IS PROVIDED AS-IS, WITH ALL FAULTS." A services model warrants work performed in a "good, professional, diligent, and workmanlike manner in accordance with industry standards". You cannot pay for outcomes and accept as-is in the same document.
• SLAs that measure accuracy, not uptime. Mayer Brown's worked example is exactly the right shape: "99% of invoices processed correctly against the purchase order", alongside "99% of support tickets actioned within the required service window". Availability SLAs are irrelevant when you are buying results.
• Audit rights over decision logs. The provider should have an "obligation to maintain decision logs for all decisions by AI agents", with the customer holding the "right to audit any agent's decision logs". This is the clause that makes the meter contestable. Specify retention, export format and a sampling right.

Add one thing Mayer Brown does not cover, because it is commercial rather than legal: a written dispute and credit mechanism. Name the log of record, the challenge window, the re-adjudication process and the remedy.

AI agent pricing per resolution vs per seat: when outcome pricing is the wrong instrument

Per-seat pricing has one virtue outcome pricing lacks: nobody argues about what a seat is. Outcome pricing is also not yet the norm. CIO Dive reported on 31 August 2026 that "only 19% of services buyers and 13% of service agreements on the seller side use such arrangements today", with Gartner's Tom Coshow saying "the increase in outcome-based pricing is more buzz than reality" and posing the question every buyer should ask: "If the vendor isn't taking on the risk, why are you bothering with outcome-based pricing?"

Foundation Capital sets out four models — access, usage, workflow ("charges per job done") and outcome — and names the structural limit: outcome pricing "aligns perfectly with value delivered" but "doesn't work if there's too much variability from company to company", because "outcome is ultimately a function of the customer's product, not just your software". If your resolution rate depends mostly on your knowledge base and your systems rather than the vendor's model, per job done is the more honest instrument — and easier to audit.

There is a finance-side consequence too. Deloitte's technology spotlight on outcome-based pricing in agentic AI (17 June 2026) explores "when a service offered in an arrangement reflects a stand-ready obligation versus a promise to deliver specified successful outcomes". Your definition of a resolution shapes revenue recognition on both sides — bring finance into the drafting, not just the price negotiation.

What to put in the contract before you sign

Procurement professionals are decision-makers in 53% of business buying cycles, alongside 13 internal stakeholders and nine external influencers (Forrester, 21 January 2026). So the definition work has an owner. Seven clauses to insist on:

• The billable event, defined positively and negatively — what counts, what explicitly does not (escalations, duplicate contacts, agent-initiated messages, test and sandbox traffic).
• The measurement window per channel, with same-issue recontact netted off inside a stated period.
• The adjudicator and the evidence — who or what classifies an outcome, what is logged, how long it is retained, and in what exportable format.
• Audit and sampling rights over decision logs, with a credit formula where classification fails.
• Accuracy SLAs with service credits, written as Mayer Brown frames them, plus performance warranties in place of as-is language.
• Price mechanics on volume movement — what the committed rate does if volume falls short or overshoots, and whether unused commitment rolls or expires.
• Change control on the definition itself. Zendesk changed its billable unit mid-year. You need notice periods and a repricing right when a vendor redefines the outcome mid-term.

Braivex's position on outcome pricing

We price on outcomes rather than logins, and we think that is the right model — but only where three things are written down: the outcome is defined in terms both sides can test, the evidence trail sits somewhere the buyer can inspect, and there is a named dispute route with a credit attached. Without those, outcome pricing is not risk transfer; it is a metered bill on a unit only one party can count.

Where the outcome depends more on your systems than on ours, we say so and quote per job done instead. That distinction shows up in how our agents are scoped and in the deployment packages on our marketplace.

If you are pricing an agent deal now, get in touch — bring the vendor's own resolution definition and we will tell you what it leaves open.