What Counts as a Result? The Sentence Every Per-Result AI Contract Depends On

The sentence most buyers struggle to write

Every per-result contract rests on one sentence that says what counts as a result. If that sentence is vague, each month ends in an argument about the count. Larry Dignan of Constellation Research expected exactly that when Salesforce launched its per-resolution Help Agent, predicting "haggling over definitions and fuzzy math".

Buyers know where the weak point is. In an a16z buyer survey cited by BCG, 47% said they struggle to define clear, measurable outcomes, and 24% said outcomes depend on factors the vendor does not control. My argument in this post is that an outcome-based contract for AI deserves more negotiating time on the definition than on the price, because the definition decides how many times the price is charged.

How the support vendors define it

Customer service vendors publish their definitions, which makes them the best examples available. Fin counts a resolution when the customer confirms the answer worked or "does not request more help after Fin answers", a case it treats as an assumed resolution. Zendesk starts from conversations that were not escalated to a person and has an LLM evaluate each transcript. Since a change that rolled out in May and June 2026, it bills only resolutions marked "Verified", which requires "additional signals confirming the outcome was complete and satisfactory".

Salesforce's Help Agent counts a resolution when "the user's question doesn't need a human escalation", and treats actions within a 10-minute window on a call as one resolution. Gorgias adds a time rule from the other direction: if someone on your team replies within 72 hours of an AI resolution, the conversation is billed as a helpdesk ticket and not as an automated interaction. Ada's documentation requires an automated resolution to be relevant, accurate, safe and contained, as assessed by Ada's own AI.

What happens when a result is reversed or handed off

A definition also has to say what happens when a result turns out to be wrong later. Fin is the only vendor in our research that publishes a reversal rule: if the customer comes back "seeking further assistance (even across billing periods)", the resolution "will be deducted and not charged". Most other vendors say nothing on the point, which leaves it to whoever drafts your contract.

Handoffs need the same care. Zendesk, Fin, Sierra and Salesforce's Help Agent generally do not bill an escalation to a person, and neither does Decagon on its per-resolution plan. Fin is the exception in one respect, since it bills a Procedure you configured to end in a handoff. Your staff still finish the escalated work, so a high handoff rate can make a cheap price per result expensive in total.

Silence is the hardest case. Ada, which sells conversation pricing and so has a stake in the argument, calls it a "containment trap" when a customer who simply goes quiet is counted as a success. That point is fair whoever makes it. If your definition counts silence, you need a reversal rule and the right to audit transcripts.

Writing the definition for document work such as invoices

Outside support there is rarely a customer to confirm anything, so the result has to be tied to what happens downstream. For supplier invoices I would write something close to this: one result is one invoice that the agent matched to its purchase order and goods receipt within the agreed tolerance and that posted to the ERP without anyone on the AP team changing a field. An auditor can test that sentence against the ERP's change log, and it charges once per invoice however many attempts the agent needed.

The exclusions deserve as much care as the main clause. Invoices with no purchase order in the system should go to a person and stay unbilled, since there is nothing for the agent to match against. Test runs and invoices resubmitted by the supplier should not count either, and the contract should say so in plain words.

Then add the reversal rule. If a posted invoice is corrected within an agreed window because the agent got a field wrong, the result is credited on the next bill. Use the same definition in the acceptance test, so the pass rate measured before launch is the pass rate you are billed against afterwards. Our acceptance criteria generator drafts criteria like these from a description of one workflow.

The terms a CFO will ask for

Once the definition is settled, finance will want the commercial terms around it. Each of the five below has published practice a buyer can point to.

• The cap, and what happens when it is reached. Zendesk's "Don't allow overage" setting pauses the AI agent and sends further requests to people, while Microsoft's Copilot Studio disables custom agents at 125% of prepaid capacity.
• The minimum, and whether unused results roll over. Fin's standalone plan has a 50-outcome monthly minimum and unused outcomes are lost, while Salesforce sells its Help Agent in prepaid packs of at least 1,000.
• How the unit price can change. Gartner expects inference costs per agentic workflow to rise more than fivefold through 2028, so a vendor will want room to reprice, and a buyer will want the price fixed for the term with notice and a limit on any change at renewal.
• Audit rights over the count. Zendesk checks conversations with an LLM, and Ada assessed only a statistical sample of conversations before July 31, 2024, so ask whether every unit is checked and whether you can read the logs. HFS Research puts the requirement plainly: "Outcome-based pricing only works when both sides trust the number."
• A dispute process with a timeline. No vendor in our research publishes one, and BCG recommends agreeing exactly what qualifies as an outcome, "backed by a dispute process".

A tighter definition costs more per result

There is a trade-off to accept going in. The stricter the definition, the fewer items count, and the vendor will price each counted result higher to cover the work it does on the items that do not count. Decagon's pricing fits that pattern, since its per-resolution option comes at "a higher fixed rate".

No definition ends every argument either. David Haber at a16z expects that "attribution debates will likely surface initially" in contracts priced on outcomes. Where other systems or people touch the same work, I would write down in advance which step owns the result. A precise sentence with a reversal rule keeps those debates short, though it will not prevent them.

How we handle it at QueryNow

We write the result definition into the acceptance criteria you sign, so it is settled before the build starts. The build of one agent workflow costs $10,000, paid only after the agent passes those criteria.

After launch you pay a fixed price for each result that passes its checks, set from the acceptance test numbers, with a $1,200 monthly minimum. Work that fails a check or that your team finishes is not billed. Several of our pharma and life sciences clients pay this way, and the terms are on our Results as a Service page.