The Meter Is Also a Model
Co-Founder & CEO of CLRT
You are about to sign for an agent that bills per resolved case, and the pitch is that the vendor now carries the risk: no resolution, no charge. Read the contract again and ask who decides what a resolution is. Not you. In 2026 three of the best-known vendors in the category each rewrote that word. Fin changed its billing unit from resolutions to outcomes in March. Zendesk split resolution into three tiers in May and bills only the tier a dedicated evaluation model confirms. Salesforce folded agent consumption back into per-user editions in September, a week before completing its purchase of Fin. The unit you are budgeting against is a definition the seller owns, reads with its own software, and revises at will. That is not a transfer of risk. It is a transfer of the audit.
Start with the customer who checked. In May 2025 a Fin customer posted on Intercom's own forum a manual audit of forty tickets the agent had touched over four weeks. Sixteen were marked resolved by Fin, a 40 percent rate, and that sixteen is what an invoice is built on. Nine of the sixteen, by the customer's reading, were not solved by Fin at all, putting the real rate, on the customer's own count, at 12.5 percent. One customer, self-reported, forty tickets: weigh it as such. What matters is the reply. Intercom's staff did not dispute the count. They restated the rule, a conversation is billed as a resolution when the end user gets an answer and does not ask for further help, and conceded a misstep in how billing had been described at a customer event. The dashboard and the customer disagreed by a factor of three, and both agreed the dashboard was working as designed.
The rule that produced the gap is the rule you will be billed under. Fin's help centre defines an assumed resolution: if a customer disengages from the conversation for 24 hours after Fin's last answer, it is counted as resolved. Zendesk's pre-tier definition ran on the same principle with different clocks, 72 hours after the last email, two hours after the last message by default, immediately on hangup for voice, after which a language model evaluated the transcript. In each case the billed event is not the customer's satisfaction but the customer's silence, measured on a timer the vendor sets and read by a model the vendor runs. A customer who gives up, phones instead, or opens a new thread is a resolution. Fin deducts the charge if the customer returns to the same conversation. It has no view of one who never comes back.
The definition has not held still for a quarter. Salesforce launched Agentforce in October 2024 at two dollars per conversation without publishing what a conversation was. Seven months later it moved to ten cents per action, conceding in its own announcement that planning usage at scale was difficult. On 3 September 2026 it folded 500,000 to 2.75 million credits into per-user editions at 195, 395 and 550 dollars a month. Between those moves its Flex Credits rate card grew from nine metered usage types on 3 February 2026 to 25 on 17 June, nineteen weeks later, with a per-resolution line now inside the per-action system and a printed reservation that usage types, tiers and multipliers may be updated from time to time. Fin renamed its unit from resolutions to outcomes on 12 March. Zendesk cut resolution into three tiers on 18 May and made two free. Three vendors, three restatements in one year.
Look at the tier Zendesk still charges for. A verified resolution, in its own words, is one the agent resolved end to end and a dedicated AI evaluation model independently confirmed. Against the silence rule that is an improvement. But notice what it makes the invoice: the verdict of a classifier the vendor trained, on data the vendor holds, against a standard the vendor wrote. A dispute about the bill is now a dispute about that classifier's judgment of a transcript. The meter is also a model. Sierra, the vendor with the purest outcome pitch, said the quiet part in June: outcome pricing only works where the software is highly autonomous and highly attributable, and where the outcome cannot be cleanly attributed to the software, you will get into an endless debate. That is a seller describing the boundary of its own model. Few buyers have checked which side their workflow sits on.
So the risk the buyer believed it transferred is precisely the one it can no longer inspect. Under seats the bill was dull and the value vague. Under outcomes the value is finally counted, by the only party whose revenue rises with the count, using a definition it can change on a help page. Salesforce's completion of its Fin purchase on 10 September said nothing about pricing; its 15 June agreement cited a 76 percent average resolution rate on the way to a 3.6 billion dollar price, and nobody published how that figure was counted. None of this is bad faith; it is what happens when one side of a contract has done this before. What protects the buyer is a definition of a resolved case written in its own words, an independent record of what happened to the customer, and a monthly reconciliation against the vendor's meter. That is contract and measurement work, and almost nobody budgets for it.
The billed event is not the customer's satisfaction. It is the customer's silence, measured on the vendor's timer and read by the vendor's model.
A deeper dive
The mechanism has three layers, each invisible from the dashboard. The first is the timer. A silence window is a modelling assumption about human behaviour dressed as an accounting rule: it assumes a customer who stops replying has been helped. That assumption is most wrong for exactly the cases that cost you most, the frustrated customer who leaves, the customer who escalates through another channel, the customer whose problem reappears next week under a new ticket. The second layer is the classifier. When a vendor says a model verifies each resolution independently, independently means independent of the agent, not of the vendor. The verifier and the agent are built by the same company on the same corpus toward the same commercial goal, and when both agree a case was resolved, that is one company's opinion twice. The third layer is the definition, which lives in a help article rather than the contract, so the unit you signed for can be widened, split or renamed between two invoices without anyone breaching anything. Each layer is defensible alone; together they place the whole of the billable truth on the vendor's side of the table.
The second-order trap is behavioural, and procurement never models it. Once silence is the billed event, every human intervention on your side becomes a billing question. A support lead who steps in early to rescue a struggling conversation may turn a free escalation into a paid resolution, or the reverse, depending on the rule of the month, and the rational response is to let the customer wait for the button. The metering rule has quietly started managing your team. The same pressure runs through the vendor's product decisions. A unit charging 9.99 dollars for a qualification and 99 cents for a resolution has priced the relative value of your outcomes for you, and a rate card that adds sixteen usage types in nineteen weeks cannot be budgeted against, only reconciled after the fact. The only durable defence is a definition of done that you wrote, evidence drawn from your own systems rather than the vendor's export, and a sampling discipline that reads real conversations rather than resolution flags. That is not a procurement exercise. It is a measurement system, and building one cheap enough to run every month and rigorous enough to stand up in a dispute is engineering judgment most in-house teams have never had to exercise.
Work with CLRT
CLRT sits on the buyer's side of this table. We write the definition of a resolved case in your words rather than the vendor's, build the independent record that shows what actually happened to your customer, and run the monthly reconciliation that tells you when the meter and the truth have parted. Where a workflow cannot carry an outcome price at all, because the outcome cannot be cleanly attributed, we say so before the contract does. If you are signing for an outcome-priced agent, talk to us first, and if you want to know which of your workflows can safely be metered this way, that is the question CLRT Ascent at ascent.clrtstudio.com was built to answer.
Vishal Sachar is the Co-Founder and CEO of CLRT, where he helps UAE businesses make sense of applied agentic AI and put it to work. He writes on agentic systems, AI governance, and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.
