In August 2024 Zendesk announced that "customers will only incur costs for issues that are resolved autonomously by AI." Customers had already been arguing about what "resolved" meant for months. On the community thread for the new automated-resolution pricing, one wrote in May 2024 that about 60% of the resolutions in their data were chats where the customer had simply left: "You're going to charge $1.5 per unresolved conversation where a client just left the conversation with a bot." Another reported around 1,300 automated resolutions in 30 days against an allotment of 210 a month, and expected to be "paying $10,000+ yearly for a service that has always been included."
The first post in this series was about deciding what a unit is. An outcome is the hardest version of that question, because something has to decide each time whether the outcome happened, and in Zendesk's case that is a language model.
A meter that reads the transcript
On May 18, 2026 Zendesk replaced the single automated resolution with three tiers. In an assisted escalation "the AI agent contributed to the interaction before a human agent completed the resolution"; in a contained resolution it "handled the interaction to completion without the customer requesting further assistance"; in a verified resolution it "successfully resolved the interaction." The first two each carry the line "This tier does not count against your resolution allowance." The third doesn't, so it is the tier that draws on the allowance and the one the customer pays for.
What separates contained from verified is a 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." The help article sets when that is: an email conversation "ends 72 hours after the first email in the conversation," a messaging one "2 hours after the first message" by default, adjustable to 72, and a voice call "immediately upon hangup." Zendesk's Relate announcement the next day added a second check: "Every resolution Zendesk charges for is verified," both by the AI agent that handled it and "independently confirmed by a dedicated AI evaluation model," with "spam and routine exchanges" excluded.
The money runs through a dollar pool. The help centre describes a resolution allowance as "a flexible currency pool used to pay for the cost of automated resolutions," worth $2 to $10 per agent seat per month depending on plan, capped at $5,000 a year. Monthly allowances "do not carry over," and "overage is billed monthly, regardless of your subscription terms." The prices that pool is spent at are harder to pin down. Zendesk's UK pricing page still lists counted resolutions, at £1.30 committed and US$2.00 pay-as-you-go, while the article on moving to allowances says its own example prices "are placeholders only, not the actual prices."
Two kinds of rule
The contrast is HubSpot, whose 72-hour timing was the subject of an earlier post. Its Customer Agent bills a resolution when the conversation includes "at least one reply from the customer agent that shares a content source" or performs an action, and there is "no qualifying handoff to a human agent within 72 hours of the last visitor response." Each condition is something a customer can check against the transcript: was an article shared, did the visitor ask for a person, when did they last write. A dispute under that rule is a dispute about facts.
Zendesk's rule asks a different question, whether the request was "satisfactorily resolved," and the answer comes from reading the conversation. Two people reading the same abandoned chat can disagree about that, which was the argument on the 2024 thread, and the 2026 tiers turn the disagreement into a classification. Fin, which Salesforce completed buying on September 10, 2026, takes a third approach and lets later evidence undo the charge: if a customer "later returns to the same conversation seeking further assistance" then "(even across billing periods) that resolution will be deducted and not charged," according to its help page.
None of the three help pages I read describes a way to contest a single resolution. Zendesk does show its working in the support product: ticket events "show the resolution types and assigned tiers," and the AI agents conversation overview "shows which tier a conversation was assigned and rationale as to why." That evidence sits next to the conversation, though, and not on the invoice.
Modelling it, on paper
As in the earlier posts, this maps the published rules onto Revenue Cloud Usage Management using the public usage-consumption skill in the SalesforceLabs repository. It's a thought experiment, not a claim about how Zendesk bills, and the platform behaviours are the repo's account of them, not something I've run.
The repo's data model lists a TransactionJournal, the usage event, by the resource it counts and its status, Pending until a summary absorbs it. The first decision is what reaches that journal at all. One option records only verified resolutions, so every row is billable and the rated summary is exactly the invoice, but the customer can't see the conversations that were judged and not charged. The other records every ended conversation against three usage resources, rates the two free tiers at zero, and keeps the denominator, which is the comparison the 2024 complaints were about: how many conversations did the AI touch, and how many of those did it bill?
Microsoft's Copilot Studio documentation shows what the second option looks like from the buyer's side, and where it stops. Its consumption reports include a "Billed credits / Non-billed credits" column, but they "don't show" usage "for a single conversation or message" or "the reason a row wasn't billed."
That reason is the second decision. The judgement arrives after the conversation has ended and, for month-end email, possibly after the period has closed, so the late-event and close problems from earlier posts apply unchanged. What's new is that the quantity on the row is the output of a model. The skill's opening warning, that "almost every mistake is silent" and that a wrong configuration produces "a plausible price that is not the one you configured," was written about setup errors, but a misclassified conversation fails the same way, as a plausible row. The previous post ended on a commitment switch that "no runtime object snapshots"; a verifier's verdict is the same kind of input, deciding the charge without appearing on the rated record unless the design puts it there. Keeping it means carrying the tier, a link to the transcript, the rationale and the version of the evaluation model alongside each journal, for as long as an invoice can still be challenged.
Zendesk has published one correction notice, undated, which says a "temporary reporting error" inflated automated-resolution counts in Admin Center and that customers were not charged for them (help). In that case the report and the bill disagreed and the bill was right. When the meter is a model that can be retrained between one billing period and the next, what does the vendor have to keep so that last quarter's verdicts can still be re-read the way they were made?
No comments yet