AI support agents should define "resolution" on the pricing page

Our view is simple: when an AI customer support product is priced "per resolution", the pricing page should say, in plain words, what counts as a resolution. At the moment that definition usually sits in a help article, a billing document or a sales call, and it differs from vendor to vendor. A buyer comparing two "per resolution" prices is not comparing like with like.

The timing matters. Pricing for AI support agents is still being rewritten. Zendesk's help centre says its automated resolution tiers were introduced on May 18, 2026. Front's help page for its Autopilot add-on shows an edit date of October 5, 2026. A pricing index run by the support vendor Macha, checked on 24 September 2026 and updated in October, counted ten of 32 AI support vendors with no usable published list price. Macha sells a competing product, so treat its figures as a signal rather than a verdict. The direction is still clear: the unit of billing is moving, and buyers have to chase it across several pages.

One word, several meanings

Take three documented examples, all from the vendors' own pages.

Zendesk says that paying per automated resolution means you pay only for requests resolved by an AI agent without escalation to a human. The same article then groups resolutions into three tiers: assisted escalation, contained resolution and verified resolution. Whether a conversation is verified is decided by a language model that evaluates the text of the conversation to confirm the request was satisfactorily resolved. The article also sets when a conversation counts as ended: 72 hours after the first email, two hours by default for messaging, and on hang-up for voice. None of that is unreasonable. It is simply a lot of detail that a headline price does not carry.

Front describes its Autopilot pricing as outcome-based, with three levels: triage, handoff and resolution. A resolution is a conversation with no follow-up message from a teammate within 24 hours on chat-style channels, or 72 hours for email and portal. A handoff, meaning the AI drafts or acts for a human to take over, is a separate charge. At the time of writing, the page lists $0.05 for triage, $0.39 for handoff and $0.89 for resolution. Front deserves credit here: its definitions are on one page, with examples. Even so, "you only pay when Autopilot delivers the outcome" and a charge for a conversation a human finishes are two different ideas of an outcome.

Gorgias documents that an automation fee and a ticket fee may apply to the same ticket when the AI Agent resolves it without human intervention. Its billing page says a ticket counts as automated if the customer does not need a human agent within 72 hours. A buyer who reads "per resolution" and budgets for one fee per resolved ticket would miss the second fee, which sits in the documentation rather than the headline.

Macha's index also reports that Help Scout and Intercom count a customer who does not ask for more help after the AI's answer. Different time windows, different handoff rules, different stacking: one phrase, many invoices.

Why this matters more than a rounding error

In most software, the vendor defines the unit, but the buyer can see the unit: a seat, a gigabyte, a message. A resolution is different, because it is an outcome the vendor measures and then bills. Macha points out that the vendor writes the definition of "resolved", and that a per-attempt price and a per-resolution price must be converted before they can be compared. Small shops that run a few hundred tickets a month may have little room to negotiate bespoke terms. They read the pricing page, pick a tool and find out the rest on the first invoice.

There is also a regulatory backdrop. A law firm's review of the US Federal Trade Commission's work says the agency is scrutinizing claims about AI performance. We are not suggesting any vendor named here is the subject of a complaint, and we are relying on the vendors' own documents. The point is narrower: clear definitions are cheap for vendors, and they protect honest sellers from being compared unfairly with vague ones.

The strongest counter-argument

Vendors have a fair reply. Per-resolution pricing is arguably the most buyer-friendly model the category has produced, because it ties cost to results rather than to seats or raw message counts. Definitions need room to change as models improve, and a pricing page crammed with time windows and edge cases would be unreadable. Documentation exists for this, it is public, and, as the examples above show, some vendors publish it in detail. Large buyers negotiate the terms in a contract anyway, so a short pricing page is a feature, not a flaw.

We accept most of that. Outcome pricing is a real step forward, and detailed help pages are better than none. But the argument proves less than it seems. If the definitions are public, then linking them next to the price costs nothing. If they change, as Zendesk's May change shows, then a dated change note belongs beside the price too. And the contract argument only helps buyers big enough to have one.

What a pricing page should carry

• The billed event: resolution, conversation, ticket or attempt, in one sentence.
• The clock: how long a customer must stay silent, by channel.
• Handoffs: whether a conversation finished by a human is billed, and at what rate.
• Stacking: whether seat or ticket fees apply on top of the AI fee.
• Who decides: customer behaviour, a model or a person, and whether the buyer can see the verdict.
• Change log: the date the definition last changed.

Until pages do that, buyers can protect themselves. Read the billing documentation, not only the pricing page. Ask each vendor to price the same sample month, including handoffs. Check whether a free or cheaper option, such as a plain help desk with a knowledge base, already covers your volume. And note the date you read each page, because these definitions are still moving.