Intercom announced a fairer Fin billing rule on stage, then took it back

The correction is the story, so here it is first.

At Intercom's "Built for You" event on 21 May 2025, the company presented a change to how Fin is billed: if a teammate replied after Fin's answer, the conversation would no longer count toward the Fin resolution bill. A customer in the audience understood it that way and said so afterwards, noting it had been stated on a large slide beside Intercom's Chief Product Officer.

On 4 June 2025, in Intercom's own community forum, an Intercom employee posted that this was not accurate. A conversation is billed when the end user gets an answer and does not ask for further help. Only an explicit request for a person stops the charge. The following day, the same employee wrote:

Correcting yourself in public, on your own forum, is rare and deserves saying so. But the position it leaves is unambiguous: the customer-friendly rule was described from a stage and does not exist. What exists instead is the rule below, and it has a branch most people do not expect.

When Fin bills, in one diagram

Fin's own help centre sets out the logic, and to its credit it is public. A resolution is counted when, after Fin's last answer, the customer either confirms the answer helped, which Intercom calls a confirmed resolution, or exits without requesting further assistance, which Intercom calls an assumed resolution.

1. Fin answers the customer
2. Did the customer ask for a person, or for more help? Yes Not billed
3. Did Fin escalate under its own default rules, or your workspace escalation rules? Yes Not billed
4. Billed as an outcome, $0.99 Including when one of your own agents steps in to fix a poor answer without the customer having asked, when a procedure you configured ends by handing the customer to your team, and when the customer simply stops replying for 24 hours. There is no route out for "gave up": from here, silence and satisfaction look the same.

The two exits are both fair, and Intercom should be credited for them. If the customer clicks through for more help, or writes back asking for a person, the conversation is not billed. Nor is it billed when Fin itself decides to escalate, for instance because it detects frustration, or because a rule you set for the whole workspace tells it to.

The branch that catches people is the one with no exit. If no explicit request was made, and one of your own agents notices a poor answer and steps in to fix it, the earlier Fin response still counts as a charged resolution. The customer never asked for a human. A human arrived anyway, precisely because the answer was not good enough. The bill does not move.

Intercom's pricing now bills per outcome rather than per resolution, and its help centre is candid about what that covers. An outcome is a resolution, or a hand-off to your team made because a procedure you configured told Fin to hand off. The same article answers the question of the frustrated customer directly: one who leaves after Fin's answer without asking for more help is counted as an assumed resolution, and billed at $0.99.

What that does to a resolution rate

Salesforce agreed to acquire Fin in June 2026 for around $3.6bn, and the announcement put a number on the case: Fin's agents resolve 76% of support volume end to end.

That figure is not disputed here, and it may well be right. The point is narrower and harder to argue with: a rate produced by this definition cannot be checked from the outside, and every source of error in it runs the same way. Conversations your team quietly rescued count as successes. Conversations the customer abandoned count as successes. None of its errors push the number down: the only ways out of the billed column are the customer asking, or Fin choosing to escalate.

That is what makes it a definitional question rather than an accuracy question. You do not need to believe anyone is being dishonest to notice that the measurement has a thumb on one side of the scale.

What one customer found when they checked

The customer who started that forum thread had done the obvious thing and read their own conversations.

A deliberate note about that figure, because this piece is about numbers people cannot check. The same customer also states an overall rate of 12.5%, which does not reconcile with subtracting 9 from 16. And a later reply in the thread, from someone selling an audit tool, restates the finding as only 3 of the 16 being genuine, which contradicts the customer's own account.

We are using none of those numbers. One sample from one company is not a resolution rate for a product used by around 30,000 businesses, and an article about unverifiable figures is a poor place to add one. What the sample does show is the mechanism working as documented, in the wild, at a scale one person could see with their own eyes.

How the rest of the market counts

Fin is not unusual here, and it is not alone in publishing its terms: all five vendors below do. The table is what each says it charges for, first checked against their own pages on 7 September 2026 and rechecked on 23 September. One of them changed its pricing in between, which is the best argument we can offer for dating every row. Follow the links rather than trusting the table.

eesel and Helply are the two to trust most, and for the same reason. Both bill for the work rather than the outcome, and both say so plainly: eesel that tasks are billed regardless of outcome, Helply that its credits are priced on complexity and not on what the answer turns out to be. You may prefer a different model, but there is no gap between what they claim and what they charge. Hugo meters escalated conversations, so you pay at the moment the product fails, though Crisp documents that openly rather than burying it. Help Scout is the only per-resolution vendor here whose definition keeps a human-assisted conversation out of the bill.

We have deliberately not filled the gaps. Where a vendor does not address the unprompted hand-off, the cell says so rather than guessing, because guessing is the behaviour this article is about.

How to check your own bill

This takes an afternoon and needs nothing but your own data. It works on any of the five.

1. Export a month of conversations where the AI was involved, with the resolution state, the reopen count, and whether a teammate replied.
2. Filter to the ones you were billed for.
3. Take a random sample of 40. Do not choose them by eye.
4. Read each one and mark it resolved, not resolved, or ambiguous, on your judgement rather than the vendor's. Count ambiguous as not resolved.
5. Compare your count with the billed count. The gap is what the definition is worth to you in pounds per month.

Two checks find most of it: conversations marked resolved that were later reopened, and conversations where one of your agents replied after the AI without the customer asking for a human. If your vendor cannot give you an export with those fields, that is itself an answer.

Where we stand

We sell a competing product, so treat this section as an argument and everything above it as sourced fact you can check line by line.

Varley bills per resolution too, at $0.50 (it was $0.49 when this piece was first published). We do not count assumed resolutions. A conversation that any human on your team joins is not billed, at any stage, whether or not the customer asked for a person, which is the branch this article is about. If a customer reopens the same issue within 30 days, the original charge is credited back in full. The complete definition, with the exclusions and the audit method, is published as its own page with a changelog.

Work short of a resolution, such as a drafted reply or an answer to a question from your team, is priced per action on a published list. A hand-off, an action you reject, a failed answer and a customer going quiet all cost nothing. We are not alone in that: Helply does not charge for failed actions or drafts your team rejects, and Help Scout does not charge when its AI stops short of a full answer. Compare the terms, not the adjectives.

We also do not publish a resolution rate. We do not yet have the volume for one to mean anything, and an unaudited percentage from the company making this argument would be worth less than nothing. When we publish one it will carry its denominator, its sample and its date.

Right of reply

We have not put these findings to Intercom, Salesforce, Crisp, Helply, eesel or Help Scout before publication. Everything here comes from their own public pages and a public forum thread, and each is linked so you can read the original rather than our summary of it. If anything on this page is wrong or has changed, write to support@dragapp.com and the correction will be published here with a date, including if it is inconvenient for us.

Changes to this piece

23 Sep 2026

Helply replaced its per-ticket pricing after this piece was first published. The original version described its page as read on 7 September, including a mismatch between its claim to charge for outcomes and a unit of any conversation processed. Neither appears on its current page, so both were removed and its row now reflects the new model. All five rows rechecked. Varley's own price moved from $0.49 to $0.50. Two claims tightened: that nothing pushes Fin's number down (the customer asking does), and that Fin publishing its definition sets it apart (all five do). Intercom now prices Fin per outcome, which includes a hand-off made under a procedure you configured: the diagram, the table and the text now say so, and the diagram shows the escalations Intercom does not bill.

7 Sep 2026

First published.

Sources

• Fin AI Agent resolutions, Fin by Intercom help centre. Read 7 September 2026, rechecked 23 September 2026.
• Fin pricing: Outcomes, Fin help centre. Read 23 September 2026.
• Fin's flawed resolution assumption, Intercom Community. Posted 27 May 2025, employee replies 4 and 5 June 2025, thread still open. Read 7 September 2026.
• Salesforce signs definitive agreement to acquire Fin, 15 June 2026.
• AI Resolutions pricing, Help Scout docs. Read 7 September 2026, rechecked 23 September 2026.
• Hugo pricing and billing explained, Crisp help centre. Read 7 September 2026, rechecked 23 September 2026.
• Helply pricing. Read 23 September 2026.
• eesel AI pricing. Read 7 September 2026, rechecked 23 September 2026.