What counts as a resolution in Fin, and what it bills

intercom, fin, pricing, ai-support, vendor-evaluation

What counts as a resolution in Fin is one of two things: the customer replied that the answer helped, or the customer went quiet for 24 hours after Fin's last message. Intercom calls the first a confirmed resolution and the second an assumed one, and bills both at $0.99. The customer who left satisfied and the customer who gave up and emailed you from another address produce the same line on the invoice.

So the number worth having before your first bill arrives is not the resolution rate. It is the split inside it.

What counts as a resolution in Fin

Fin's help centre defines a resolution as a conversation where Fin answers and the customer either confirms it was answered — replying something like "that helped, thanks" — or exits without requesting more help. The second case has a clock on it: if the customer disengages for 24 hours after Fin's last answer, it is an assumed resolution, and it is billed exactly like the first.

Fin's own product behaviour shows which of the two Intercom would rather have. By default it asks "Is that what you were looking for?" on a first answer and "Did that answer your question?" on later ones — prompts whose only job is to turn a would-be assumed resolution into a confirmed one. A vendor does not build that if the two are equally good evidence.

What else Fin bills for

Resolutions are one of four billable outcomes:

• Resolution — $0.99. Fin answers; the customer confirms or asks for nothing more.
• Procedure handoff — $0.99. A Procedure you configured runs to completion and ends in a handoff to a human or a workflow.
• Disqualification — $0.99. Fin decides a prospect does not match your qualification criteria.
• Qualification — $9.99. Fin matches a prospect to your criteria and routes them onward.

You are charged for at most one outcome per conversation, however many actions Fin takes inside it. Standalone Fin — the version that plugs into Zendesk, Salesforce, HubSpot or Freshdesk instead of Intercom — carries a 50-outcome monthly minimum, a floor of about $49.50, and unused outcomes do not carry into next month. Inside Intercom, outcome charges sit on top of seats, which start at $29 per seat per month on annual billing and run to $85 and $132 on higher plans.

Why "assumed" is doing most of the work

Silence after an answer is consistent with at least five stories: the customer got what they needed; got distracted; found the answer somewhere else; gave up and opened a support email; or gave up and started a fresh chat. One of those is a resolution in the ordinary sense of the word. Billing treats all five identically, because from inside the conversation they are identical.

The gap this opens shows up in Intercom's own community. In a thread titled "Fin's flawed assumed resolved & pricing design", customers report confirmed-resolution rates in the 6–7% range against assumed rates near 60%. In a second thread, one customer counted 40 conversations Fin touched: 16 came back marked resolved by Fin — a 40% resolution rate — and on reading them, 9 of the 16 had not been solved at all, leaving 12.5%.

Those are single-account reports, not benchmarks, and we cannot audit them. What makes them worth reading is where the disagreement sits: nobody disputes the count, only what the assumed half of it means — and a metric that sums both kinds cannot express that disagreement at all.

What Fin does not bill for

The list of non-outcomes is genuinely reasonable, and it is the part most of the angry threads predate. You are not charged when a customer explicitly asks for a human, when a Procedure fails to complete, or when a conversation is passed to your team without an outcome. Fin's default frustration detection escalates to a human, and that escalation is free too.

Then the rule that answers most of the original complaint: if a conversation was resolved — confirmed or assumed — and the customer later returns to that same conversation asking for more help, the resolution is deducted and not charged. It applies across billing periods, so a January charge can come off a February invoice. Intercom shipped this in May 2025, after the threads above.

Read the wording, because it is doing precise work: the same conversation. A customer who abandons the chat and starts a new one — new session, new email, different channel — has not returned to it. On the rule as documented, that path bills the assumed resolution and then bills whatever the second conversation ends as. We could not verify how Fin handles a re-contact that arrives separately: the help centre does not say, and we have no account to test it in. Which makes it the most valuable thing to measure in your own data — step four below.

How to audit your first month

You do not need Fin's dashboard to agree with you. The raw material is in the API.

1. Pull the state, not the rate. Every conversation carries an ai_agent object, and ai_agent.resolution_state holds the outcome: confirmed_resolution, assumed_resolution, routed_to_team, procedure_handoff, and negative_feedback (the state formerly called abandoned). It is on the conversation model from API version 2.11 onward. For a whole month, use the Reporting Data Export instead — get_datasets and enqueue, both under /export/reporting_data/: the first lists the fields available, the second starts an export of the Conversations dataset.
2. Compute two numbers. Assumed ÷ (assumed + confirmed) is how much of your deflection is inference rather than testimony. Assumed × $0.99 is what that inference cost this month. Put both on the same line of the same report, every month.
3. Join to satisfaction. The conversation-rating dataset records which kind of agent was rated — teammate, Fin, or a chatbot — so you can read the ratings left on assumed resolutions specifically. A poor rating on a conversation billed as resolved is the cleanest contradiction available to you.
4. Hunt for re-contacts outside the thread. Take every contact whose conversation ended as assumed_resolution and list the other conversations that contact opened in the next 72 hours. Those are customers who came back, but not to the same place, so no deduction fired. Their count times $0.99 is the part of your bill the refund rule does not reach.
5. Read twenty of them. Twenty transcripts is enough to tell "answered and left" from "asked twice, got the same article, left".

If assumed resolutions are more than half your total, check two things before arguing about price. Whether Fin is asking the confirmation question at all in your setup — an unanswered "Did that answer your question?" is still silence, but a missing one guarantees it. And your knowledge base: sort content by conversations involved against conversations resolved, and the articles with the widest gap are the ones producing the silence. If your customers write in several languages, check which language your index is actually in before blaming the model — a retrieval miss and a bad answer end the same way, in an assumed resolution.

What this changes when you compare prices

Per-outcome pricing gets compared against per-seat pricing as though the two were the same kind of number. They are not: one of them moves with how often your customers stop replying.

Take 2,000 resolutions in a month. At $0.99 that is $1,980, and if 60% of them are assumed — the ratio those threads describe — then $1,188 of the bill is priced on silence. Three seats on a flat plan do not move when the knowledge base gets worse. This is not an argument that per-outcome pricing is wrong; at low volume it is plainly cheaper, and a floor of about $49.50 beats any seat-based plan on the market. It is an argument that the two models fail differently, and that the outcome model's failure is invisible until you split confirmed from assumed.

One more input, which nobody controls: Salesforce signed a definitive agreement to acquire Fin for about $3.6 billion in June 2026, with closing expected in its fiscal Q4 2027. Pricing models rarely survive an acquisition unchanged. If you are re-evaluating anyway, we wrote up the fifteen questions to put to a vendor's documentation before shortlisting from a listicle.

Where Replium stands on this

The method above is only fair if we run it on ourselves, so here are the same questions answered about us, starting with the ones where the answer is no.

Do we bill per resolution? No — we do not bill at all. There is no billing in Replium today: no plans, no card, no invoice to audit.

Do we have an AI agent that could resolve a conversation on its own? No. We have not connected an LLM provider to the product. Nothing on our side would generate a resolution to charge for, assumed or otherwise.

Who should stop reading here? Anyone who needs ticket deflection this quarter. Fin does that, well enough that 30,000 companies pay for it, and we are not a substitute for it today.

What Replium is right now: a multi-tenant help desk with a shared inbox, an embeddable chat widget, and a public knowledge base that is server-rendered so search engines read the text instead of an empty shell — in English and Russian. Small, honest about being small, and every one of those pieces exists in the product you can sign up for today.

The one thing a vendor this size is genuinely better at is distance. Your report does not queue behind thirty thousand accounts; it reaches the people writing the code, and small things ship in days. That is worth exactly what it sounds like — no more — but it is true, and it is not something anyone can buy their way to.

If that trade sounds right, create a workspace. It is self-serve, it takes an email confirmation, and since there is no billing there is nothing to cancel.