Intercom Fin and the Resolution Number
Fin launched in 2023 with a genuinely new idea: pay 99 cents only when the bot resolves a conversation. Three years on the price is unchanged, but the number keeps rising. A lookback, using Intercom's own figures, at what "resolution" has actually counted.

When Intercom shipped Fin in 2023, it did something no rival would touch: it put a price on the outcome. Not a seat, not an API call — 99 cents every time the bot resolved a conversation.
The pitch was disarmingly fair. If Fin can't answer, Fin is free. You pay only when the customer's problem goes away.
Three years on, the price per unit hasn't budged. Almost everything else about the number has. This is a lookback — using Intercom's own published figures — at what Fin's "resolution rate" actually counted between 2023 and 2026, and why a metric that only ever rises deserves a second look.
Start with the definition, because it does the heavy lifting. Intercom counts a resolution two ways. A confirmed resolution is when the customer types something affirmative — "thanks, that helped." An assumed resolution is when the customer simply leaves after Fin's last answer without asking for anything more.
Both bill at 99 cents. The distinction matters: a customer who gives up and closes the tab looks, to the meter, exactly like one who was helped. Intercom is transparent about this — greetings, clarifying questions, and human escalations aren't billed — but "assumed" is carrying real weight in a figure people quote as if it were verified.
Now the climb. At launch, Intercom's CFO later put Fin's average resolution rate at roughly a quarter of conversations.
By the Fin 2 launch in October 2024 — "the first AI agent that delivers human-quality service," with a claimed 99.9% accuracy — the average was 51%.
Fin 3, a year later, reached 66% across more than 6,000 customers.
By early 2026 Intercom was citing 76%.
Some of that is real: better models, better retrieval, more of the customer's own content wired in. But two accounting changes rode alongside the capability gains, and both push the number the same way.
The first was semantic. In March 2026 Intercom announced it was moving "from resolutions to outcomes." A resolution — Fin handling an issue end to end — became one type of outcome. Procedures, where Fin gathers information and takes an action, became another. The billable unit widened while the headline word stayed familiar.
A number that keeps rising while the definition of what it counts keeps changing isn't a benchmark. It's a moving target you're being billed against.
The second was arithmetic. In mid-2026 Intercom changed how the rate is computed, excluding "Fin Constrained" conversations — where Fin was active but never got to attempt an answer — from the denominator. Its own help center is blunt about the effect: resolution rate "will increase, because Fin Constrained conversations are no longer counted in the denominator."
Fewer conversations on the bottom of the fraction, the same resolutions on top, a higher percentage. There's a defensible case for the change. It also means the 76% measured before it and the number measured after are not the same measurement.
None of this makes Fin a bad product or the pricing dishonest. Outcome-based pricing was a genuinely fair idea, and paying only when the bot works still beats paying per seat for software nobody opens — the tradeoffs laid out in the support tooling field guide. The point is narrower: a vendor's resolution rate is a figure the vendor computes, and Intercom has revised what it counts at least twice while the figure rose. That's the pattern we flagged in what "AI resolved 70%" actually means and why deflection is a vanity metric.
If you're buying, don't argue with the headline — reproduce it. Ask what share of "resolutions" were confirmed versus assumed. Track your own reopen rate on the tickets Fin closed. Watch containment against the escalations you can actually see. A truly autonomous resolution is one you can verify — not one the invoice asserts. Fin's number climbed for good reasons and for accounting reasons, and only your own data will tell you which share you're paying for.