Why Is Intercom's Fin AI So Expensive Per Resolution?
Intercom charges $0.99 per resolution for Fin AI because it's pricing the outcome, not the compute. A resolution counts once, when a customer confirms an answer worked or simply doesn't ask a follow-up question, per Intercom's own pricing documentation. The token cost of generating that answer, using GPT, Claude, or Gemini under the hood, is a few tenths of a cent. The other 99 cents is the packaging: the retrieval over your help center, the conversation handling, the escalation logic, the analytics, and Intercom's margin on all of it.
That gap between token cost and sticker price is exactly what shows up when a team compares Fin to running its own model key. A founder building a competing helpdesk laid the math out on Hacker News: Fin runs about $0.99 per resolved conversation, while bringing your own OpenAI, Anthropic, or Gemini key and paying the model directly runs roughly $0.001 to $0.003 per resolved conversation. At 5,000 AI replies a month, that's about $5,000 in Fin charges against $5 to $15 in model spend for, in his framing, "the same job."
What $0.99 is actually buying
That gap looks too large to be honest, so what does "the same job" actually leave out? The HN comparison comes from a founder pitching a competing product and has reason to flatten the difference. Fin's resolution price bundles several things a bare API call doesn't include:
- Retrieval against your help center and past conversations, kept current as articles change
- Escalation logic that hands off to a human teammate with context, mid-conversation
- A support-specific UI: the inbox, transcripts, and reporting that already exist on top of Intercom
- Intercom absorbing the liability and support burden of a wrong answer, not you
A team paying $0.002 per model call still has to build the retrieval pipeline, the handoff logic, and the inbox themselves, or use another product that already has them. The BYO-key number is real, but it's the cost of the model, not the cost of a working support agent.
The marketplace math that breaks
The comparison bites hardest for businesses with thin per-transaction margins, because $0.99 is a flat tax regardless of what the transaction is worth. Here's a hypothetical that plays out at a lot of thin-margin marketplaces: a package-forwarding operator, call it Parcelwise, clears a little under two dollars in margin per shipment and turns on Fin AI Agent for shipping-status and customs questions. A month in, someone runs the numbers for the CFO:
Fin resolved 2,100 conversations in March. That's roughly $2,080 in Fin charges, on a product where our average order nets us $1.80. We're spending more per resolved question than we make on the shipment it's about.
That's not a Fin problem so much as a margin problem that Fin's flat pricing makes visible fast. A SaaS company charging $200 a month per seat can absorb $0.99 a resolution without noticing. A marketplace clearing pennies per order can't, and the Deploy Fin community forum includes a real example along the same lines: a marketplace founder who stopped using Fin after roughly $12,000 in spend, citing margins of around €1 per user per month against Fin's flat per-resolution price.
Why Intercom prices it this way anyway
Outcome-based pricing isn't unique to Intercom. It's the same logic Zendesk and other vendors have moved toward for their AI agents: charge for a resolved conversation instead of for seats or tokens, because a resolved conversation is the thing the customer actually wanted to buy. A high-margin B2B SaaS company can treat $0.99 to deflect a ticket as a clear win against a human agent's time; a low-margin marketplace or e-commerce operation can watch that same $0.99 exceed the profit on the thing being discussed. The pricing model isn't broken; it's tuned for one kind of business and mismatched for another.
Where the BYO-key workaround stops working
Teams that do the math and decide to skip Fin entirely, wiring their own model key into a homegrown bot, usually hit the same wall a few months in. The $0.002 per call is real, but it doesn't include:
- Retrieval quality drifting as the help center grows. Someone has to keep the embeddings current, or the bot starts confidently answering from stale articles.
- Escalation handling. A bare API call doesn't know when to hand off, or how to pass conversation context to the human who picks it up.
- Model migrations. Swapping providers when pricing or quality shifts means re-testing prompts, not just changing an API key.
None of that is about resolution pricing. It's the cost of building and maintaining a support agent, which Fin's $0.99 was quietly covering. Below a few thousand resolutions a month, most teams find the DIY math genuinely wins. Past that, the maintenance burden starts to look a lot like the thing they were trying to avoid paying for.
Both routes share a blind spot that has nothing to do with price. Whether Fin resolves a conversation, a BYO-key bot does, or a human teammate takes over, none of that pipeline captures what the conversation was actually about once it closes. A customer asking about multi-carrier rates or reporting a bug gets a $0.99 answer or a $0.002 answer either way, and that answer is the last anyone downstream hears about the request unless a person manually writes it down.
Where Modem fits, and where it doesn't
That gap, a resolved conversation that leaves no trace of what was asked, is a different problem than resolution pricing, but it's the one Modem is built for. Modem isn't a cheaper way to resolve conversations, and it isn't a Fin alternative. Whether Fin answers a question for $0.99, your own model key answers it for two-tenths of a cent, or a human teammate answers it for free, the underlying fact worth capturing is the same: a customer asked for something.
“we need SSO before rollout”
“any update on single sign-on?”
“SSO came up twice on this call”
Linear issue, quotes attached
insight report in NotionModem reads Intercom conversations regardless of how they closed, whether Fin resolved them, escalated them, or a person handled the whole thing, and pulls out the requests, bugs, and recurring themes inside them. It dedupes the same ask across Slack, Zendesk, and Intercom into one counted topic with the requesters attached, then files it as a Linear, Jira, or GitHub issue so the resolution cost debate upstream has no bearing on whether the signal downstream keeps flowing to engineering.
That's a different problem than the one this guide has been answering. If the question is "how do we make Fin cheaper," the honest answer is to negotiate volume pricing with Intercom, narrow Fin to the conversation types where it clearly pays for itself, or move those conversations to a BYO-key setup and accept the maintenance cost. If the question is "how do we make sure a resolved conversation about a missing feature doesn't just disappear once it's marked closed," that's the point where a tool like Modem enters, alongside the customer context graph it builds from Intercom, Slack, and the rest of your support stack. We build Modem, so weigh that recommendation with the source in mind; the broader field of tools built for this specific job is compared in best tools to mine feedback from Intercom conversations.
The short version
Fin costs $0.99 per resolution because Intercom is pricing a finished outcome, not a model call, and that price is real value for a business where a resolved ticket is worth far more than a dollar. It's a bad fit for thin-margin, high-volume operations like Parcelwise, where the per-resolution tax can exceed the margin on the thing being discussed. Running your own model key gets the token cost down to fractions of a cent, but only after absorbing the retrieval, escalation, and maintenance work Fin was doing underneath the price tag. Neither choice affects whether the requests inside those conversations get tracked once the conversation closes, which is a separate problem worth solving either way.
