· AI

Intercom Charges $0.99 for an AI Support Outcome. The Model Call Behind One Costs Under a Tenth of a Cent.

Renting a packaged AI feature by the outcome made sense when running the same call yourself cost nearly as much. That stopped being true, and most small businesses have not noticed yet.

Intercom's Fin AI agent is priced at $0.99 per outcome: a conversation where the customer confirms the problem is solved, stops asking for more help, or completes a workflow Fin was built to run. You are not charged for a failed attempt, which is a genuinely well-built pricing model. It is also, on a pure compute basis, a markup of somewhere between two and three orders of magnitude over what the model call underneath it costs to run yourself.

My position: once a support task has already cleared the basic test of being worth automating at all, a business with even modest engineering capacity now comes out ahead building the narrow version against a cheap open model instead of renting the packaged outcome. That was not true two or three years ago, when running your own model cost close to what the SaaS markup charged for the same job. It is true now because the price of the underlying call has kept falling while the per-outcome sticker price has not moved much at all.

The case for paying $0.99 anyway

That case is real, and it does not rest on Intercom overcharging anyone. The fee buys retrieval over your actual knowledge base, a workflow engine that knows how to hand off to a human, delivery across email, SMS, WhatsApp, and phone (each billed separately, per Intercom's own pricing page, though inbound email is included), and a service someone else is paid to keep running at 2 a.m. It also buys outcome-aligned billing: you pay for a resolution, not for every message Fin sends while trying to get there. Replacing all of that, not just the model call, is real engineering work, and a business with nobody to build or maintain it is not actually choosing between $0.99 and a fraction of a cent. It is choosing between $0.99 and however many hours of someone's time it takes to build, and keep working, something Intercom built once and sells to every customer at once.

The gap, in dollars

Here is the math. Together.ai's published pricing lists Llama 3 8B Instruct Lite, a small but genuinely usable open model, at $0.14 per million input tokens and $0.14 per million output tokens. A realistic support resolution, with a knowledge base article or two pulled into context and a short back and forth, runs on the order of four to five thousand input tokens and a few hundred output tokens, by my own rough estimate from running comparable workloads. That puts the raw cost of the model call at roughly seven hundredths of a cent, call it a tenth of a cent to be generous. Fin charges $0.99 for the same outcome, on top of either a seat-priced Intercom plan or, used standalone with your existing helpdesk, a minimum monthly commitment that Intercom's own pricing page gives as fifty outcomes as an example floor.

Run five hundred resolutions in a month, an ordinary volume for even a small team's support inbox, and you are comparing roughly $495 in outcome fees to well under a dollar in inference. That is not a rounding error. That is the entire price of the convenience, paid every single month, for as long as you keep renting it.

When buying still wins

Buy still wins in a few honest cases. If your support surface is broad and unscoped, every channel, every failure mode, no single well-defined task, you are not really buying a model call. You are buying an integration and retrieval problem that a packaged product has already solved and you have not. If your volume sits below the minimum commitment, the per-outcome math barely matters because you are paying close to a flat fee either way. And if nobody on your team can build or maintain even a narrow version, the hourly cost of acquiring that skill can easily exceed the markup you are trying to avoid.

Build wins when the opposite holds: one task, narrow enough that you can describe exactly what counts as success, real volume, and someone who can wire a model call to your helpdesk and keep an eye on what it gets wrong. That last part matters as much as the price. A model that resolves cheaply but answers confidently wrong is not actually cheap, it is just a cost you have not counted yet.

This is the exact question behind MojoAI's automation work: not whether a model can do the task, but whether the task is worth automating at the real price of running it yourself, once you have been honest about what either path costs in engineering time. If you want a second opinion on which side of that math your task falls on, that conversation is free.

Sources

References used in this article. Links also appear alongside the relevant claims.

Let's talk

Let's make it happen.

You don't need a polished brief. A couple of lines about where you want your company to go is plenty, and we'll come back with what we'd tackle first.

We scope the work around your goals and timeline before quoting anything, so you know exactly what you're getting.

LocationBoca Raton, Florida
CoverageSouth Florida + remote nationwide
Status Now accepting clients