Intercom built its support bot to say it does not know

The conversation

On 13 June 2023, Intercom's Bob Wallis wrote that the company's new support bot, Fin, powered by a mix of large language models including OpenAI's GPT-4, was now available to everyone. Intercom's own announcement described Fin as ingesting a company's support content to resolve queries, and set out three design goals: Fin should be trustworthy, controllable and seamless. On trustworthiness, Intercom wrote plainly about the risk it designed against, calling hallucinations the fly in the GPT soup when it started working on Fin. A little over a year later, a second Intercom announcement, dated 10 October 2024, introduced Fin 2, adding a Knowledge Hub able to learn from internal content, external websites, PDFs or databases, and a pricing model charging 99 cents per resolution while free to use when it cannot answer.

What the documents show

Intercom's own 2023 post is explicit that Fin's answers are bounded by design: the company says it limited Fin to answering based solely on a business's own support content. The same post states the fallback in Intercom's words: Fin has no problem saying it does not know, and swiftly passes a question to a support team member if it cannot find the answer in that content. The 2024 post adds a vendor-reported figure, stating Fin achieved an average resolution rate of 51 percent, with an accuracy rate of 99.9 percent, Intercom's own claim, over a period and population the post does not further specify, not an independently audited measurement.

The system boundary

Fin's documented boundary is content restriction plus a named fallback: the bot is designed to answer only from material a company has chosen to give it, and to hand off rather than improvise when that material does not cover a question. The 2023 post frames this as a deliberate rejection of what it calls a first-order solution of reskinning an existing large-language-model chat experience for customer service, on the grounds that an unconstrained chat interface fails to account for the unique needs of customer service teams. Fin's current site, retrieved 16 September 2026, shows the product has since been given its own domain and now advertises that it hands off to a human team across multiple helpdesk platforms, not only Intercom's own.

Where it fails

The 51 percent and 99.9 percent figures in the 2024 announcement are Intercom's own reported statistics, not a third-party audit, and the post does not state the exact time period, customer sample or definition of accuracy behind them.

• Over what period, and across what mix of customers, was any vendor-reported resolution or accuracy figure actually measured?
• What happens to a conversation the moment Fin hands it to a human team, and how much context actually transfers?
• Has the support content Fin draws from been audited for the kind of contradictions or gaps that would make a trustworthy-sounding answer wrong anyway?

Intercom's own materials describe a system built to say it does not know rather than invent an answer; whether that boundary holds in a specific deployment depends on how well a company's own support content actually covers what customers ask.