Intercom Fin: a Claude-powered support agent resolving up to 86% of questions

• Intercom — Claude customer story (Fin AI Agent) — Anthropic
• Introducing Fin: Intercom's breakthrough AI chatbot, built on GPT-4 — Intercom (блог), 2023-03-14
• The Fin AI Engine™ — официальное описание архитектуры — Intercom Help Center
• From resolutions to outcomes: Evolving how Fin delivers value — Intercom (блог), 2026-03

Background

Intercom is a customer support and communications platform used by more than 25,000 companies. In 2023 the company made the most radical bet in its history: it declared itself 'AI-first' and began rebuilding the entire product around an AI agent that resolves customer inquiries itself rather than just assisting a human agent.

Intercom moved first in the industry. Fin was announced on March 14, 2023 — the same day GPT-4 was released — as the first customer service AI bot built on that model. The timing was no accident: the previous model generation (GPT-3.5) hallucinated too often to be trusted with a customer conversation on behalf of a brand, and Intercom publicly named GPT-4's reduced hallucinations as the condition that made the product possible.

The second pioneering decision was the monetization model: Intercom priced Fin at $0.99 per successful resolution — money charged only when Fin fully resolves a customer's issue with no human involved. For a SaaS market accustomed to per-seat pricing this inverted the economics: for the first time, the vendor took on the quality risk of its own AI.

Later Intercom changed the model foundation too: Fin moved from GPT-4 to Anthropic's Claude models. VP of AI Fergal Reid explained the switch plainly: 'We switched to Anthropic because its Claude models are the best at providing high-quality answers to our customers at scale.' That migration story is rare public evidence that a product like Fin can be model-agnostic: the value lives in the application layer, and the underlying model is replaceable.

Problem

Support is expensive, and its cost grows almost linearly with the business. Intercom's own framing of the pain: 'Customer support is very expensive to deliver. Support reps spend a lot of time searching for authoritative answers, even for questions they've seen before' (Fergal Reid, VP of AI). A large share of an agent's day is not helping the customer but re-finding knowledge that already exists.

Classic scripted chatbots did not solve the problem — they often made it worse: rigid decision trees frustrated customers, failed on rephrasings, and escalated to a human almost anything outside the prepared branches. First-generation 'automation' lowered satisfaction while barely lowering the load.

Response time remained the main source of dissatisfaction: per the Anthropic case study, before Fin customers waited around 30 minutes for an answer — versus seconds after. For SaaS companies, where support directly affects subscription retention, every such queue is churn.

Intercom as a vendor had a specific problem of its own: you can only ship an AI agent that answers customers on behalf of other brands if hallucinations are tightly controlled. A mistake by Intercom's bot damages two companies at once — the client and the platform itself. That requirement — 'resolution without invention' — shaped the product's entire architecture.

Solution

Fin answers from the client company's knowledge base and data, works in more than 45 languages, and is built into Intercom's support funnel: it resolves routine questions itself and hands complex ones to a human agent with the accumulated conversation context. Onboarding is deliberately simple: Fin learns from the company's existing help articles and switches on with a toggle, without months of setup.

Under the hood is the patented Fin AI Engine™, which Intercom describes as a six-layer architecture purpose-built for customer service. The official documentation discloses three key processes. First, query refinement: before generating anything, the engine reformulates the customer's question using dialogue history and account context, because LLM answer quality depends directly on input clarity. Second, response generation via a bespoke, enhanced RAG architecture: relevant fragments are retrieved from connected knowledge sources and fed to the model with the query. Proprietary models — fin-cx-retrieval and fin-cx-reranker, trained specifically on customer service tasks — work here. Third, validation: before sending, the answer passes accuracy and safety checks, and low-confidence dialogues are escalated to a human.

Generation runs on Anthropic's Claude models. Intercom staff ML scientist Pedro Tabacof named three reasons for the choice: answer quality, reliability at scale, and the quality of the strategic partnership. The company does not disclose specific model versions — a deliberate stance: Fin is sold as a product with a resolution metric, not as a 'wrapper around model X'.

The product is organized around four capabilities. Knowledge: Fin learns the entire product knowledge base. Behavior: it follows company policies and tone. Actions: it executes tasks — processes refunds, changes account data; it does, not just answers. Insights: it generates analytics on inquiry topics. Actions is what turns Fin from a 'smart FAQ' into an agent: with every release a growing share of inquiries is closed by an action rather than text.

The product economics evolved too. In March 2026 Intercom moved from paying 'per resolution' to paying 'per outcome': besides a full no-human resolution, a billable outcome now includes a 'procedure' — a case where Fin gathered the context and did the heavy lifting, but final confirmation stayed with a human (for example, in transaction disputes or subscription changes where compliance requires a human decision). Intercom's phrasing: 'success stopped being binary'. The company also stated its principle publicly: 'trust is the currency of AI', and pricing must stay fair and predictable.

Result

Out of the box Fin resolves 51% of inquiries on average; tuned for a specific company, resolution reaches 86% — a gap that itself shows knowledge-base quality is decisive. Response times drop from ~30 minutes to seconds. Per Intercom's blog in 2026, more than 7,000 teams use Fin, the average resolution rate across customers has reached 76% and keeps growing month over month — even as the share of complex queries rises; the product pages state over a million conversations per week.

Published customer results from the Anthropic case study give three different frames of scale. Synthesia (a startup): in 6 months Fin closed over 6,000 conversations, saved over 1,300 support hours, and self-serve reached 87%. Fundrise (a growing fintech): over 50% of volume automated within 3 months at 95% response accuracy, with seasonal case peaks roughly halved year over year. Lightspeed (enterprise): resolution up to 65%, AI involved in 99% of conversations, and agents closing 31% more conversations per day — AI removes the routine even from dialogues a human still leads.

It is important to frame the metrics: a 'resolution' in Intercom's reporting is a conversation fully closed without human involvement (since March 2026 — an 'outcome', which also includes procedures with final human confirmation). The 51%, 76%, and 86% figures are vendor and customer data, not an independent audit; resolution depends heavily on the quality of a specific company's knowledge base and its inquiry mix.

Also telling is how the resolution ceiling depends on the type of business: startup Synthesia reached 87% self-serve, while enterprise client Lightspeed tops out at 65%. In our view this is not a difference in tuning quality but a natural difference in product and inquiry complexity: the more complex the product and the higher the cost of error, the more cases still require a human — and the more it matters that AI accelerates those dialogues too (hence the +31% conversations closed by Lightspeed's agents).

In our view, Intercom's main contribution to the industry is not the resolution percentages but the economic model. Outcome-based pricing ($0.99 per resolution, later the outcome model) aligns vendor and customer incentives: Intercom earns only when its AI actually works, so it cannot afford a 'chatty bot' that simulates activity. This model has become the de facto industry standard for support AI agents — competitors across the market are copying it.

The second transferable takeaway: Fin's migration from GPT-4 to Claude shows that for an application company the base model is a replaceable component, while the competitive advantage lives in the layer above it — retrieval models, validation, action integrations, and data from millions of real dialogues. For companies building their own agents, that argues for architectures where the model can be swapped without rewriting the product.

Lessons learned

1. Sell (and measure) resolution, not 'answers': Fin is judged by the share of fully resolved inquiries — an honest metric a chatty bot can't game.
2. Outcome-based pricing aligns incentives: a vendor that earns only on successful resolutions must invest in quality, not demos.
3. The 51% → 86% gap is company-specific tuning: knowledge-base quality matters more than the model.
4. Reliability architecture is a pipeline, not a single model call: query refinement → RAG with specialized retrieval models → answer validation before sending.
5. The base model is a replaceable component: Fin survived a GPT-4 → Claude migration without changing the product; the advantage lives in the application layer and the data.
6. An AI agent also raises human productivity: Lightspeed agents close 31% more conversations daily — AI removes routine instead of competing.
7. The next step after Q&A is actions and 'procedures': Intercom already bills cases where AI does the heavy lifting and a human only confirms the final step — that is the path from ~50% resolution toward 80%+.

Frequently asked questions