Intercom is the conversational helpdesk for product-led teams, and the AI workflow that compounds inside it has three moving parts. An agent reads the inbound conversation, classifies intent, and drafts the reply. A human steps in at the right moment. The rationale behind every classification and handoff is written somewhere durable, so the next conversation gets smarter rather than starting from scratch. Intercom's REST API, the Fin AI Agent, and a thin coordination layer are enough to run it.
The workflow
1. Intent classification on inbound. When a conversation hits Intercom, a webhook fires to your classifier. Most teams use Claude or ChatGPT through a small Zapier or Workato workflow, or post directly to an internal endpoint. The classifier tags the conversation with intent (billing_question, outage_report, cancel_intent, feature_request), urgency, and customer tier. Tags get written back through the Intercom REST API on the conversation object, which makes them visible to every teammate.
2. First-response drafting. Fin AI Agent handles the deflectable layer: known answers from your help center, account lookups, refund status. For anything Fin defers on, a side workflow drafts a reply with Claude or ChatGPT using the conversation history and a retrieved knowledge snippet, then posts it as a note on the conversation. The human teammate either sends it, edits it, or escalates. Ada and Forethought run similar patterns if you want a heavier orchestration layer.
3. Routing and assignment. The classifier output drives assignment rules. Billing goes to the billing team. Outage reports flag a broadcast. Cancel intent routes to retention. This is the same logic covered in escalation routing, and it works identically whether the system of record is Intercom, Zendesk, Freshdesk, Help Scout, Front, or Kustomer.
4. Handoff with context. When Fin escalates to a human, the conversation history goes with it. What does not go with it, by default, is the reasoning: why Fin chose to escalate, what it tried, what the customer said two days ago in a different thread. This is the persistent-state pain, and it shows up in every conversational-AI deployment.
5. CSAT and review. Post-resolution, sample 5 to 10 percent of AI-touched conversations into a review queue. Score the draft quality, the classification accuracy, the handoff timing. Feed the misses back into the prompt or the knowledge base. The audit pattern is in agent audit and compliance.
Worked example: a cancel-intent conversation
A customer opens a chat saying "I want to cancel my Pro plan." The classifier tags cancel_intent, urgency high, tier paid. Fin checks the account, sees an active subscription, and defers. A retention-aware draft posts to the conversation as a note: a one-sentence acknowledgement, a question about the reason, and a pointer to a downgrade option. The conversation routes to retention. The retention teammate sends the draft, gets a reason, logs the outcome. Three weeks later the same customer opens a new conversation about an unrelated bug. The retention rationale from the last interaction needs to be visible, or the new teammate treats them as a fresh ticket.
The persistent-state pain
Intercom holds the conversation. Fin holds the deflection log. Your CRM holds the account. Nothing holds the AI agent's running interpretation across all of them: why it escalated last time, which macro it tried, which CSAT pattern this customer fits. One way to solve this is a workspace like Dock that holds the interpretation layer (rows for classifications, escalation rationale, CSAT analysis) with pointers like intercom_conversation_id linking back to the Intercom record. The conversation stays in Intercom. The reasoning becomes queryable. The setup is covered in Dock for customer support, and the identity hygiene each agent needs is in agent identity lifecycle.
Why it matters
Intercom rewards teams that respond fast and remember context. The AI layer makes the first part trivial. The second part is what separates a deflection bot from a support team that compounds. The same architectural split applies in Zendesk workflows.
Read the full AI customer support playbook to see how the helpdesk-specific patterns connect.
FAQ
Does Fin AI Agent replace the need for a custom classifier? No. Fin handles deflection well. Routing, urgency scoring, and tier-aware assignment usually need a thin classifier on top, written against the Intercom REST API.
How do I prevent Fin from escalating with no context? Write the Fin reasoning to a durable store on every handoff. The conversation history transfers automatically; the deflection rationale does not. According to Intercom's customer service trends research, context loss at handoff is one of the top drivers of low CSAT in AI-assisted support.
Can I use Claude or ChatGPT alongside Fin? Yes. Fin handles the deflectable layer. Claude or ChatGPT through Zapier or a direct API call handles drafting, summarization, and the harder reasoning Fin defers on.
What about Ada or Forethought as an alternative? Ada and Forethought run heavier orchestration. If your support volume is high enough that you need workflow tooling beyond Fin, they are reasonable options. The persistent-state problem is identical across all of them.