What makes Intercom easy, and slightly awkward, to automate
Intercom’s API covers contacts, companies, conversations, tickets, help center articles, tags, segments, data attributes and data events. Data events are the useful part for automation: your product can send “invoice_paid” or “integration_connected” and Intercom can segment and message on them.
The awkward parts are practical. The built-in n8n node only handles contacts and companies, so anything involving conversations means raw API calls. Webhooks are configured on a Developer Hub app rather than in a simple settings screen, and the API is versioned, so a workflow written against one version can break when the app’s version changes. Zapier and Make cover the common triggers (new conversation, new lead, tag added) well enough for light workflows.
The AI layer, and why Fin usually comes first
Intercom is unusual among helpdesks in that its AI agent is the product’s center of gravity. Fin answers from your help center and other content, can be given procedures and actions, and is billed per resolution. If you run Intercom and have not tried Fin on your real volume, do that before paying anyone to build a custom agent. We explain the tradeoff in detail in Intercom Fin vs a custom support agent.
Where custom work still makes sense:
- Actions in your systems. Fin can call external endpoints, but someone has to build and secure those endpoints. We build the order lookup, refund eligibility check or plan change API that Fin calls.
- Work outside the conversation. Routing, enrichment, CRM sync and churn scoring happen around Intercom, not inside the chat.
- Analysis. Reading every handed-off conversation and telling you which articles to write is cheap with an LLM and rarely done by hand.
- Multi-tool support stacks. If chat is in Intercom but email tickets are in another tool, one shared agent may be simpler than two vendor bots.
Limits to plan around
- API versioning. Pin the version. Test upgrades against a copy of your workflow before switching.
- Rate limits apply per app and per workspace. Bulk exports of conversation history should page slowly and back off on 429 responses.
- Webhook delivery. Treat webhooks as best-effort. Make your handler idempotent and run a periodic reconciliation job that fetches recently updated conversations.
- Conversation shape. A conversation is a list of parts (comments, notes, assignments, state changes). Flatten it to customer and agent messages before sending it to a model, or you waste tokens on system events.
- Contact types. Users and leads are both contacts with a role field. Lead qualification logic must not overwrite paying users.
How Lumeor typically connects it
For most Intercom clients we leave Fin in place and build two things: secure action endpoints Fin can call, and an n8n layer that receives webhooks, tags and routes conversations, and syncs qualified leads to the CRM. Conversation reads and writes go through the HTTP Request node with the workspace token. Model calls use Claude Sonnet 5.5 or OpenAI’s GPT models with structured output, and results land as tags or internal notes so agents see them in Intercom.
If Fin’s resolution rate stalls and per-resolution fees climb, we scope a custom agent and run it side by side on a slice of traffic before switching. See how to build an AI customer support agent or book a call.