Major

Intercom AI Agent: Build a Triage-and-Draft Workflow You Control

An Intercom AI agent can classify conversations and draft replies, but production support needs durable state, approvals, and auditability. Here is the governed workflow to build.

Jose GironUpdated
Intercom AI Agent: Build a Triage-and-Draft Workflow You Control

What does an Intercom AI agent do?

An Intercom AI agent reads incoming conversations, classifies intent and urgency, pulls account and knowledge-base context, drafts a reply, and routes the case to the right person. The version you can trust in production keeps the draft behind a human approval step and logs every decision.

Only two of those jobs need a model: classifying messy customer language and drafting the reply. Fetching the thread, looking up the plan tier, applying the routing table, and writing the audit record should run identically on the ten-thousandth conversation as on the first. That work belongs in code.

This is where Major fits. Major is the enterprise platform where agents build the software they run on. The agent reasons once about how your billing triage should work, builds a governed app for the repeatable parts, and runs that app from then on. The model is left with classification and drafting. This guide is for support and ops leads who want that workflow built on Intercom without handing the inbox to a black box.

Four Intercom agent workflows that hold up

Inbound triage. The agent tags and assigns new conversations by intent and urgency. Humans own the routing table and review misroutes weekly. The agent never closes a conversation on its own.

Draft-for-approval replies. For how-to and billing questions, the agent writes a draft and a rep approves, edits, or rejects it. Nothing reaches the customer without that click.

Risk escalation. Refund requests above a threshold, legal language, security incidents, and cancellations go to a named owner with context attached and no draft. Handoff speed is the whole job.

Customer-signal detection. The agent flags repeated complaints about one feature, or a spike from one company, and posts the pattern to the account owner.

Autonomous sending is fine for narrow, measured, low-risk intents like "where do I reset my password." Anything ambiguous or involving money gets a person. Classification will be wrong some of the time, and the approval step catches that error before it is emailed. Measure classifier accuracy on a labeled sample of your own inbox before you loosen any approval rule.

For the same pattern on another helpdesk, see our Zendesk AI agent walkthrough.

What the Intercom API gives you

Auth. For your own workspace, create a private app in the Intercom Developer Hub and use its access token as a bearer token. A triage-and-draft agent needs read conversations, write conversations, read admins, and tag scopes. It does not need the delete scope, which Intercom marks as irreversible.

Conversations. GET /conversations/{id} returns the full thread. POST /conversations/{id}/reply accepts an admin reply with message_type set to comment (customer-visible) or note (internal only), plus the author's admin_id. That note type is the hinge of the approval design. The draft lands as an internal note first, and only an approved draft is sent as a comment. Every request needs an Intercom-Version header.

A draft note is one call: POST /conversations/215472/reply with {"message_type": "note", "type": "admin", "admin_id": "3156780", "body": "<p>Draft pending approval...</p>"}.

Triggers. Subscribe to the conversation.user.created and conversation.user.replied webhook topics and verify the X-Hub-Signature header. Intercom expects a response within 5 seconds, so acknowledge fast and do the work asynchronously.

Retries. A retried webhook means the same event can arrive twice. Key every run on the conversation part ID and refuse to draft twice for the same part. Watch X-RateLimit-Remaining and back off with jitter on a 429. Check current limits for your workspace before you depend on them.

Worked example: the billing triage-and-draft agent

A customer writes into the Billing inbox: "I was charged twice this month, can you fix it?"

  1. The conversation.user.created webhook arrives. The app verifies the signature, returns 200, and records the event keyed on the part ID.
  2. The app fetches the thread and the company attributes (plan, seat count, account owner).
  3. The model classifies the message: billing, duplicate charge, urgency medium, confidence high.
  4. The app checks the routing table. Duplicate charges under the refund threshold go to draft. Anything above it, or any mention of a chargeback, goes to escalation.
  5. The app pulls the duplicate-charge article from the Notion billing knowledge base, and the model drafts a reply grounded in it.
  6. The app writes the draft to Intercom as an internal note and posts an approval card to the billing Slack channel.
  7. A rep approves. The app sends the text as a comment and logs who approved, when, and what changed.

Two steps out of seven use the model. The rest is code that behaves the same every run. Each conversation also gets a row: event ID, classification, confidence, route, draft, approver, final text, timestamps. Heavy rep edits become a diff you can review. A misroute shows which rule fired.

This pattern does not cover multilingual drafting, Fin configuration, or ticket (as opposed to conversation) workflows. Those change classification and routing enough to need their own treatment.

Build the Intercom triage agent on Major

Here is the build, piece by piece.

  • Connectors: Intercom for threads and replies, Notion for the knowledge base, Slack for approval cards. Each connection uses its own scoped credential.
  • The app: a managed database table holds one row per conversation (event, label, route, draft, approver, final text), plus a review view for misroutes and heavy edits.
  • The agent's judgment: classify intent and urgency, and draft the grounded reply. Nothing else.
  • Deterministic code: webhook verification, dedupe on part ID, the routing table, the Slack approval gate, the write-back to Intercom.
  • Trigger: the Intercom conversation event starts a workflow run. No polling.
  • Governance: the credential proxy keeps the Intercom token out of every prompt. RBAC limits who edits the routing table. Every approval and send is an audit row. See AI agent governance for how those controls apply where the agent acts.

Start building free and describe the workflow in a sentence.

An agent that re-reasons every conversation pays tokens for the dedupe, the lookups, and the routing on every run, and leaves you guessing why it chose a route. On Major that work is an app your team can read and change, so cost is front-loaded and then flat. If a rep reads every message in a small inbox, you may not need an agent yet. Once triage is a real queue, start with one inbox, keep approval on for every intent, and loosen it only where the audit log shows the drafts have earned it.

Start building on Major

Build the Intercom triage-and-draft agent as a governed app: approvals, state, and audit included from the first deploy, with Intercom, Notion, and Slack connected through scoped credentials.

Related articles

Frequently asked questions

What does an Intercom AI agent do?
An Intercom AI agent reads incoming conversations, classifies intent and urgency, retrieves account details and knowledge-base articles, drafts a grounded reply, and routes the case to the right rep or queue. In production the model handles classification and drafting, while code handles fetching data, routing rules, approval, and logging.
Can an Intercom AI agent send replies automatically?
It can, but only for narrow, low-risk intents you have measured, such as password resets. Ambiguous questions, billing disputes, refunds, legal language, and cancellations should go through human approval or straight to an owner. A safe design posts the draft as an internal note and sends a customer-visible reply only after a rep approves.
How do you build an AI agent for Intercom?
Create a private app in the Intercom Developer Hub with minimal scopes, subscribe to conversation webhooks, verify the signature, and dedupe on the conversation part ID. Store each conversation's state in a database, use a model to classify and draft, post drafts as internal notes, require approval before replying, and log every action. Major builds this as a governed app.
What tools can an Intercom agent connect to?
Any system with an API. The worked example uses a Notion knowledge base to ground drafts and a Slack channel for approval cards. Each connection needs its own scoped credential, such as read access to selected Notion pages and post access to one Slack channel.