How to Set Up AI Auto-Responses in Intercom

90% of customers say an immediate reply is essential when they have a support question, and 60% define immediate as ten minutes or less. The average email first response time across industries is still seven to ten hours. The gap is wide enough that the team who closes it consistently has a customer-experience advantage their competitors have stopped trying to match.

This is the case for Intercom AI auto-responses: the cost of being slow is real, and the technology to fix it now sits inside the helpdesk. Setting it up well takes more thought than flipping the Fin toggle. The difference between a noisy auto-reply and an actual auto-resolution is mostly in the design choices we cover below.

TL;DR

• Auto-responses on Intercom are a three-layer stack: instant acknowledgment, AI-grounded answer, clean handoff. Skipping layers two and three caps the resolution rate.
• The native path is Fin in a workflow. Start with the "Let Fin AI Agent answer first" template.
• Prerequisites: clean knowledge base, intent map for the top 20 topics, explicit escalation rules.
• The handoff is the part that consistently gets underbuilt. Pass history, summary, account context, and what the AI tried.
• Track first-true-response time, not first-acknowledgment time. Auto-acks don't count.

What "AI auto-response" actually means in Intercom

There are three layers people lump under the auto-response umbrella, and the distinction matters because each has a different role.

Layer 1: Instant acknowledgment. A workflow message that confirms receipt within seconds. "Got it, looking into this for you." Useful for the customer's sense of being seen, but it doesn't address the question. The first-response-time clock keeps ticking.

Layer 2: AI-grounded answer. Fin (or another AI agent) interprets the message, retrieves relevant knowledge base passages, and composes a grounded answer. This is the auto-response that resolves the conversation. The classification-plus-answer cycle runs in 100 to 500 milliseconds on most modern AI agents, so the customer experiences an instant, useful reply.

Layer 3: Clean handoff. When the AI can't or shouldn't resolve, it routes the customer to the right human with a summary of what was tried, the customer's account context, and the failed attempt. The human starts with full context from minute one.

A well-built setup ships all three. Stopping at layer 1 plus the easy slice of layer 2 is the floor; the leverage compounds in layers 2 and 3.

Prerequisites (do these first)

Three things should be in place before you wire up the workflow.

A clean knowledge base. The auto-response is only as good as the content the model is reading. Knowledge base quality is the biggest predictor of resolution rate, so this is the prerequisite to the prerequisite. The companion piece on building an Intercom knowledge base for AI covers the prep workflow in detail.

Defined intents. Pull the top 20 conversation topics from the last 90 days. For each, decide:

• Should the AI handle this fully? (e.g., "where is my order")
• Should the AI gather context, then route? (e.g., "I want a refund")
• Should this go straight to a human? (e.g., "I want to cancel" for a top-LTV account)

Without this map, the workflow treats everything the same way and the resolution rate caps at the easiest 30%.

Escalation rules. Define when the AI must stop trying and hand to a human. Sentiment (anger, distress). Compliance topics (fraud, legal). Account tier (top-N% LTV always to human). Customer asks for a human explicitly. Without explicit rules, the AI will keep trying to resolve until it has exhausted the customer's patience.

Setting it up with Fin

The native Intercom path, step by step.

Step 1: Open the Fin workflow builder

In the Intercom side nav, go to Fin AI Agent > Workflows. Click + New workflow. Intercom ships a template called "Let Fin AI Agent answer first"; start there for most use cases.

Step 2: Set the trigger

Two triggers work for auto-responses:

• When customer opens a new conversation in the Messenger. Use if you want Fin to introduce itself immediately on the in-app chat.
• When customer sends their first message. Use this for alternative channels (email, WhatsApp, Instagram, SMS), where the conversation starts with a message rather than an interface action.

Pick the channels in the trigger settings.

Step 3: Add the "Let Fin handle" step

This is the step where the AI takes over. Fin reads the customer message, retrieves from your knowledge sources, and composes the response. Intercom documents the Let Fin handle step in detail. The step accepts upstream context from earlier branches in the workflow, so a triage layer can pass through customer attributes Fin uses to personalize the answer.

Step 4: Configure handoff conditions

After Let Fin handle, branch the workflow:

• Fin resolves: end the workflow.
• Customer asks for a human, sentiment turns negative, or Fin's confidence drops below threshold: route to a team or specific agent. Include conversation context, the customer's profile attributes, and a summary of what Fin tried.

Step 5: Test before launching

Run your red-team query set (the 30 to 50 real customer questions you built during knowledge base prep) through the workflow in preview mode. Score the responses. Push the workflow live only when 80%+ are landing correctly.

Optional: Procedures for multi-step flows

For account-aware or action-led automation (refunds, address updates, subscription changes), Fin uses Procedures: structured multi-step plans the AI follows for specific intents. Procedures are configuration-heavy. The work is what unlocks layer 4 and layer 5 automation that pure retrieval cannot handle.

Setting it up with a third-party AI agent

To push past what Fin's retrieval can do natively, third-party AI agents (open.cx and others on the Intercom App Store) layer on top of Intercom and run the conversation while Intercom keeps handling ticketing and the agent inbox. No migration.

The setup pattern is similar:

1. Connect the AI agent to your Intercom workspace via the app or API.
2. Point it at your knowledge sources (help center, public URLs, internal docs).
3. Configure the routing rules in Intercom workflows: route incoming conversations to the AI first, and let the AI hand off to a human team when it escalates.
4. Wire the AI's actions to your back-end systems (order lookup, refund, subscription change) so it can resolve action-led tickets end to end.

The differentiator at this layer is less about which AI engine you pick and more about how deep you wire the action-led integrations. Mollie automates over 60% of conversations for 250,000+ businesses across Europe on this pattern, with the AI handling payment lookups, subscription changes, and refund flows natively against the payments platform's own systems.

Triggers, conditions, and timing windows

A few practical defaults that hold up.

Trigger on first message, not first conversation event. "When customer opens conversation" can fire before the customer has even typed. The AI then guesses at intent without input. "When customer sends first message" gives the AI something to work with.

Use audience rules. Logged-in customers can be matched to account attributes and treated with context. Anonymous chats should get a different first-turn flow (often "ask for email + question" before involving the AI heavily). Email predicates let you run different workflows for different inbound addresses (support@, billing@).

Set a hold before escalating. When the AI hands off, give the customer 30 seconds to read and respond before the human inbox picks up the conversation. Many handoffs resolve themselves when the customer reads the AI's answer carefully.

Don't auto-respond after hours unless the answer is good. A 3am acknowledgment that defers a question to "next business day" is worse than no message. If the AI can resolve, let it. If it can't, decide whether the customer is better off seeing nothing until a human is on shift.

Designing the handoff (the part that gets underbuilt)

Handoff design is the part of an auto-response setup that consistently falls short. The AI handles the easy turn, then drops a confused customer on a human with no context. The human re-asks the questions the AI already asked. The customer's frustration compounds.

A good handoff carries forward:

• The full conversation history. Visible to the human at the top of the ticket. Intercom does this by default; the work is presenting it cleanly.
• A one-line summary of the issue. "Customer is asking about refunding a duplicate charge from May 8. Fin verified the charge but can't process the refund because the account is flagged for manual review."
• Customer account context. Tier, lifetime value, recent issues. The difference between treating customers in batch and treating them in context.
• What the AI already tried. So the human doesn't repeat it.
• The customer's emotional state. A sentiment tag in the handoff helps the human pick the right opening line.

Fin can summarize the conversation and pass attributes natively. Third-party agents typically expose this through Intercom's conversation attributes API. Either way, the handoff package is a deliberate design choice, not a default.

Common failure modes

A short field guide to the patterns that sink auto-response setups.

The "Got it!" auto-acknowledgment trap. Layer 1 alone, no layer 2 or 3. The customer gets an instant ack, then waits hours for a human. CSAT drops because the false positive of "fast reply" makes the slow real reply feel worse.

Confident wrong answers. Layer 2 without the knowledge base prep. The AI quotes outdated policy or improvises on edge cases. Fix the content, then revisit the workflow.

The handoff black hole. Layer 3 missing. AI escalates without a summary; human re-asks; customer re-explains; loop.

Routing every conversation to AI first. High-LTV customers don't want to be triaged by a bot. Carve out an attribute-based exception for top accounts and route them straight to a human queue.

No "I want a human" exit. The AI keeps trying because there's no escape clause. Every workflow needs a phrase-based or button-based path to a human.

Set-and-forget. Workflows drift. Run the red-team query set monthly and fix what's regressed.

What "well-built" looks like

A well-built auto-response setup on Intercom looks like this from the customer's side: they send a message, a useful answer arrives within seconds, and either the issue is resolved or a human is on the conversation within a minute with full context. The customer experiences one response. The team sees the resolution rate climb, the cost-per-conversation fall, and CSAT stay steady or improve.

That's the goal. The pieces are documented, the prerequisites are mechanical, the failure modes are predictable. Build it once carefully, then maintain it.