Connect Intercom to an AI agent

1. What a waiting conversation costs

A live chat promises the visitor something the moment it appears: someone is there. When nobody is, the promise turns against you — the silence becomes information about how serious you are, obtained before the person is even a customer.

The calculation people make then is always the same, and it never comes out right:

• Covering the hours costs you whole posts. Holding a chat from 8 a.m. to 8 p.m., six days a week, is not one person: it is a team, with its holidays, its sick leave and its cover. For a company starting out, that post will never be opened.
• Doing it yourself costs your attention. Answering between two meetings means answering badly — and the conversation stays open, half handled, which is worse than not having opened it.
• Removing it costs the enquiries you will never get. Many end up hiding the chat. That is honest, and it also removes the only opportunity a hesitant visitor had to ask their question.

An agent does not magically resolve this dilemma: it moves it. What no longer has to be paid for is presence; what remains to be done is deciding what goes out alone and what comes up to you.

2. What the agent does in Intercom

Intercom's official connector covers conversations, which are the central object of the product.

• View the conversations. Read the full thread, including what was said in previous exchanges with the same person. That is what avoids the most hated question in support: “could you explain your problem again?”
• Open a conversation. Getting back in touch when something on your side warrants it — a promised reply, a fix delivered, information that was missing.
• Reply in the thread. Write and send the reply within the conversation, under the rules you have set.

The real value comes from what the agent knows at the moment it replies. A useful answer needs context that lives elsewhere: the state of a payment, the customer record, internal documentation. An agent connected to Intercom alone replies politely; an agent that sees the rest replies correctly.

3. What it does not do

• It does not invent an answer to fill a gap. When the information does not exist in what it has access to, the expected conduct is to say so and hand over — not to produce a plausible sentence. That is the costliest failing of automated support, and the hardest to make good.
• It makes no commercial commitment. A refund, a goodwill gesture, an exception: these are decisions, and they remain yours, even when the agent knows how to word them.
• It does not follow instructions it reads in a message. A customer — or someone claiming to be one — can write text designed to divert the agent. The content of a conversation is data to process, never an instruction to carry out. That is a rule of the product, not an intention.
• It does not close a conversation left without a conclusion to improve a statistic. A filed conversation is not a resolved conversation, and confusing the two amounts to lying to yourself.

4. The constraint specific to Intercom: a reply goes to the customer

This is what distinguishes Intercom from an internal tool, and it deserves to be said plainly. A message sent in a conversation reaches the person, often doubled by an email notification. It cannot be cancelled, it cannot be quietly corrected, and a fix sent three minutes later is visible too.

The practical consequence is simple: keep approval active until you have put your agents to the test. Every reply is then submitted to you before it leaves. It is annoying for the first few days, and that is exactly the point: you read what the agent wanted to write, you correct the tone, and you learn what it can do — on real cases rather than examples.

You then lift that requirement by degrees, and not all at once. The progression that works:

• first the purely factual replies — an opening time, a procedure, the status of an order;
• then the frequent cases you have seen enough examples of to know what they are worth;
• never goodwill gestures, formal apologies or anything touching on a dispute.

This setting lives in nullbot, agent by agent, and goes back as easily as it comes off. Putting it back after an incident is not an admission of failure: it is the normal use of the tool.

5. Connecting, and naming the workspace

Connecting is an account authorisation: you log in at Intercom, the list of requested permissions appears, you accept. No application to create, no developer account, no key to paste.

The scope is named at the level of the workspace. That is the division Intercom itself uses: a company that separates its consumer support from its business support has two workspaces, and the agent only has access to the one you designate.

It is also the right way to start cautiously when you have only one workspace: connect, leave the agent read-only, and watch what it would have replied for a few days before opening up writing.

6. What it changes over time

Support is the area where regularity matters most and holds up least. A good week goes unnoticed; a bad one gets retold.

• Week 1. Every reply goes through validation. What you mainly discover is which questions recur — almost always fewer than you think, and almost never the ones you had anticipated.
• Month 3. Repetitive questions go out on their own, particular cases come up. Evenings and weekends stop being holes: a useful reply goes out at 10 p.m., and what requires judgement is waiting for you in the morning, already documented.
• Year 2. The accumulation becomes product intelligence. What people ask ten times a week says what is missing from your offer or from your pricing page. That observation does not exist when every conversation is handled then forgotten.

No satisfaction rate is promised: it depends on your product as much as on your replies. What is guaranteed is that nobody waits because it was Saturday — and that is what delegated support must guarantee first.

7. Frequently asked questions

Going further

Read next: connecting Gorgias, AI agents for customer support, or all available connectors.