Skip to content
Customer service Working integration

Connecting Intercom to AI: who is this visitor and what has already been said

Intercom can be connected to AI. The assistant looks up contacts by email address or by other attributes, pulls in the running conversations with their state and records an internal note on a contact. It can also send a message to a user or lead, and that genuinely reaches that person. The last one only happens on explicit request and after the text has been shown.

Intercom usually sits in the bottom right corner of a website or app, which makes the conversations there short and fast. Someone types a question in the middle of doing something else, the answer has to come straight away, and the colleague picking it up has no time to open three screens first to see who is on the other end. That is exactly where the loss is. An answer goes out without anyone knowing which plan this customer is on, when they were last in touch and what it was about. The Intercom connector in IntraGPT is therefore built around looking things up. The assistant first finds the contact, by email address or by another attribute you keep in Intercom, and then pulls in the conversations belonging to that person. It looks at the latest messages and at the state, because a conversation that has been snoozed means something different from one that is closed. What it can do next falls into two kinds you never want to mix up. A note is internal and stays inside the team. A message is external and reaches the customer. The assistant always says which of the two it means, and for a message it waits for approval.

What the assistant looks up and does in Intercom

Look up contacts

Searching for users and leads works on email address, but equally on the other attributes you keep in Intercom. That is where nearly every question starts: first know who you are dealing with.

Pull in running conversations

The assistant retrieves the conversations that are still open or that belong to a particular colleague. So you get one person's work, or only what is still live, instead of the whole stream.

Weigh the state of a conversation

A conversation in Intercom is open, snoozed or closed. Snoozed means it comes back later, not that it has been handled. The assistant uses that distinction before drawing conclusions about what is still outstanding.

Record an internal note

A note hangs on a contact and is visible only to colleagues. Handy for recording what was found in another system, without the customer noticing anything.

Send a message to a user

A message goes to a user or lead, with a subject, a text and a sender. It genuinely reaches that person, so the assistant first shows what it says and who it goes to.

Where this makes a difference in practice

SaaS

Situation
A user asks a question about a feature in the chat and the support colleague does not know which plan this customer is on.
What the agent does
The assistant looks up the contact by email address, reads the attributes and the earlier conversations and puts the relevant context into the draft answer.
Result
An answer that is right for this account, instead of a general one that has to be corrected later.

E-commerce

Situation
Someone chats about an order, but the order details live in the webshop and not in Intercom.
What the agent does
The assistant finds the contact, retrieves the order from the connected system and records what it found as an internal note.
Result
The next colleague to pick up this conversation does not have to search all over again.

Customer success

Situation
At the start of the week a customer success manager wants to know which conversations are still in her name.
What the agent does
The assistant retrieves her conversations and splits them into what is open and what is snoozed, with the last movement on each.
Result
A work list in which snoozed conversations do not quietly disappear among the handled ones.

Onboarding

Situation
A group of new users has not yet taken an important step in the setup.
What the agent does
The assistant finds the contacts with that attribute and drafts a message per person explaining what is still missing.
Result
Targeted help at the right moment, and nothing goes out before someone has read the text.

Setting it up for the support team

  1. 1

    Connecting through the platform

    An administrator records access to your Intercom workspace in IntraGPT once. If that access is withdrawn, the connector stops at that same moment.

  2. 2

    Looking up and writing set separately

    Looking up, recording a note and sending a message are three separate choices. Many teams first let the assistant only look up and take notes, and add sending messages later.

  3. 3

    Using the assistant

    After that you simply ask who this visitor is, what was discussed before or which conversations are still open. The assistant pulls in what it needs and says where it found it.

  4. 4

    Everything in the log

    Every retrieval and every act lands in the log. For a sent message it records when it happened and at whose request, which is exactly what you want to be able to check.

Connected securely

A chat window is an open line to the customer

Intercom holds personal data, behaviour and conversations the customer does not need to read back twice. That is why the line between internal and external is built in hard here.

Where the data lives: Nederland of de EU

Internal and external are two things

A note is for colleagues, a message is for the customer. The assistant always says which of the two it is recording, and never sends a message without an explicit request.

Who may use the assistant

You decide per role who calls it. An assistant that can send messages belongs with a smaller group than one that only looks things up.

Visible what was retrieved

The log records which contacts and conversations were looked up and what was noted or sent. That makes it explainable afterwards.

The model runs on your own server

Conversation content is processed on our own server in the Netherlands with open-weight models. That content does not go to an external model provider and is not used to train models.

Questions about Intercom and AI

Can I connect Intercom to ChatGPT?

Not in a way where you decide who may do what. You would give a public service access to your contacts and conversations, without roles and without a log of what was looked up or sent. IntraGPT connects your own Intercom workspace, with permissions you choose, and processes the contents on our own server.

Does IntraGPT work with Intercom?

Yes. Intercom is in the connector catalogue and in use at customers. The assistant can look up contacts, pull in conversations, record an internal note and send a message to a user or lead. Which of those are switched on is up to you.

What is the difference between a note and a message?

A note hangs on a contact and is visible only to your team. A message goes to the user or lead themselves and arrives in their Messenger or mailbox. Those are two different acts. The assistant always names which one it is about to perform, and a message only goes out after approval.

Can the assistant reply inside a running chat?

No. It reads conversations and can send a message to a user or lead, but it does not post a reply in the middle of an existing conversation. In practice that works well: the assistant drafts the answer and the colleague pastes it into the chat, or it records its findings as a note for whoever takes the conversation over.

What does a snoozed conversation mean exactly?

Snoozed is a conversation deliberately pushed to later; it resurfaces by itself at the agreed moment. So it is not the same as closed. When you ask what is still outstanding, such a conversation counts, and the assistant keeps that distinction in its overviews.

On which data can the assistant find contacts?

On the email address and on the other attributes you keep in Intercom, including the custom fields you added yourself. That is useful when you want to find a group by plan, country or a status from your own application.

What does an Intercom connector cost?

Connecting is part of the project we run together. What drives the work is whether the assistant may only look up or also record and send, which attributes are involved in searching and which roles may use it.

How long does setup take?

The connection itself costs an administrator little time and can be tested the same day. The lead time sits in deciding what the assistant may do and in agreeing who may have a message sent.

For which sectors

Other connectors

See all connectors

Curious what this connector would deliver for you?

Book a free thirty minute AI call. We look at your systems, the permissions around them and the first use case that saves time or money.

Book a free AI consultation