How do you export Intercom data for an AI data license?
Intercom holds the chats, tickets and help articles of a company that talks to customers in a messenger. The export routes differ a great deal: some return only reporting columns, and one returns the full text of each conversation. This guide sets out which is which, from Intercom's own help pages, and what has to be cleaned first.
What an Intercom workspace usually holds
| Conversations | Customer messages and teammate replies in order, from the messenger, email and other channels connected to the inbox. |
| Tickets | Structured support cases with their type, state and the exchange that closed them. |
| Internal notes | Comments that teammates left for each other inside a conversation. |
| Help center articles | Public how-to articles that the team wrote and keeps up to date. |
| Contacts and companies | The people and accounts behind the chats, with the attributes the team stored on them. |
| Reporting datasets | Counts, timestamps and metrics about conversations, useful for describing the set rather than as training text. |
How conversations leave Intercom, route by route
| Tool | What an export gives you | What is missing or needs an admin |
|---|---|---|
| Single or bulk export from the inbox (text or PDF) | The conversation in time order with participants and metadata, one file per conversation. Intercom's article gives a range of 200 to 300 conversations per bulk operation. | Internal notes are not in transcript exports, and card numbers are redacted. A person has to run it, so it suits a sample, not a full history. |
| Conversations API | Full conversation content for a developer who lists conversation IDs and then retrieves each one. The article says the export API accepts a window of up to 90 days per request. | Needs a developer and an access token. The pages we read do not state the output format or rate limits for retrieval. |
| Saved View CSV from the inbox | A snapshot of the columns in a saved view, with links back to each conversation or ticket. | No message text. It needs the export permission and an inbox seat, and the article says very large exports are refused. |
| Dataset export (Reports) | Reporting columns such as IDs, timestamps, attributes and metrics, as CSV, and it can be scheduled. Data is available for up to two years, per the article. | No message or conversation content. |
| Amazon S3 export | Structured conversation and ticket data in JSON, either as a historical run covering up to two years or as periodic runs for new and updated items. | The article states that it contains no transcript, though another line mentions message content, so confirm what a test file holds. It cannot be imported into a new workspace. |
| Articles API | Help center articles listed in pages, newest update first, with a maximum page size stated in the reference. | The page we read does not say what happens to article translations. Contacts and tickets sit under other endpoints that we did not read in full. |
Tool names belong to their owners and are used only to say where records usually live. Export features change by plan and version, so check the current rules for your account.
Picking the route for the text itself
If the goal is the words people wrote, only two routes in the documentation return them for a whole history: the Conversations API, and the inbox export one conversation at a time. The CSV and dataset routes are for counting and describing, and the S3 route is described as structured data without a transcript.
Intercom changes these features over time, and permissions depend on the plan and on how seats were set up. Check the current article for your workspace, and ask the workspace admin who holds the export permission.
Counting chats and tickets, and the dates to write down
The unit to report is the conversation, with tickets counted on their own line. The Conversations API list response carries a total count, which makes a verified figure. A number read from a dashboard is an estimate, so label it that way.
Note the first and last conversation dates, the channels in use and the share that has an internal note. Also record whether conversations were ever bulk-closed or deleted by an automation, since gaps from that cannot be filled later.
What to clean before chats leave
Chat is informal, and customers type their name, email, phone, address and order details straight into the box. Intercom also attaches contact attributes. Replace each with a de-identified label that keeps one customer's thread together, and remove anything that cannot be cleaned reliably.
Customers may also paste screenshots or files. Attachments are screened separately and left out of the license unless they can be cleaned. General information, not legal advice: ask counsel what the privacy notice told customers about their messages.
What stays out of an Intercom set
| Contact details and attributes | Names, emails, phone numbers and stored attributes, removed or replaced with de-identified labels. |
| Payment and health detail typed into a chat | Removed. A thread where it cannot be removed with confidence is left out of the license. |
| Attachments and screenshots | Left out unless they can be cleaned, because images often show an account page. |
| Internal notes about named people | Notes on a customer's behavior or a teammate's performance are reviewed by the owner before any are included. |
A sample, worked through
A software company with a busy in-app messenger counts its conversations through the API total and its tickets from a saved view, and marks both as read from the screen. A developer retrieves conversations in 90 day windows with a read-only token, and the admin exports the help center articles through the Articles API. The set removes contact fields, replaces names with de-identified labels and drops attachments. The owner reads a sample of the cleaned threads and decides what moves on.
Questions owners ask about this
Is the CSV from a saved view enough to license?
It lists conversations but carries no message text. It helps with counting and dating the set, while the words come from the API or the inbox export.
Can the S3 export replace the API pull?
Intercom's article says the S3 files hold no transcript, so it may not give the chat text. Ask for a small test file and read it before choosing.
Do internal notes come out with the rest?
The article says internal notes are left out of transcript exports. Whether the API returns them was not stated on the pages we read, so a developer should check a sample.
How far back can the history go?
Some routes mention up to two years of data and the API route is limited by window size per request. We did not find a published overall retention limit for conversations.
Didn't find your question? Ask us directly.
Sources
The vendor's own pages this guide draws on. Some help centers did not open for us, and where a step rests on a summary and not the page itself, the guide says so. Menus and limits change by plan and version, so check the current pages for your account.
Keep reading
Related kinds of records
- Support tickets: The whole thread, wrong guesses included, with customers replaced by stand-ins.
- Team chat: Slack and Teams threads, where one person must stay one person and direct messages stay out.
- SOPs and knowledge bases: Procedures with their edit history and comments; published help articles add nothing.
- Zendesk data export: How to get tickets with every reply and internal note out of Zendesk, and what customers’ details need first.