How to Export Intercom Conversations and Support Records

Match the export to the intended task

Intercom holds customer conversations, conversation parts, participant references, assignment events and support metadata. A performance report is useful for queue analysis; it is not necessarily the conversation itself. For AI preparation, identify whether the task needs the original text, a response sequence, internal handoff context or only aggregate metrics.

Download content through the Inbox

The official export guide describes opening a conversation, using its three-dot menu and choosing text or PDF export. For an approved batch, filter the Inbox, select conversations and use the bulk action menu’s export option. Transcript exports exclude internal notes, so inspect what your chosen export actually contains.

Reports → Dataset export provides reporting CSV with selected attributes and time windows. It does not include full conversation text. Use reporting CSV as metadata, not as a substitute for message collection. Check the permission set for the selected method before requesting access.

Collect bulk content with a versioned API

Intercom documents listing conversation IDs and retrieving each conversation’s content. Keep original response JSON, the API version and a log of successful IDs. The preview retrieval reference documents a 500-part cap; verify your selected stable version and test long exchanges. Do not treat a returned parts count as proof that no earlier parts existed.

The export documentation contains conflicting descriptions of S3 transcript coverage. This guide therefore does not promise that an S3 export provides complete transcripts. Validate a known conversation or use the documented per-conversation collection method with explicit completeness checks.

Structure the conversation rather than flattening it

Keep conversation ID, source message, part IDs, author type, timestamps, channel, attachment references and final state. Store contact and company relationships as separate edges. A contact may change companies; decide whether the dataset represents collection-time or event-time relationships rather than silently choosing one.

A synthetic exchange might include a bot answer, a customer correction, an agent response and a reopen event. Preserve those roles separately. The bot’s first answer is not automatically a supervised target, and a closed state does not prove that the customer confirmed a solution. Keep internal notes out of a candidate customer-facing response sequence unless a distinct, authorized task needs them.

Clean and review sensitive content

Normalize HTML in a derived field while retaining source formatting for audit. Remove repeated greetings or signatures only after checking whether they contain product context or author cues the task needs. Review uploads, account identifiers, links, payment-related text and copied logs. Do not rely on product-level redaction to establish that every remaining record is safe to license.

Customer and employee contributions, purchased contact enrichment and third-party content can carry separate restrictions. Export permission is a technical capability; licensing needs an explicit rights decision. Report retained dates, known part caps, missing assets and human-versus-automated coverage in the assessment.

Troubleshoot and assess completeness

Why is my CSV full of metrics but no messages? Reporting exports and content exports serve different purposes. Select a known conversation and compare the methods.

Why does an unusually long exchange begin mid-story? Check the API version’s part limit and the chosen export method; flag truncation instead of inventing context.

Potential applications include support handoff analysis and answer-quality evaluation. Interest depends on coherent context, reliable outcomes and permitted use, not the number of exported report rows.