How to export a ticket archive for AI training

Export only what a signed scope asks for: ticket threads in order, timestamps, channel, tags, status and reusable reply text. Strip names, emails, phone numbers, order and account numbers, card digits and attachments before anything leaves your systems. Your helpdesk's own admin documentation decides what your plan can export.

Platform by platform

These notes describe what each platform's records usually hold and what to do with them. They are general on purpose. Check the platform's documentation for formats, limits and which plans include a full export.

Zendesk

Ticket-first. Each ticket carries a thread of public replies and internal notes, plus tags, custom fields and a requester record.

What the records hold: Ticket records with their comment threads, requester and organization records, tags, custom field values, satisfaction ratings where you collect them, and macros your team wrote.
What buyers ask for: The full thread in order with timestamps, who spoke (customer or agent, not their name), tags, status changes, and the macro text agents reused.
Strip before export leaves: Requester names and emails, phone numbers, organization names, anything typed into custom fields that holds an order or account number, and every attachment.
Watch for: Years of history are a lot of records. Check in your admin settings and Zendesk documentation which export your plan includes before promising a buyer anything.

Freshdesk

Ticket-first, much like Zendesk: conversations, private notes, statuses, groups and custom ticket fields.

What the records hold: Tickets with their conversations and notes, contact and company records, ticket field values, and canned responses.
What buyers ask for: Conversation order, the split between customer messages and agent replies, resolution status, ticket type and category fields, and canned response text.
Strip before export leaves: Contact names, emails and phone numbers, company names, and any notes where agents pasted account details.
Watch for: Private notes are where agents write freely, so they hold the most personal detail and the most useful reasoning. Decide early whether they are in scope.

Intercom

Conversation-first. Chat threads mix bot turns, customer messages and human agent replies, often in one session.

What the records hold: Conversations with their parts, user and lead records with custom attributes, tags, and the articles your help center published.
What buyers ask for: Whole chat sessions with turn order, which turns came from a bot and which from a person, the hand-off point, and the help article that answered the question.
Strip before export leaves: User attributes (names, emails, company, plan and location data you track), and any identifiers your product passes into the messenger.
Watch for: Custom attributes are where product data seeps into support data. List them and decide which are safe before anything else.

Help Scout

Mailbox-first. Conversations behave like shared email threads, with notes, tags and saved replies.

What the records hold: Conversations with their threads, customer profiles, tags, notes and saved replies, organized by mailbox.
What buyers ask for: Thread order, customer versus agent turns, tags, which mailbox a conversation came in on (as a category, not an address) and saved reply text.
Strip before export leaves: Customer profiles, email signatures (they carry phone numbers and titles), quoted earlier emails and attachments.
Watch for: Email threads quote each other. Redaction has to catch names and numbers repeated in quoted text, not only the newest message.

Jira Service Management

Requests are Jira issues, with comments split between customer-facing replies and internal ones, often linked to engineering tickets.

What the records hold: Request issues with comments, request types, statuses and workflow history, linked issues, and knowledge base pages if you use them.
What buyers ask for: The request, the customer conversation, the internal diagnosis, the status path to resolution and the link to the bug or change that fixed it.
Strip before export leaves: Reporter and assignee names, customer organizations, internal hostnames, credentials or keys pasted into comments, and links into private systems.
Watch for: Internal IT and engineering detail sits next to customer text. Check that you are not licensing security details along with the conversations.

Gorgias

E-commerce first. Tickets sit next to the customer's store orders, so order data is never far from the conversation.

What the records hold: Tickets with messages across email, chat and social channels, customer records, tags, macros and the order context linked from your store.
What buyers ask for: Message order across channels, the intent behind each ticket (refund, where is my order, sizing), tags, macros and the outcome.
Strip before export leaves: Order numbers, shipping addresses, names, emails, phone numbers, payment details and social handles.
Watch for: Order and address data is the main risk. Treat every number in a Gorgias ticket as personal until checked.

What order should the work go in?

Agree the scope with a buyer in writing: which years, which channels, which fields.
Check your privacy policy and customer terms allow it (see what buyers redact for the field list).
Run the export with the narrowest field list that meets the scope.
Redact, then spot-check a sample by hand. Automated redaction misses names in signatures and numbers in quoted replies.
Transfer only through the method in the agreement, never by email attachment.

If your helpdesk changed platforms, the older years may sit in a different system or in an import with flattened threads. Say so when describing the archive; buyers ask.

Questions about exporting

Should we export the whole archive before talking to a buyer?

No. A buyer first wants a description: platform, ticket count, years, channels, languages. Exporting everything early creates a large file full of personal data that you then have to secure. Export when there is a signed scope, and only the fields in it.

Do internal notes belong in the export?

Buyers value them because they show how an agent reasoned. They also hold the most unguarded personal detail. Many companies leave them out of a first dataset and add them later, after a redaction pass they have checked by hand.

What about attachments and screenshots?

Leave them out by default. Screenshots and attached files are hard to redact reliably and often show account pages, addresses or card details.

Who should run the export?

Someone with admin rights who can also answer for your privacy policy: usually the head of support with your security or legal lead. Keep a written record of what was exported, when and by whom.

Will a buyer do the export for us?

Some offer to work on your side of the fence. Scale AI describes its anonymization running within the seller's systems before extraction. Before granting any buyer access, read exactly what its page and draft contract say it will touch.