Define the unit as a resolved interaction
Support data is most interpretable when it connects the customer’s question, the clarifying exchange, the proposed solution, and the eventual outcome. A pile of isolated replies may contain fluent language without reliable evidence that the advice worked. Decide whether your collection is intended for response generation, classification, retrieval, or evaluation before choosing fields.
Describe the product versions, supported languages, customer segments, and date range. A ticket resolved for an old interface may still be useful for reasoning about troubleshooting, but it should not be sold as current product guidance.
Separate resolution evidence from administrative status
A closed ticket is not automatically a successful resolution. It may have timed out, merged into another case, or been closed after a refund. Preserve reason codes where trustworthy and distinguish agent-confirmed fixes from customer-confirmed outcomes. Be candid about cases where no outcome evidence exists.
Inspect duplicates, canned macros, auto-responses, and internal notes. Repeated boilerplate can make a large collection much smaller in terms of unique instruction. Internal notes may add useful context, but they can also contain escalation details or judgments that were never intended to be shared externally.
Redact the difficult surfaces
Tickets routinely include screenshots, logs, receipts, credentials pasted by mistake, and recordings. Treat these as separate assets with their own review steps. A body-text scanner does not clear a screenshot or a compressed diagnostic bundle. When excluding an attachment, mark that omission so the remaining conversation is not mistaken for a complete record.
Have customer terms and privacy commitments assessed for the proposed license. The FTC’s guidance on AI-related data use emphasizes that existing commitments matter. Ask whether selected examples can be transformed into lower-risk task pairs without preserving unnecessary customer identity.
- Keep the original issue and resolution linked.
- Separate customer text, agent text, and automated system events.
- Record redacted attachments and missing diagnostic context.
Run a bounded evaluation
Offer a reviewed sample across issue categories, not only easy success stories. Let the buyer inspect complexity, completeness, and domain fit using agreed access restrictions. If human labeling is needed, document the rubric, disagreement handling, and whether the labeler saw the eventual outcome.
Define acceptance in terms of eligible conversations, usable attachments, label coverage, and agreed exclusions. If refreshes are requested, account for ongoing privacy review and product-version changes. A support archive becomes a maintainable product only when someone owns those updates after the first delivery.
Common questions
Are internal agent notes valuable?
They can explain escalation and diagnosis, but need separate review for confidentiality, unsupported assertions, and sensitive information.
Should unresolved tickets be discarded?
Not necessarily. They may suit escalation or uncertainty tasks if clearly labeled. Do not present them as verified solutions.