Data types

Can you license company email for AI training?

A practical review of email ownership, thread quality, attachments, and safer sampling before discussing a dataset with a buyer.

Sources and guidance · Reviewed October 2026

Start with a bounded mailbox collection

A company inbox is not a single clean asset. It mixes employee writing, customer messages, forwarded documents, automated alerts, signatures, and correspondence received under confidentiality expectations. Begin with one clearly defined business workflow, such as resolved product questions, rather than offering every mailbox the company can export.

Write down the date range, departments, languages, approximate thread count, and original collection purpose. Explain what a buyer could learn from the sequence: issue recognition, clarification, resolution, or escalation. That is a more useful description than a compressed archive size.

Review both sides of the conversation

Account administration is not proof that you can license every message. Separate employee-authored material from third-party contributions and attachments. Ask your legal reviewer to examine employment agreements, customer terms, confidentiality commitments, and the intended training use. The FTC has specifically warned AI companies to honor privacy and confidentiality promises.

Exclude personal correspondence, privileged legal discussions, authentication information, payment details, and sensitive employee matters from the proposed scope. A narrow collection with a documented decision trail is easier to evaluate than a broad collection accompanied by sweeping assurances.

Preserve useful structure without unnecessary identity

A buyer evaluating conversation quality needs ordering and context, not necessarily real names or email addresses. Consider consistent role labels, relative timestamps, stable thread identifiers, and explicit attachment references. Remove quoted duplicates carefully so the reply still makes sense. Keep the unmodified source in a restricted environment while preparing a separate review copy.

Test redaction on more than message bodies. Sender fields, subject lines, filenames, signatures, quoted replies, calendar links, and document previews can disclose identities or secrets. Record how each transformation changes the usefulness of the conversation.

  • Count unique threads separately from messages.
  • Report missing attachments and incomplete conversations.
  • Distinguish human replies from templates and automated notifications.

Use a staged buyer review

Begin with a schema, aggregate coverage statistics, and a small sample already cleared for disclosure. Do not upload raw mailboxes to an unsolicited intake form. Agree on sample access, evaluation purpose, retention, and whether any model training is permitted before transfer.

If the sample demonstrates value, define acceptance using complete eligible threads and documented exclusions. Specify whether later customer deletion requests can be propagated, who handles them, and what happens to derived artifacts. A paid pilot can test delivery effort without assuming the entire archive is licensable.

Common questions

Can I sell my work inbox personally?

Do not assume so. An individual’s access to a work account does not establish ownership or permission to commercialize company and third-party correspondence.

Is deleting names sufficient?

No. Distinctive facts, signatures, attachments, and conversation context may still identify people. Review the full export and the intended recipient context.

Sources & further reading

FTC: AI companies must uphold privacy and confidentiality commitments ↗
Use this as a preparation framework. It is not legal advice, a valuation or a promise that a buyer will accept your data.