Evaluating CRM data for AI licensing

Choose a workflow instead of a contact dump

A CRM contains several different products: account attributes, opportunity histories, sales notes, email activity, call transcripts, and outcome labels. Start by asking what learning task a buyer is evaluating. Forecasting an opportunity, summarizing a handoff, and drafting a follow-up require different records and carry different disclosure concerns.

A list of named prospects is particularly difficult to justify as training material when those identities are irrelevant to the task. Consider whether a carefully transformed process dataset can demonstrate the same business value using account categories, role labels, and relative timing.

Trace where each field came from

Distinguish employee-entered observations from customer submissions, purchased enrichment, imported partner lists, and synchronized communications. A field being present in your CRM does not show that your company can sublicense it. Keep provenance at the field or source-table level, especially for third-party data that arrived under a separate commercial agreement.

Have the intended use reviewed against collection notices, contracts, applicable privacy requirements, and confidentiality obligations. Exclude sensitive free-text notes and personal details that are not needed for the buyer’s stated task. Removing the obvious contact columns does not address information embedded in activities or attachments.

Avoid accidental label leakage

If the buyer wants to predict an outcome, provide a clear observation cutoff. A note written after a deal closed can reveal the answer even if the explicit closed-won field was removed. Preserve timestamps for stage changes, record edits, and imported activities where available.

Explain what an outcome means in your system. A lost deal might represent a competitor win, a postponed project, a duplicate opportunity, or an administrative cleanup. Treat these as different labels when the underlying evidence supports it. Do not silently recode ambiguous outcomes to improve apparent dataset quality.

• Document changes to stage definitions over time.
• Separate account, contact, activity, and opportunity identifiers.
• Report missing outcomes and unresolved duplicates.

Price a deliverable that can be checked

Use eligible, connected opportunity histories as a candidate unit rather than raw CRM rows. A useful package includes the event sequence, permitted notes, a label dictionary, and a clear mapping between related tables. Quote optional annotation or update work separately from the base license.

For a pilot, agree on a representative slice that includes ordinary records and difficult cases. Require the buyer to state what failed: missing joins, weak labels, irrelevant domain coverage, or insufficient rights. That feedback is more actionable than an unexplained request for another unrestricted export.