How do we decide what data the AI actually needs access to?

Why default integrations over-collect

Selecting fields takes thought; requesting the whole object does not. So the typical integration pulls entire customer records, entire job records, entire invoice records, and stores them, because at build time nobody knew which parts would be needed later.

It feels prudent. What it actually does is convert every future privacy question, retention decision and access review into a bigger problem than it needed to be, permanently.

Work backwards from the decision

Name the decision, then list what it requires. A booking-rate analysis by marketing source needs the call outcome, the source, the date, the job type and the location. It does not need the customer's full service address, their payment history, their equipment records or their account notes.

Run the same exercise for each output the system produces. The list of fields that survives across all of them is your actual scope, and it is usually far shorter than what is currently syncing. This is a design conversation worth having during integration scoping, when changing it costs nothing.

Three questions per field

• What decision changes if this is absent? If you cannot name one, it does not belong in scope.
• Could a derived version do the job? A tenure bucket instead of a signup date, a ZIP instead of a street address, a range instead of an exact ticket. Derived values usually answer the question with a fraction of the sensitivity.
• What obligations does holding it create? Some field types pull in specific handling requirements. Deciding not to hold them is often cheaper than holding them correctly.

Enforce it at the sync layer

Minimization applied at the model call is theater — the data is already in your database. The place it has to happen is the sync: only the selected fields are ever fetched and stored, so the sensitive ones never exist in your environment at all.

That is a real engineering constraint with a real payoff. Data you never collected cannot leak, cannot be subpoenaed from your systems, cannot be exported to a laptop and does not need a deletion path. It is the cheapest control available and the one most often skipped, because it requires making decisions early. Doing it well is a large part of what separates a considered data integration from a connector someone switched on.