Personalization that feels like a leak is worse than no personalization.

Personalization that feels like a leak is worse than no personalization.

A model can read more history than a team, pick a next offer, and swap the copy while the page loads. That only works if the file is clean, the ask is narrow, and you said what you would do with the data. A retailer once used purchase patterns to infer something private. The model was right enough to be creepy. The brand paid for it.

Use the data for the job you named. Let people opt out. Check the recs for who they skip. Keep a person on the decisions that actually matter.

Four ways this blows up

Free tools often train on what you paste. Customer records in a consumer tab become someone else's pile. Pay for the enterprise version and set it so your data stays yours.

Bias: if the history over-represents one group, the recs will too. Product, credit, even who gets a human agent. Review the output before customers see it. If a model denies a loan or flags an account, a person should be able to say why. You own the outcome, not the vendor. Write down who owns each system and what happens when the chatbot is wrong.

A one-page rule is enough to start

Fair: watch who gets the worse result, and let them appeal. Accountable: a named owner, AI as a draft, a person on the call. Clear: label the bot and the AI offer. Trust: collect what you need, vet the vendor, know if they train on you. Safe: test the ugly cases, watch for drift, have a manual fallback.

Trusted brands keep people. A leak or a biased rec does the opposite. Two or three use cases. A sentence that says what the system does, why, and how to opt out. An owner. Then scale what did not embarrass you.