AI Features in Existing Products

Three things. Identity-document capture and field extraction in the live hotel check-in system, with a manual correction path for low-confidence reads. A natural-language-to-SQL assistant answering questions over a reporting database. And a content generation pipeline with a review step before anything publishes. That is the real list — it does not include autonomous agent platforms or voice systems.

Mostly by not asking it to remember anything. Answers are generated from retrieved records rather than model knowledge, outputs are parsed into a fixed structure instead of free text, and low-confidence results route to a person rather than through to the user. Where a wrong answer would be costly, there is an approval step by default.

Not under the commercial API terms of the major providers, which exclude API traffic from training by default. If that is not enough for your risk position — and for some regulated data it should not be — the same features can run against a locally hosted open-weight model on your own infrastructure. That trades some quality for full data custody.

Because it gets measured before it ships to everyone. A feature goes out behind a flag to a small slice of real usage with token cost and latency logged per call, and the honest comparison is against the manual process it replaces. Sometimes the answer is that a form field and a lookup table would have been better, and that is a valid outcome.