Shipping AI Features People Trust

IIIt has never been easier to wire a language model into a product, and never harder to do it responsibly. The demo takes an afternoon; the trust takes the rest of the project. For a regional logistics operator we built an assistant that drafts shipment exception notices — and most of the work had nothing to do with the model and everything to do with the system wrapped around it.

Open-ended generation is a liability in an operational tool. We narrow the model’s job to the smallest useful surface: structured inputs, a tight schema for outputs, and validation that rejects anything off-shape before it reaches a human. The model proposes; the system disposes. When the output must match an enum, a date, or a known reference, we check it — and we fail loudly rather than guessing.

• Every model output is schema-validated before it is shown or stored.
• Retrieval is grounded in the customer’s own data, with citations back to the source record.
• Prompts and responses are logged for review under least-privilege access.
• A confidence threshold routes uncertain cases to a human instead of bluffing.

The fastest way to lose a user’s trust is a confident wrong answer. So we spend real design effort on the unhappy path: a clear “I’m not sure — here’s why” state, a one-tap escalation to a person, and never hiding the fact that a draft came from a model. Honesty about uncertainty is not a weakness in the UX; it is the feature that makes people willing to rely on the rest.

We treat AI features as systems with regressions, not magic with vibes. That means an evaluation set drawn from real (anonymized) cases, a dashboard that tracks acceptance and override rates, and a rule that no prompt change ships without re-running the evals. The result is unglamorous and durable: a feature that gets measurably better, and a team that can prove it.