AI Data Guard

The product runs in the window beside this column, and every control in it works.

The product

Ten weeks of AI use at a bank, with the controls on

AI Data Guard sees sensitive banking data going into public AI tools, coaches the person at the moment they paste it, and climbs from logging to blocking one policy rung at a time. What runs beside this column is one deployment of it, ten weeks in and deliberately still near the bottom of that ladder: Alder Creek Community Bank, $850M and 11 branches in Ohio, with every surface dated to Aug 14, 2026.

The bank is fictional and every figure on every surface is a projection of one seeded event log — the tiles, the report and the assistant read the same object, so they cannot disagree about what happened. Nothing you do here is written to any system of record.

Detection is code before it is AI: validators, patterns, the institution’s dictionary. The intercept answers with no API key configured.
No browser extension is installed and none exists here. The Moment simulates the sensor and says so; everything behind it is the product’s own code.
Every figure drills — a tile, an outcome figure or a tool row opens the events it counted, filtered to exactly that set; an event opens the actor, the spans, the policy version, the transform and the destination. Every one of those states is a URL you can paste to a colleague.
The model phrases the weekly brief at build time and answers questions at request time; it produces no figure, and an answer that fails its checks becomes a named refusal instead. The brief carries a byline saying which of the two wrote it — the model, or ordinary code from the same closed pack of figures.