An AI tool that can't make up numbers

Sales reps built event proposals by hand. An AI editor is fast, but it can invent prices or policies.

Reps fill a form, see a live two-page proposal, then edit it in plain English ("change the DJ to TBD"). Guards in code run before anything reaches the screen. AI edits are kept separately, with an "Undo assistant edits" button.

• Also for Premier a lead and revenue dashboard (who came in, from which channel, what they booked) on a tested, privacy-scoped reporting layer with a nightly booking-system sync; a database security lockdown; and a corporate-events page that opens with a scroll-driven 3D flight from downtown Austin to the lake.

The proposal builder, dashboard and lockdown are mine. They're built inside Premier's own site and booking system, which their team wrote; the lead attribution data is theirs.

• Guards in code, not in the prompt, because a prompt is only probably obeyed.
• Buffered responses instead of streaming, so the guards run before the rep sees anything.

What went wrong: The audit found three ways a wrong price could get through: the house rate retyped in a different format, a dollar figure tucked into a sentence, and any number in the rep's message, even a year, counting as approval. The first two are closed and now have tests. The third is a known limit of the design, so the rep still checks every price before a proposal goes out.

Reps edit proposals in plain English, and the AI can't put a price on a proposal that the rep didn't type.