Why no LLM touches a dollar in our quotes

The fastest way to lose a charter operator’s trust is to let a language model invent a number. A quote is a contract. A compliance finding can ground a flight. Both have to be exact, and both have to be reproducible by hand. Probabilistic text generation is the wrong tool for either.

Where AI earns its keep

We use a model for exactly one job: reading an inbound charter request — a forwarded Avinode lead, a broker chain, a one-line text — and turning it into structured fields: origin, destination, dates, passengers, aircraft preference. Anything it extracts below 70% confidence is flagged for a human before it moves. That’s the whole mandate. The model never sees a dollar figure or a duty limit.

Where arithmetic takes over

From the structured trip, everything is plain math: billable hours × your hourly rate, your daily minimum, a taxi pad per leg, positioning. Then the tax layer — 7.5% federal excise tax (IRC §4261(a)) and per-passenger segment fees (§4261(b)). The 135.267 duty engine is gated by a set of hand-written test vectors; if a refactor changes a number, a test goes red before it ever reaches an operator.

Deterministic where it counts

“AI-powered” is doing a lot of dishonest work in our corner of aviation right now. The honest version is narrower and more useful: AI where it helps (reading messy email), arithmetic where it counts (money and legality). The day a model “rounds” a rest requirement is the day an operator stops trusting the tool — and they would be right to.