PREVIEWYou are looking at sample content. Members, companies and threads are illustrative.
Request to joinWe average 38 minutes from inquiry to quote on a typical light-jet trip. Our scheduling system gives us the tail and the numbers fast; the slow part is everything around them: the write-up, the options (tech stop vs. non-stop, catering, ground), and the follow-up email.
Has anyone put an LLM in the middle of this without embarrassing themselves in front of a client? What did you keep the model away from?
Accepted answer
We did exactly this over the summer. The rule that made it work: the model never touches a number. Pricing, times and tail availability come from the scheduling system as structured fields. The model only writes prose around them: the summary of options, the caveats, the follow-up email.
Turnaround went from about 35 minutes to 12. We review every quote before it goes out; the review takes two or three minutes because you are checking wording, not arithmetic.
Two things I'd do again: give it three or four of your best historical quotes as examples of tone, and make it list assumptions at the bottom so the client sees them.
The "list your assumptions" idea is the part I was missing. Thank you, marking this as the answer.
From the maintenance side: whatever you build, make sure it can't quote an aircraft that's in a work order. We had a near miss with a sales tool that read the schedule but not the maintenance status.
Seconding this. Our reconciliation caught a quote for a tail that was AOG. Nobody had connected the two systems.
We tried letting the model draft the options itself and it invented a tech stop that made no sense. Elena's approach of feeding it the options as data and letting it only explain them is the right split.