A boot-removal contract runs on a one-hour clock. The order arrives by email, the boot has to come off within the hour, and a late job goes unpaid. Cornell Transportation sends these orders to a small local towing company, one to twenty a day depending on the season. Whoever was dispatching had to catch each email and retype the plate, make, model, color, location and customer into the dispatch software, every time.
I built a watcher to take that job off the desk. It runs on the email address and dispatch software the company already had, with no new vendor. It reads the order from the PDF, creates the dispatch call, sends the client a confirmation and logs the job.
The flow
A retry could send a second truck to a boot that is already off, so the email is marked handled as soon as dispatch confirms the call. The confirmation to the client repeats the order exactly as the watcher read it, so a wrong plate shows up in an inbox right away, not at the curb twenty minutes later. If the parser is unsure of any field, nothing goes out and the order stays in the inbox. That has never happened on a real call, only on false alarms and inputs nobody expected.
Any hard failure opens an issue with the details automatically. That reaches me as a notification, and I'm usually in a terminal within fifteen minutes. Most months there are none. A second watcher follows each call until the boot is off, then sends the "boot removed" follow-up. Every job is written to an append-only log along the way.
Orders now reach dispatch within minutes, which leaves the driver more of the contract hour. It has run unattended since November 2025. The calls that used to be the most tedious are now just calls.