← The journal

From email chaos to one inbox: how AI reads your orders

Orders arrive as emails, PDFs, and phone calls you type up later. Here is how the inbox reads each one, drafts the order at 99% confidence, and keeps a human on the approve button.

Your last order today probably did not arrive through a form. It arrived as a forwarded email with a trip sheet PDF attached, a tail number in the subject line, and the real delivery time buried in the third paragraph. Someone typed it into your system at 9:40 p.m., after the kitchen closed.

We watched caterers do this for months before we built anything. The pattern was the same everywhere: the order exists twice, once in the message and once in your system, and every retype is a chance for the two to disagree.

How orders arrive today

Across the kitchens we studied, a typical week of 41 orders broke down like this:

  • Forwarded emails, often three replies deep, with the details scattered.
  • PDF attachments: trip sheets, request forms, screenshots of both.
  • Phone calls typed up after the fact, sometimes hours later.

Retyping a single order takes eight to twelve minutes when nothing goes wrong. When something does, a missed nut allergy, a delivery time read as an arrival time, the cost is not minutes. It is a refund, an apology, and sometimes the account.

What the inbox reads

Connect your catering address and every message gets read the same way, whether it is plain text, a PDF, or a note typed after a call. The reader extracts the fields that make an order an order:

  • Items and quantities, matched to your own menu.
  • Tail number, like N604VZ, and the operator behind it.
  • FBO and airport, so a TEB order never routes to HPN.
  • Delivery time, normalized to local time at the field.
  • Special notes: allergies, crew meals versus pax, plating requests.

From those fields it builds a draft order and scores its own confidence on each one. A request from Bluewake Jet Group that names the items, the tail, the FBO, and the time drafts at 99%. A vague reply that says "same as last Tuesday" does not.

A human stays on approve

RULE OF THUMB

When the model is not sure about a field, it flags it and leaves it blank. A blank you can see beats a guess you cannot.

No draft becomes an order until someone in your kitchen taps approve. High-confidence drafts take one glance and one tap, about 20 seconds in practice. Low-confidence fields are highlighted, so you check exactly the parts worth checking instead of rereading the whole thread.

Replies become updates

The second half of the problem is the follow-up. "Make that wheels up at 0700" used to arrive as a fresh email and become a duplicate order, or worse, a second entry nobody reconciled. Reply detection recognizes that a message belongs to an existing order and drafts an update to that order instead: same record, new time, full change history.

The practical takeaway: you do not need to change how your clients order. Let them keep emailing, attaching, and calling. Put a reader in front of your inbox, keep your best people on the approve button, and let the retyping end. That is what the PlaneCater inbox is built to do.

Keep reading
404 · No such entry

That post isn't on the record.

It may have moved, or the address was mistyped. Everything we've published lives in the journal.

Back to the journal