What the AI is not allowed to do.
Every freight vendor now says it has AI agents. Almost none will tell you where their model is forbidden to act. Here is ours, in full.
The one rule
The AI reads. It never decides a value, and it never commits anything that can't be undone.
Everything below follows from that sentence.
The reason it falls exactly there: ordinary software is bad at precisely one thing — turning an unpredictable document or a sentence a human typed into structured facts. That is what the model is for. Everything after that point is arithmetic, rules and lookups, which ordinary code does more accurately, more cheaply, and in a way you can audit. Using the model any wider than that would buy you nothing and cost you certainty.
The line
The AI may
- Read a booking email, a B/L, a commercial invoice, a packing list or a scan — including layouts it has never seen before — and pull out the fields.
- Turn a sentence into a draft: "3 pallets auto parts Shanghai to LA, ready Tuesday, prefer sea."
- Work out what an incoming message is — a booking, a status chase, a document, a rate request, a dispute.
- Suggest which of your parties or ports a name on a document probably refers to, for you to confirm.
- Decide what needs doing next — book, re-quote, raise an exception, reply.
- Draft a status reply or summarise a shipment for a person to review.
The AI may never
- Calculate anything. Chargeable weight, FX, duty, VAT, margin, an invoice total — all computed by ordinary code, every time.
- Decide a deadline. ISF, cutoffs, filing windows — date arithmetic, not judgement.
- Commit anything irreversible. An invoice, a customs filing, a payment, a released B/L — it proposes, you commit.
- Clear its own sanctions hit. A denied-party match is a hard stop for a person, always.
- Decide two names are the same because they look similar. Exact match binds; a resemblance is only ever a suggestion.
- Fill a blank with something plausible. Ever.
Three rules everything obeys
Grounded — it can't invent your data
A party, a port or a rate the AI produces is matched against what's actually in your records. If it isn't there, that's a gap it names — not something it conjures.
Verified — its own confidence counts for nothing
Any value that can be computed or checked, is — by ordinary code, before anything is saved. The model doesn't get to mark its own work.
Gated and named — uncertainty becomes a task, not a guess
Irreversible actions stop for a human no matter how confident the system is. And when something is missing or unclear, it becomes a named item on your queue — never a silent zero, never a default, never a plausible fill.
What that looks like on a Tuesday
This is a real exception from the system. Note what it does: it doesn't proceed, it doesn't guess, and it says exactly why it can't be sure.
Denied-party / sanctions match — human decision required
Raw-document denied-party screen: NOT screened — no denied-party list is loaded, so a sanctioned name in the document cannot be asserted clear.
"Cannot be asserted clear" — not "clear". The difference is the whole product.
Why publish this
Because it's a system of record. It touches your money, your customers' cargo, and declarations that carry legal weight. A wrong invoice or a wrong entry is not a bad user experience — it's an event with consequences.
And because in this industry the filing act attaches to a named, licensed human being. The law already assumes a person is answerable. Software that quietly takes that away isn't helping you.
If another vendor tells you their agent is autonomous, the useful question is the one this page answers: where is your model allowed to decide a number?
Now that you know exactly what it won't do —
bring it one real booking.
And if any of this doesn't match what the system actually does, that's a defect and I want to know: write to me.