Human-in-the-loop AI for load booking: why $150k freight can’t hallucinate

There is a version of freight automation that sounds impressive in a conference pitch: an agent that reads the email, prices the lane, picks the carrier and books the truck, with nobody watching. We do not build that, and this note explains why.

Two different kinds of system

A language model is probabilistic. Ask it the same question twice and you may get two answers, each delivered with equal confidence. Freight execution is the opposite: a rate confirmation is a commitment, a pickup number is either right or wrong, and a truck sent to the wrong dock is a real cost to a real customer.

Putting a probabilistic system in direct control of a deterministic commitment is how you get dangerous hallucinations booking $150,000 loads unsupervised. The engineering problem is to get the benefit of the first without letting it touch the second unchecked.

Gate one: every field has a source

The first control is traceability. Every extracted value — rate, commodity, zip code, weight, special equipment instruction — is stored with a reference to the exact line of the original email or PDF it was read from. Click the field and the source is highlighted.

This does two things. It makes review fast, because the reviewer checks a highlight instead of re-reading a thread. And it makes fabrication structurally hard: a value with no source span is not a value, it is a gap, and gaps are flagged.

Gate two: deterministic validation

Extraction is followed by checks that have nothing to do with language models. Do the postal codes exist and match the stated cities? Is the weight legal for the equipment? Is the rate within a sane band for the lane? These are rules, and rules give the same answer every time.

Gate three: thresholds your brokerage sets

The last control is policy. BrokerOS proposes quotes, matches carriers and extracts rate cons, but final tenders and rate agreements require broker sign-off or strictly defined threshold policies. A brokerage might allow a repeat lane with a trusted carrier to confirm automatically inside a margin band, and require a human for everything else.

The important property is that the brokerage writes those rules, can read them, and can change them. Nothing is a black box.

Why this is the faster design, not the slower one

It is tempting to see human review as friction. In practice the opposite is true. A system that nobody trusts gets double-checked end to end, which is slower than doing the work by hand. A system that shows its sources and enforces its own limits gets a one-click approval.

Computers excel at parsing thousands of email threads and reconciling zip codes. Humans excel at carrier relationships, nuanced exception handling and commercial judgment. Keeping each on its own side of the gate is what makes the whole thing safe to run at volume.