Building AI agents you can actually trust in regulated freight workflows

Trust is the actual feature

A demo where AI drafts a perfect reply is not interesting. The interesting product question is: what happens the first time the AI is wrong? In a brokerage, "wrong" can mean a misread destination, a stale lane rate, or a reply sent in the wrong tone to a top-five customer. Trust is built by what happens around the model, not by the model itself.

The four guardrails we ship

1. Structured extraction, not freeform JSON

Every extracted shipment is constrained to a typed schema: origin and destination zips, commodity class, equipment type, weight, dates. The model fills the schema; it does not invent fields. When confidence on a field is low, the field renders as "needs review" instead of being silently filled in.

2. Retrieval over the customer's own data

Pricing context comes from the brokerage's historical loads, not a global model. This sounds like an architectural detail. It is the entire trust story: a generic model that hallucinates a rate based on training data is not auditable. A retrieval system that returns the 12 specific past loads it priced from, with rates and dates, is.

3. Human-in-the-loop by default

Drafts are never sent automatically. The broker reviews, edits, and clicks send. Every send is logged with the broker's identity, the model version, the retrieved context, and the diff between the AI draft and the final outgoing email. That log is the audit trail an enterprise legal team will ask for in week two of evaluation.

4. No model training on customer data

We do not train shared models on customer emails or customer-specific lane data. Customer-specific intelligence stays customer-specific. This is non-negotiable in regulated industries and increasingly non-negotiable for any enterprise procurement process.

What this looks like in practice

When a broker opens a draft, the side panel shows the extracted fields with confidence scores, the suggested rate with the lane history that produced it, and the comparable spot rate from the broker's preferred index. Nothing is hidden. The broker can look at any number on the screen and answer "where did this come from" in five seconds.

That is the engineering bar. Anything less feels like a black box, and brokers do not — and should not — hand customer relationships to a black box.