Inside the inbox: what 15,000 quote emails a day taught us about freight email

The dataset

Across pilot deployments with U.S. freight brokerages, FreightSurge ingests more than 15,000 inbound quote-related emails every day, covering dry van, reefer, flatbed, and intermodal moves. A sampled subset is reviewed by experienced brokers for shipment fields, urgency, and intent — and that labeled data is what shapes the rest of this post.

What we found

1. Quote requests are messier than docs suggest

Public datasets and EDI standards make freight email look orderly. The reality is closer to the way humans actually write — abbreviations, missing pickup dates, ambiguous origins, "same as last time" implicit references, and trailing email signatures with phone numbers and disclaimers that confuse naive extractors.

2. Context lives outside the email

Roughly 38% of the requests we see rely on context that does not appear in the message body — a previous quote, a recurring lane, a known commodity. A model that only reads the email gets the obvious facts right and the important details wrong.

3. Tone matters as much as numbers

A reply that quotes the right rate but breaks tone with a known customer reads worse than a slightly off rate in the right voice. Brokers told us this directly when they edited drafts: "the math is fine, the words are not."

How this changes the system

These findings forced us to make three architectural choices that a generic chat product would not:

1. Extraction is grounded in a freight-specific schema, not freeform JSON the model invents.
2. Lane and customer context is retrieved at draft time from the broker's own data — not baked into the model.
3. Drafts are generated against a per-customer style profile so the tone matches existing rapport.

That last point is what shapes the rest of the product: every quote shows the lane history that informed it, the spot rate it was benchmarked against, and the confidence score on each extracted field. The AI is the assistant. The broker still ships.