Every dispatcher knows the routine. Broker sends a PDF. You open it, manually enter pickup address, delivery address, rate, reference number, pickup window, delivery window, equipment type, and any special instructions into your TMS. Then you file the PDF somewhere. Seven minutes of work that produces zero value — it's just data moving from one place to another because someone didn't build an API.
Rate Con AI was built to eliminate those seven minutes. Here's exactly how it works.
The technical process
When a rate confirmation PDF arrives — either uploaded directly in ESSE Portal or forwarded to your ESSE email address — the system extracts the text content of the document. This sounds simple, but PDFs are infuriatingly inconsistent: some are text-based, some are images of text, some are combinations, and every broker uses a slightly different layout.
For text-based PDFs, we extract the raw text and pass it to GPT-4o with a structured prompt that asks it to identify specific fields: origin city/state, destination city/state, pickup date and time window, delivery date and time window, rate, equipment type, weight, broker company name, broker contact, and reference number.
For image-based PDFs — scanned documents or PDFs generated from photos — we use OCR processing first, then the same extraction pipeline.
The 5% that breaks it
The failures are interesting. Here's what causes them:
Ambiguous addresses. "Memphis" could be Memphis, TN or Memphis, TX. Usually context makes it clear — a dry van load starting in Chicago is not going to Memphis, Texas. But sometimes it's not clear, and the AI makes a guess. We flag low-confidence address extractions for human review.
Non-standard date formats. Most brokers write "04/22/2026" or "April 22, 2026." Some write "22APR26" or "4-22" with no year. The AI handles most of these, but not all.
Multi-stop loads with unusual formatting. A standard A-to-B rate con is easy. A rate con with four stops, two pickups, and complex delivery windows across a two-week period is hard. We handle these less reliably and always flag them for review.
Broker company names that don't match our database. We maintain a broker contact database. When a rate con comes from a broker we've never seen before, we create a new contact record but flag it as unverified.
What happens when the AI is wrong
Every extracted rate con shows the original PDF side-by-side with the extracted fields before the load is created. The dispatcher reviews and corrects any errors with one click. Even when the AI gets something wrong, the dispatcher is correcting a pre-filled form rather than filling a blank one — which is still significantly faster than manual entry.
We log every correction. Those corrections feed back into our training data to improve future accuracy. The system gets better the more it's used.