Intelligent document processing can only read what somebody wrote down.
Ask any engine how to automate freight document processing and you get the same answer: run IDP over the inbound PDFs, extract the fields, validate them against the load record, write them into the TMS, escalate the exceptions. That answer is correct and it is not the whole job. It assumes the document contains the thing you need — and in freight, the fields that decide who pays are usually the blank ones.
Below: five eligibility conditions taken word for word out of executed US freight contracts, and what each one implies about what your automation has to capture rather than read.
Why does document automation stall on freight paperwork?
Because a detention or cargo claim is decided by conditions the paperwork has to satisfy, not by data the paperwork already holds. We read 10 US freight documents end to end rather than summarising what other sites say about them. Of the five detention eligibility conditions we extracted, four depend on a person doing something at the dock and it being recorded there. None of them can be recovered by reading the file afterwards, at any level of model quality.
| What the contract says | Source | Can IDP recover it? |
|---|---|---|
| “To be eligible for detention pay, drivers must obtain accurate in and out times that are stamped or written on the Bill of Lading (BOL) by the receiver or shipper.” | Leonard's Express, executed rate confirmation, order 2453170, page 2 | No |
| “To be eligible for detention pay, the Carrier's driver must obtain preapproval authorization by contacting the Broker […], and further in order to receive the detention pay Carrier must provide accurate “in and out times” stamped or written on the bill of lading by the receiver or the shipper.” | Leonard's Express, executed rate confirmation, order 2453170, page 5 | No |
| “Failure of driver to accept tracking will result in the denial of detention. Additionally, Leonard's Express will have the authority to deduct up to $150.00 from the contracted rate stated within this contract.” | Leonard's Express, executed rate confirmation, order 2453170, page 1 | No |
| “Carrier will not be entitled to detention charges when Carrier arrives at the location of loading or unloading after the scheduled appointment time.” | Campbell's, motor carrier accessorial exhibit, Exhibit 2, Accessorial Charges and Fuel Recovery, page 4 | No |
Read the right-hand column again. A model handed the bill of lading can tell you the times printed on it. It cannot tell you whether the receiver signed them, whether the driver called for preapproval, or whether the tracker was on — because those are events, and an event that nobody recorded leaves no trace in any document to extract from.
How do you automate freight document processing with AI?
Audit what is being lost for not being written down, capture it at the point of the event, then read and reconcile on top of what is now captured. catuh builds agents for three jobs: reading the documents, capturing the missing facts and matching invoice lines to their proof. The sequence is not a preference. Running stage three first is why these projects get quietly abandoned nine months in: the extraction works, the numbers do not move, and nobody can say why.
- 01 / 03
Read the documents that arrive by email
Rate confirmations, bills of lading, proofs of delivery and carrier packets come in as attachments. The agent reads each one and writes its fields into the system you already use, with anything it cannot read escalated to a person.
What you get: Fewer hands re-keying the same paperwork.
- 02 / 03
Capture the facts no document records
Arrival and departure times, the exception at the pickup and the carrier verification are not on any form because nobody wrote them down. The agent records each one as the load moves, with its source and the time it was captured.
What you get: A dated record for the facts that decide a dispute.
- 03 / 03
Match every invoice line to the document that proves it
Before an invoice goes out, the agent checks each line against the document that supports it. A line with no proof behind it is held back for a person instead of being sent and disputed later.
What you get: Invoices that carry their own evidence.
Can AI do rate confirmation and BOL data entry on its own?
Yes for the rate confirmation, no for half of what the BOL is supposed to prove. A rate confirmation arrives complete, so extraction solves it outright and that is where IDP earns its cost. A bill of lading arrives with the fields that decide a dispute left blank, and no model fills a blank.
| Task | IDP on its own | Capture first, then IDP |
|---|---|---|
| Retyping a rate confirmation | Solved. The fields are on the page. | Solved, same way. This is where IDP genuinely earns its cost. |
| Knowing when the truck arrived | Only if somebody wrote it down. Usually a text message that never reaches the file. | Captured at the stop with its source, so the extraction has something to find. |
| Proving the receiver signed the times | Reads the signature if there is one. Cannot tell you it is missing until you already lost. | The blank is flagged while the driver is still at the dock and can fix it. |
| Exception noted at pickup | Reads the note. Most BOLs are signed clean, so there is no note to read. | The agent asks for the note while the driver is still at the pickup, which makes leaving it empty a decision. |
| Carrier vetting a year later | Extracts today's status from today's certificate. | Holds what you saw on the day you awarded the load, which is the only version that defends you. |
How does OCR work on freight documents, and how accurate is it?
OCR turns the image of a freight document into text and fields, and its accuracy depends on what is printed or written on the page, not on the model alone. Current document models can read handwriting on a bill of lading, including the exception note in the margin. No accuracy rate is quoted here because we have not measured one, and the published rates that circulate for freight OCR trace back to vendor marketing rather than to a study. Whatever the rate, OCR can only return a field that somebody filled in.
What problems come up when automating scanned or low-quality freight documents?
The scan is rarely the main problem; the blank field is. A skewed, faded or photographed page costs extraction reliability, which is why every serious setup escalates low-confidence reads to a person rather than publishing a figure for how often that happens. A clean scan of a bill of lading signed without times, or with the exception box left empty, still gives automation nothing to read about who pays for the load.
How do I get started automating document processing in a small company?
Start by finding which documents your people retype and which fields are lost for not being written down, before you buy any automation. The order we work in is to audit what is lost, capture it at the point of the event, and only then read and reconcile on top of what is captured. Retyping a complete rate confirmation is the quick win, and the saving is arithmetic you can do yourself: documents re-keyed per day, times minutes each, times your loaded hourly cost.
Frequently asked questions
How do you automate freight document processing with AI?
With intelligent document processing: a model reads the inbound PDF or email, extracts the fields, validates them against the load record and writes them into the TMS, with exceptions escalated to a person. That part is settled and every vendor does it. What decides whether it pays off is whether the fields you need were written down in the first place, because extraction cannot recover a time nobody recorded.
Can AI read a handwritten bill of lading?
Yes. Modern document models handle handwriting on a BOL, including the exception note in the margin, which is the field that decides a Carmack cargo claim. The harder problem is not reading the note but the fact that most BOLs are signed clean with the box left empty, so there is nothing to read.
Is intelligent document processing worth it for a small brokerage?
It pays for itself on volume of retyping, so it depends on how many documents a person keys in each day rather than on how many trucks you run. A twenty-person brokerage keying rate confirmations and PODs by hand usually clears the cost. The order still matters: automate the retyping of documents that already contain what you need, and fix capture separately for the fields that are blank.
What does a TMS not do?
A TMS records the load and issues the invoice. It does not build the file that makes the invoice collectible: the arrival time with its source, the signature that was or was not obtained, the exception noted at pickup, each kept with the date it was created. That file is what a payer asks for when it disputes the charge, and it is assembled from documents rather than from the load record.
References
Every figure and every quotation on this page traces to one of these. Each was retrieved and read rather than cited second hand.
- Leonard's Express executed rate confirmation, order 2453170 Source of the detention eligibility clauses quoted above.
- Campbell's motor carrier accessorial exhibit (Exhibit 2)
- Saia LTL Freight uniform straight bill of lading
- Estes Express Lines bill of lading
- YRC Freight straight bill of lading (OP-097 09/21) The carrier is defunct; the host may disappear.
- ArcBest master bill of lading, December 2023
- GS1/VICS standard bill of lading (NTS Logistics)
- Landstar published tariff LAND 100-A, March 2020 Detention is Rule 500 with power, Rule 501 without.
Stage one is a conversation, not a pilot.
Tell us which documents your people retype and which numbers you bill but do not collect. We will tell you in writing which half is an extraction problem and which half is a capture problem, and roughly what each takes. If it is all extraction, you do not need us for it and we will say so.
The reason to start with the audit rather than the tooling: extraction lowers what the typing costs, and only the capture stops the cost moving with whoever is on shift.