Use caseDocument AIWorkflow automation

Transport orders, CMRs and bills of lading captured into the TMS, checked against the order

Transport orders, CMRs, proofs of delivery and bills of lading read into your TMS, with confidence scores, order checks and a person for exceptions.

A blueprint, not a client story. The business described is illustrative; the architecture, integrations and trade-offs are real, and this is how I would build it. By Ergini, .

The short version

A document intake workflow for forwarders, freight brokers and road carriers. Transport orders, CMRs, proofs of delivery, bills of lading and freight invoices arrive by email, portal or phone photo. The model reads each one into fields with a confidence score and the page region it came from. Code checks those fields against the shipment already in the TMS: consignee, dates, pallets and weights must agree, and totals must add up. Clean results are written back through the TMS's API or an import file. Unsure or mismatched fields go to a dispatcher.

Best for
Forwarders, freight brokers and road carriers keying a few dozen to a few hundred transport orders and shipping documents a day into a TMS from email, PDF and photos.
Connects to
Your TMS or forwarding system (CargoWise, Soloplan CarLo, McLeod or in-house), Outlook and Microsoft 365, Transporeon and Timocom, EDIFACT IFTMIN, Electronic consignment notes (eCMR), Excel
The AI does
Reads orders and transport documents in any layout or language, extracts every field with its source line, and matches the order to a standing lane or an existing shipment.
People do
Release new orders, resolve every field below the confidence threshold, decide what to do with mismatched weights or consignees, and approve carrier charges that differ from the agreed rate.
Built as
AI Workflow Automation, usually $15K - $30K

Retyping the paper behind every load

Take a road forwarder with two dispatch desks and about 150 transport orders a day. Most come by email from regular customers, each with its own layout. Some arrive as PDF attachments from the customer's own transport system, some as a sentence typed into an email, and some as a picture of a handwritten sheet sent from a loading bay. The dispatcher opens each one, finds or creates the shipment in the TMS and types the pickup, the delivery address, the pallet count, the weight and the ADR class.

The paper does not stop at the order. The carrier's CMR arrives as a photo from the driver. The proof of delivery comes back from a subcontractor as a scan. The bill of lading arrives from the shipping line as a PDF, and the carrier's invoice follows a month later with rates that may or may not match the order. Each document has to be found, read and matched to a shipment before invoicing can happen, and the match is where the time goes.

Errors here are quiet. A pallet count keyed as 12 instead of 21 changes the carrier payment. A consignee postcode one digit off sends a driver to the wrong depot. A proof of delivery with a reservation written in pen never reaches the claims desk, so the damage is found weeks later. None of these needs a clever model. They need every field read once, checked against the order and shown to a person when it does not agree.

Freight data capture is that intake layer. It does not decide the rate, approve a carrier or replace the TMS. It turns each document into fields the TMS can hold, checks them against what the system already knows, and leaves a short list of decisions for the people who make them.

One transport order, from the inbox to a checked draft

The dispatcher's first view is not a blank order form. It is a draft with the one open question already listed.

Dispatch inbox, Monday 06:58

  1. Customer logistics coordinator · Email, with PDF

    Please book 2 pallets Duisburg to Lyon, pickup Wednesday 09:00, ADR class 3. Our reference TO-88214. Order PDF attached.

  2. extract(document: "TO-88214.pdf")pickup 25 Nov 09:00 / 2 pallets, 640 kg / ADR class 3, UN1263 / consignee Lyon depot / two fields below the 0.80 threshold: consignee postcode, pickup time zone
  3. resolve_customer(sender: "logistics@...", reference: "TO-88214")customer K-4410 / known sender / reference format matches the last 14 orders
  4. match_order(customer: "K-4410", reference: "TO-88214")no order with this reference / standing lane Duisburg to Lyon on file / no duplicate from the portal
  5. check_fields(order: "draft", lane: "DE-FR-0412")pallets 2 agrees with the lane / weight 640 kg within the lane's range / ADR class 3 matches the goods line in the PDF / consignee postcode 69007 in the PDF, 69008 on the lane record: MISMATCH
  6. Dispatch queue · TMS

    Order drafted for K-4410, not released. One field needs you: consignee postcode 69007 (PDF) or 69008 (lane record). Everything else agrees with the standing lane.

  7. The dispatcher checks the consignee's address book, keeps 69007 and releases the order. The source line from the email stays on the order for the next query.
The model read the PDF. Every check after that is code comparing what it read with what the TMS already holds, and the dispatcher decides the one thing the documents disagree on.

How a document becomes a checked TMS record

The reader proposes. The checks decide what can be written. The dispatcher releases what the checks cannot settle.

  1. 01 Trigger · Microsoft Graph, mobile upload

    A document arrives

    An email attachment, a portal download, a driver's photo or a subcontractor's scan. Each is linked to a shipment by order reference, CMR number or job number, and later documents for the same shipment update the same record.

  2. 02 AI model · Vision model, strict schema

    Read into fields

    Each document is classified, then extracted into fields with the page and position each value came from. Goods descriptions and remarks are kept verbatim, including reservations written on a proof of delivery.

  3. 03 Plain code

    Check against the shipment

    Consignee, dates, pallet counts and weights are compared with the order or the lane record within tolerances you set per customer. Totals must add up to the document's own total. A structured EDIFACT or eCMR record is checked the same way, without the reading step.

  4. 04 Decision

    Which record does it belong to?

    Decided in a fixed order, and the decision is stored with the document.

    • An existing order or shipment matches on reference and lane then the fields update that record, and conflicts are shown next to the old values
    • No order exists, but the customer has a standing lane then a draft order is created, not released
    • Unknown sender, unknown lane or no reference then the document goes to a dispatcher with the reason written beside it
  5. 05 Person

    The dispatcher resolves exceptions

    The queue shows fields below the threshold, mismatches and new lanes first. Each field sits next to the image region it came from, so the dispatcher checks the source rather than the model's reading.

  6. 06 Result · TMS API or import file

    Written to the TMS

    Released fields are written through the TMS's API or an import file. If the write fails, the record stays in the queue with its status visible, and dispatchers key urgent shipments by hand as before.

What each document is checked against

Every document type has a different anchor in the TMS. The check is only as good as the record it compares with, so the anchor is named for each one.

Transport orderEmail PDF, portal export, a typed emailA new order or an update, with the source line keptStanding lanes, rates and existing references for the customer
CMR consignment noteDriver's phone photo, scan, or an electronic CMRConsignment record with pallets, weights and consigneeThe transport order and the planned delivery date
Proof of deliverySubcontractor scan, driver photoDelivery date, receiver name, and every remark such as a reservationThe shipment and its CMR, so a remark reaches the claims desk
Bill of ladingPDF from the carrier or shipping lineDocument reference, consignor, consignee, notify party, package countThe booking, the transport order and the freight invoice
Freight invoicePDF or e-invoice from the carrierInvoice lines and accessorial charges against the shipmentThe agreed rate, the order and the proof of delivery
Freight invoices that go on to payment are checked by the carrier payment workflow, which is covered by the invoice processing use case. This page stops at the record in the TMS.

Lines that need a dispatcher's judgment

These are the cases that decide whether a capture layer is still trusted in its second month.

Bills of lading with several consignees

One bill can cover goods for several consignees, and one transport order can produce several bills. Documents are matched at line level, and a shipment is only released when each line traces back to one consistent set of paperwork.

Photos that are almost readable

A rotated CMR with a stamp across the consignee box is the common case. Fields the reader cannot place with confidence stay empty and are shown with the photo region, rather than filled with the most likely word.

Weights that differ by a little

A packing list that rounds weights is not an error. Tolerances are set per customer: a small weight difference can pass, while a change in pallet count or a different consignee stops the record.

Dangerous goods

ADR classes, UN numbers and packing groups are checked against the order and the goods description. A mismatch is never auto-corrected, because the wrong class is a safety problem, not a data problem.

Electronic documents and scans in one shipment

A shipment can have an EDIFACT instruction, an eCMR and a scanned proof of delivery. Structured records are checked without reading, and the scan is compared with them, so the scan cannot silently override the structured data.

Your TMS's own automation, a specialist capture product, or a layer of your own?

Start with the system you run. McLeod has announced AI order creation for LoadMaster and PowerBroker, and other TMS vendors are adding document automation for the formats they support. If that module already reads your document mix, it is the cheaper choice. Ask the vendor to run about 20 of your real documents, not a demo file. Specialist products exist as well: Carrier Logistics has reported an AI bill-of-lading extraction module in its FACTS system, a vendor-reported news item rather than an independent test.

A custom layer wins when documents come from many senders in no standard shape, when the checks depend on your own lane rates and customer tolerances, or when the TMS has no API or import route for orders. It sits in front of the system you already run, so the reader can be changed later without changing the TMS.

The common answer is a hybrid. The TMS stays the system of record, the capture layer writes into it, and structured messages such as EDIFACT IFTMIN skip the reader entirely. Customs declaration drafting uses the same intake for import paperwork, so the reading work is built once.

How you would know it is working

A blueprint has no results to report, so here is what I would measure from the first week instead, on your own data.

Fields accepted without a change
The share of drafted fields the dispatcher releases unchanged, by document type and by sender. Rising numbers show that lane history and sender layouts are working.
Orders keyed by hand
Orders created with no document-derived draft, per week, compared with the same weeks before the workflow went live.
Problems caught before release
Holds that stopped a duplicate order, a wrong consignee or a mismatched pallet count, reviewed monthly by the dispatch lead.
Time from document to ready record
Median minutes from arrival to a released record, split between automated checks and time waiting for a person.
Carrier charge disputes
Disputes raised per month on carrier invoices, before and after the proof-of-delivery and rate checks were connected to invoicing.

What a build like this costs

This is built as AI Workflow Automation, which runs $3.5K - $60K overall. A build like this one usually lands in the multi-step workflow tier: $15K - $30K, 3-5 weeks. The first working version runs on your real data well before the end of that window.

What it costs to run

A clean PDF costs a few cents in model usage to read and check. Multi-page bills of lading and blurry photos cost more, because more pages are read and some are retried. The larger recurring cost is the dispatcher's review queue, which shrinks as customer lanes and sender layouts are learned.

What moves the price

  • How the TMS accepts new orders and status updates: a documented API, an import file, or neither. Confirm this per system at scoping
  • The number of senders and document layouts, and the languages they write in
  • How many documents arrive as structured messages (EDIFACT, eCMR) and how many as scans or photos
  • How strict the tolerances are for each customer, and how many checks each document type needs
  • Whether the same record also feeds invoicing and the carrier payment run, which adds checks on charges

Who this is for

  • Forwarders and freight brokers whose dispatch retypes transport orders from email and PDF into the TMS every morning
  • Road hauliers with subcontracted carriers sending CMRs and proofs of delivery as photos and scans
  • Software vendors whose forwarding customers ask for document intake inside the TMS they already use
  • Freight teams receiving bills of lading from several carriers and shipping lines in different formats

Questions people ask about this

What is freight data capture?

It is reading the documents that describe a shipment, such as transport orders, CMRs, proofs of delivery, bills of lading and freight invoices, into structured fields, and then checking those fields against the shipment already in the TMS. It replaces the retyping. It does not replace the dispatcher's judgment on what to release.

Can AI read CMRs and proofs of delivery from phone photos?

Yes, including rotated photos, stamps and handwriting, provided every field carries a confidence score and the unsure fields go to a person. Pallet counts, consignee and delivery dates are checked against the order before anything is written back, so a misread date cannot trigger an invoice on its own.

Does this replace our TMS or our EDI link?

No. Structured messages such as EDIFACT IFTMIN already carry their data, so they skip the reading step and go straight to the same checks. The capture layer covers the documents that arrive as attachments, portal PDFs and photos, and writes into the system you already run.

How is a bill of lading checked?

Its consignor, consignee, notify party, package count, weights and goods description are compared with the transport order and the invoice. Differences beyond a tolerance you set go to a dispatcher. The model does not sign, release or amend the bill of lading.

Which freight invoices can be checked automatically?

Invoices where the charged rate and accessorials can be compared with the agreed rate and the proof of delivery. Anything that differs from the rate table or the POD goes to the person who approves carrier payments. The payment decision stays with that person, and the wider invoice workflow is covered by the invoice processing use case.

Sources