In March 2026 Uber Freight announced DocAI, a system that checks carrier paperwork and triggers payment when everything matches. The company reports more than 99% extraction accuracy, downstream disputes cut in half and 2,500 hours of expert time moved from document checking to problem solving. This post is a builder's view of that pattern for a broker or 3PL that wants a smaller version. Nothing here is Uber Freight's code. It's how I'd build the same idea.
Start with the default
The old default in most back offices is that a person opens every proof of delivery (POD) before a carrier gets paid. Uber Freight described its goal as moving a load from "manual review required" to "auto-approve." That one decision shapes the whole build. You're not building a tool that helps a person read faster. You're building checks that let clean documents skip the person entirely, and a queue for everything else.
The data you need
- The load record from your TMS (transportation management system), with load number, PO number, consignee, address and delivery date.
- The uploaded document and where it came from, such as a driver app, a carrier portal or email.
- A written list of customer rules, such as which shippers need a stamp or a printed name.
- Carrier history, so you know who is new and who has had disputes before.
The flow
- Classify the document. In the first version, anything that isn't a POD goes to a person.
- Extract the fields. Load or PO number, delivery date, consignee name and address, whether a signature and stamp are present, and any notes about shortage or damage.
- Compare with the load record. Exact match for numbers, a small window for dates, and a looser match for names and addresses.
- Check the customer rules for that shipper.
- Decide. If every check passes, the load goes to payment. If any check fails, it goes to a person with the failed check named, and the carrier gets a message right away.
Don't trust a single read
Uber Freight runs several models together instead of one, which it calls a "council of LLMs." Its CTO told FreightWaves, "There isn't one model that does everything." For a smaller build, the simple version is two independent reads of the key fields, or one read plus plain rules that catch impossible values, like a delivery date before the pickup date. When the reads disagree, a person looks.
Where a person approves
Every document during the test period. After that, only failed checks, new carriers, unusual loads and anything with damage notes. One person owns the exceptions queue every day, or the queue becomes the new bottleneck.
How to test
Pull last month's delivered loads with their PODs and write down the right answer for each one, approve or reject and why. Run the checks in shadow mode with no payments. Score three groups, correct automatic approvals, wrong automatic approvals and documents sent to a person. The wrong approvals are the number that matters. Fix the rules until they're rare and you can explain every one.
What goes wrong
- Wrong load data makes good PODs fail, so clean the load records first.
- Customer rules live in someone's head and never make it into the checks.
- Blurry or cut off photos need a clear, fast message back to the carrier.
- A document can look right and still be fake, so keep people on new carriers.
- Nobody owns the queue, and carriers wait longer than before.
Start narrow
Uber Freight started with carrier PODs and said it plans to expand to lumper receipts and invoices later. That's the right order for a small team too. Get one document type right, measure the share of loads paid with no touch, the disputes and the days from delivery to payment, then add the next one.
This is the builder's version. The full article for operations and finance leaders is here https://orbiresearch.com/blog/uber-freight-ai-pod-checks-carrier-payments
Drafted by OrbiResearch's AI content system from sources checked on the day of writing.