#1: How we used AI to automate 90% of POD validation
PODs (proof of deliveries) are the documents that confirm a shipment was delivered and crucial to executing all the payments related to any truckload shipment.
It sounds simple, but today the process for validating a POD before executing a payment is one of the most broken, manual workflows in freight and until recently, we were handling it the same way most logistics companies do:
A carrier would send a blurry photo or scanned PDF.
Our ops team had to track it down, upload it manually into the system, interpret what it said, cross-check each line against shipment details, and
Only then mark the load “invoice ready” so a payment could get executed
This was happening across thousands of shipments per month. It was slow, inconsistent, error-prone, and wasted countless hours for our ops teams every week.
The pain
Specifically we were seeing these painpoints:
Dozens of hours per week lost to parsing POD emails and matching fields manually
High error rates and missed data points, depending on carrier and doc format
2-3-day delays to reach “Ready to Invoice” impacting lead times for working capital
Zero structured data to track performance, benchmark vendors, or resolve disputes quickly
What we built
We didn’t plug in a generic OCR tool or document parser.
Instead, over several months, our R&D team ran hundreds of experiments combining multiple vision-language models (VLMs), custom prompt architectures, and proprietary datasets. We built dynamic validation logic based on specific customer needs and continuously retrained the system using patterns from human-rejected edge cases.
We ran side-by-side experiments on real PODs and found Gemini 2.0 Flash consistently outperformed other models on visual parsing, especially with hard-to-read scans and lower-resolution docs. GPT-4.1, on the other hand, was more reliable for text validation and generating explainable outputs. So we combined both.
What we came up with was a fully integrated three-module pipeline inside NuvoOS: Reception → Validation → Action.
1. Smart intake
Emails are sent to a dedicated inbox. From there, GPT-4o extracts shipment numbers, sender names, contact info, and identifies associated loads. The system routes single and multi-load documents automatically, pulls metadata from NuvoOS, and archives everything in Google Drive and Airtable.
This removes the need for email triage and data entry, which used to make up nearly all of the manual effort in processing incoming PODs.
2. Intelligent Validation
We use Gemini 2.0 Flash to extract structured data from POD documents, then apply GPT-4.1 Mini to validate each field against expected values. Every pass/fail comes with a clear, explainable output.
The system creates a fully auditable trail, flags anomalies without slowing down the flow, and gives our teams complete visibility for QA and reporting.
3. Automated Fulfillment
Validated docs are instantly marked “invoice-ready” in NuvoOS and stored appropriately. Exceptions are routed to Slack for review, without interrupting the flow. Human feedback loops directly feed model retraining and continuous improvement.
The results
Today, this system runs live across thousands of shipments per month and the impact is very real:
95%+ field-level accuracy
90%+ of POD flows are fully automated, end to end
PODs are now uploaded and approved 3.4× faster, saving 8+ hours of ops time per day and accelerating downstream workflows
Fewer disputes, cleaner inboxes, and ops capacity redirected to higher-leverage work
If you want a demo of this workflow or of Nuvo OS (our AI native operating system for North America freight), please book some time with our team here.
We’re always excited to share any of our learnings and how we are helping North American shippers reduce freight spend, boost customer service levels, and adopt AI across their entire North America supply chain.