What Happened When We Put an AI Agent on Customs Docs

I manage operations at a mid-sized freight forwarding company. We handle about 400 shipments a month, mostly UK-EU and UK-Asia. Before last quarter, customs documentation alone took two full-time staff members — and they were still stretched.

The Breaking Point

Every shipment generates the same paperwork: commercial invoices, packing lists, certificates of origin, and customs declarations. Each document has to be checked against the specific requirements of the destination country, the commodity code, and the Incoterms. Get any of it wrong, and the shipment gets held at customs — costing us time, money, and client trust.

We tried off-the-shelf software. Most of it was just template generators — they created the forms but didn't check the data. A bad number in a commodity code field wouldn't be flagged until the shipment was already stuck at Dover. We were spending more time fixing errors than actually moving freight.

The AI Agent Experiment

A colleague suggested we look at AI agents — not chatbots, but systems that could actually read the documents, cross-reference them against customs rules, and flag problems before submission. I was sceptical. We'd been burned by "AI" products before.

We started with one agent on one trade lane: UK to UAE. The setup took about a week, mostly because we had to feed it our SOPs, client templates, and the specific customs regulations for that route. It wasn't plug-and-play, but it wasn't a six-month implementation either.

The agent was running on a small edge box in our office. No cloud, no monthly subscription beyond the base software. I was honestly surprised how quiet the hardware was.

What Actually Happened

Within the first month, the agent was handling about 60% of our customs documentation for that lane — automatically checking documents, flagging errors, and generating the completed forms ready for human review. The two staff members who used to do this work shifted to checking the agent's output and handling the complex cases.

The numbers that mattered:

• Document processing time dropped from about 45 minutes per shipment to under 10
• Error rate on customs forms went from roughly 8% to under 1%
• One full-time role was freed up for client-facing work

The biggest unexpected win was inconsistency. The human team was experienced, but different staff handled the same forms slightly differently. The agent was perfectly consistent. Once we tuned the templates, every document came out the same way. Customs brokers noticed and started complimenting our accuracy.

What I'd Tell Someone Else Considering This

Three things I learned that I wish someone had told me upfront:

1. Start with one route. Don't try to automate everything at once. Pick your highest-volume trade lane and get that right first. The lessons from one lane apply to the others.
2. Your SOPs need to be clean. The agent can only be as consistent as the rules you give it. We spent as much time cleaning up our internal documentation as we did setting up the AI. That was time well spent.
3. On-premise matters for customs data. Shipping manifests and customs forms contain commercially sensitive information. I would not have been comfortable sending that data to a cloud AI service. Running it locally was a non-negotiable for us.

We're now rolling out the agent to three more trade lanes. Each one gets faster to set up than the last. The team has gone from sceptical to actively suggesting new workflows to automate. That shift in mindset — from "will this work?" to "what else can it do?" — has been the most valuable outcome.