Insights · AI in trade

AI agents in customs: what they should do, and what a person must still approve

AI agents can take most of the reading, checking and re-keying out of customs work. The legal responsibility does not move, so the design has to keep a licensed person in charge.

Customs work is document-heavy, rule-heavy and time-critical: exactly the kind of work AI agents are good at. But a customs declaration is a legal statement. In Australia, New Zealand, the UK, the USA and elsewhere, the person who makes it, and the broker who lodges it under their licence, carry the responsibility for its accuracy. Software changes how the work is done. It does not change who answers for it.

What agents do well

  • Reading documents: pulling parties, values, quantities, weights, origins and Incoterms® rules from invoices, packing lists and bills of lading, whatever their layout.
  • Cross-checking: spotting that the packing list shows 412 cartons and the invoice 420, or that the origin on the certificate does not match the invoice.
  • Suggesting classifications: proposing a tariff code with the legal text and reasoning behind it, and a confidence score.
  • Screening: checking parties against sanctions and denied-party lists and flagging goods that may need permits.
  • Watching shipments: following container milestones and warning people before free time or deadlines run out.

What a person should approve

  • Lodging or amending a customs declaration or cargo report.
  • Paying duty, taxes or fees.
  • Anything that commits the business to another company, such as accepting a quote or confirming a booking.
  • Low-confidence classifications, valuations with related-party or royalty questions, and any sanctions match.

Good systems make these approval points settings, not afterthoughts, and let a business move them as it gains confidence on each lane.

Agents talking to agents

More companies now run their own agents. Two open protocols are becoming common: the Model Context Protocol (MCP), introduced in 2024, which lets an agent use tools and data from other systems, and the Agent2Agent protocol (A2A), introduced in 2025, which lets agents from different companies exchange messages and tasks. Used well, a supplier's agent can answer a question about an invoice in minutes. Used carelessly, an agent could act on a message it should not trust. So every agent needs a verified identity, permissions limited to its job, and a full log of what it did and why.

A simple test for any AI customs tool

  • Does it show the source and reasoning for every value it suggests?
  • Can you set which actions need approval, by country, client and value?
  • Is every action logged in a way an auditor or customs officer could follow?
  • Does the licensed broker stay the declarant, with their own credentials?

How this shows up in GenticTrade

GenticTrade is built on these rules. Agents prepare and check the work, approval gates are set per workspace, every action is logged, and licensed brokers lodge under their own licence. Our agents connect through MCP and A2A, inside the same approval gates.

Sources and further reading

  • Australian Border Force, licensed customs brokers and the responsibilities of importers (abf.gov.au).
  • Model Context Protocol specification (modelcontextprotocol.io).
  • Agent2Agent (A2A) protocol, an open project of the Linux Foundation.

General information as at 21 November 2026, not legal, customs or financial advice. Check the official source or a licensed customs broker before acting.

Keep reading

More from Insights

Launching in Australia and New Zealand first

Let agents do the paperwork.

GenticTrade gives you live shipment visibility and AI agents that prepare customs work for a licensed broker to approve.