Frontiers of Trade: AI agents as traders  

Welcome to the second in our Frontiers of Trade series. We're looking beyond today's headlines on tariffs and geopolitical tensions to the questions businesses will need to confront over the next decade. This blog examines what happens when AI agents begin executing cross-border trade without human oversight, and why the legal frameworks governing those trades are not ready. 

Imagine the scene: A retail group's AI procurement agent autonomously identifies a lower-cost electronics supplier in Vietnam, negotiates terms, raises a purchase order, and arranges shipment to the US. The agent classifies the goods using a six-digit HS code it has used successfully before, drawing on its own record of a code that cleared in the past rather than checking the live schedule. What it does not know is that the specific product configuration, properly classified at the eight-digit level, falls under a Section 301 tariff schedule carrying a US duty north of 50 per cent. The shipment clears. The customs invoice arrives. The company owes millions in duties it never budgeted for, on a deal nobody signed off on. 

Who is responsible for the error? Who pays? And, more pressingly, what exactly was the legal status of the entity that made the decision? The instinctive reaction is that this is simply a badly built agent, and that better engineering would have caught the error. That instinct is not wrong. But it is only half the story: even a well-run agent raises questions that our current trade system cannot yet answer. 

Trade law was built for humans making decisions 

Autonomous AI agents are already executing real commercial transactions, negotiating supplier contracts, triggering purchase orders, arranging logistics, and filing trade documentation, without a human approving each step. It is already happening, across supply chains in financial services, retail, pharmaceuticals, and advanced manufacturing. 

The entire architecture of international trade law assumes a human being made a decision. Rules of origin, import liability, sanctions compliance, competition law, export controls; all of them trace liability back to a person. When the decision-maker is an algorithm does it default to whoever built or deployed the machine? We identify seven emerging gaps in the current framework. 

It is worth being precise about which kind of liability is in play, because the two travel on different tracks. Commercial liability can largely be resolved privately through contract terms, indemnities and insurance. Legal and regulatory liability, the question of who has breached a trade rule, cannot be reassigned by private agreement; closing the gaps requires governments to change the rules.  

Not all AI is the same 

Narrow AI has been embedded in customs processes for years: rules-based classification engines, risk-scoring algorithms and automated document checks. These systems have been highly defined; they behave much like a database. The new risk stems from the probabilistic nature of the general-purpose AI capabilities that some customs processing platforms are now building on top of. It is here that risks emerge, such as outdated training data, and outputs that sound plausible but are wrong. 

General-purpose AI can make the older narrow systems more powerful, chiefly by coherently linking processes that previously sat in silos. To do that safely, however, agentic platforms must embed the narrow components, definitively confirming the latest tariff codes, sanctions lists and licensing requirements, within the agentic workflow. The better platforms are already doing this. The risk stems from those who choose to build on top of a frontier model without any step of their own to verify the data the model returns. That is where hallucinated tariff codes and stale export controls screening will surface. 

The seven gaps 

1. Who is the importer? Customs law normally requires a named, legally responsible Importer of Record for every cross-border shipment. The US is moving forward more quickly than others. A January 2026 US Customs ruling confirmed that AI cannot conduct "customs business" itself, meaning the liability sits with the human deployer, even if that person had no knowledge the transaction was happening. As AI agents grow more autonomous, this position will become harder to sustain. 

2. When two agents transact, who owns the outcome? Increasingly, an AI buyer is able to negotiate with an AI seller both serving different companies in different jurisdictions. If the resulting contract violates a country’s rules, no existing framework determines how fault is allocated across two separate supply chains. 

3. The tariff schedule updates; the agent doesn't. HS classifications, Section 301 lists, and preferential tariff schedules are updated continuously. AI agents trained on historical trade data can apply outdated classifications to new product configurations. The exposure is greatest where a platform relies on what a general-purpose model recalls from its training data rather than a deterministic look-up of the current schedule.  

4. You cannot sanction an algorithm. Sanctions are instruments designed for persons: freeze assets, restrict travel, impose reputational cost. An AI agent has none of these vulnerabilities. If an agent transacts with a prohibited counterparty, enforcement action runs to the deployer, who may have had no visibility over the transaction at the time it occurred.  

5. Can you prosecute a cartel that never had a meeting? Competition law, both the Sherman Act in the US and Article 101 TFEU in the EU, requires evidence of agreement or coordinated intent. AI procurement agents trained on similar data can independently converge on identical pricing or sourcing strategies without any communication between them. A recent OECD report on Algorithmic pricing and competition in G7 jurisdictions demonstrates how existing competition practices are not equipped. The international trade dimension compounds this problem as there is no multilateral enforcement mechanism for cross-border algorithmic convergence. 

6. Is the agent a broker, a service, or a product? If a firm offers an AI agent to manage third-party procurement across borders, it may simultaneously be providing a financial service, a logistics service, a data service, and, per the January 2026 CBP ruling, unlicensed customs brokerage. The WTO has no classification for this. Neither does any current free trade agreement. 

7. What happens if the agent routes around export controls? An AI optimising for cost and speed may chain together suppliers and intermediaries in ways that trigger export licensing requirements for controlled goods or dual-use technology, without any mechanism to flag that it has done so. Under US EAR and ITAR, and equivalent EU regimes, ignorance is not a defence. The deploying firm carries strict liability for a routing decision it did not take and may not be able to reconstruct. 

What companies need to be thinking about 

Governments will eventually close these gaps. The question is whether your business is exposed. We have identified four practical steps companies should be taking now. 

• Map their autonomous exposure. Identify processes where an AI system is initiating, influencing, or completing a cross-border transaction without mandatory human sign-off. Most companies doing this exercise find the footprint is larger than expected. 

• Treat the deployer liability as real. Legal liability for AI-executed transactions sits with the deploying organisation. That means trade compliance obligations (sanctions screening, export classification, tariff classification accuracy) apply in full to transactions you did not directly authorise.  

• Build classification and screening into the agent. Embed the narrow, deterministic components, live tariff schedule look-ups and current sanctions and export control screening, into any agentic workflow. Design agentic procurement workflows to require human review at defined risk thresholds, and to log every automated decision in a form that can be audited later. 

• Engage policymakers on the data infrastructure. Companies have a concrete ask here: government agencies and customs authorities should publish robust, fully up-to-date APIs for tariff schedules, sanctions lists and licensing requirements, so that the deterministic components of an agentic system can draw on authoritative data at the moment of transaction. Because agents transact across borders, unilateral national fixes will only go so far; the greater prize is a multilateral to align these interfaces around common, interoperable standards. Candidate organizations include the WTO, OECD, building on initiatives like UNCITRAL’s model laws , or a plurilateral route through digital trade arrangements like Digital Economy Partnership Agreement (DEPA).  Much of the official data in this area remains unstructured and poorly suited to machine consumption. Fixing that would de-risk autonomous agents. 

Back to the border 

Most businesses survive an unexpected duty bill from our opening example. The harder question is what happens when the agent's mistake triggers a sanctions violation, creates an undetected cartel, or routes a dual-use export through an unlicensed intermediary.  

Regulators may not even be the first to force the issue. Insurance underwriters are likely to move sooner, insisting on contracts that clearly specify where autonomous decision-making ends and human accountability begins, verification that agents are drawing on live rather than historical data, and full auditability of automated decisions.  

There is further reason to expect the policy conversation to intensify. As autonomous agents become load-bearing infrastructure in global supply chains, governments will increasingly view them through the lens of resilience and security, not compliance alone. When thousands of firms depend on the same agents, those systems start to look like a single point of failure: a corrupted data feed could propagate across critical supply chains at machine speed. Expect policymakers to scrutinise the security and resilience of AI trade agents accordingly. 

Companies that map their exposure, update their compliance architecture, and engage with the policy conversations now will be better positioned than those waiting for regulators to act first. In trade, as in customs classification, the cost of the wrong answer can arrive before the right framework does. 

George Riddell is a Senior Advisor specialised in supporting clients in international trade, sustainable supply chains and geopolitical fragmentation. Joe Armitage is Flint's Head of AI Product & Partnerships. Chris Martin leads Flint's tech practice across Asia-Pacific.

Enjoyed this blog? Sign up for our Trade Monthly newsletter for analysis of the political, policy and regulatory developments shaping UK, EU and global trade.