Agentic AI in Trade Compliance: What the Word Actually Means
Agentic means a system that acts, not one that answers. Gartner found just 15% of IT leaders pursuing fully autonomous agents — governance, not capability.
The Word Is Doing Two Jobs
"Agentic" has become the word vendors reach for when "AI" stops impressing, and it has drifted far enough from its technical meaning that buyers should insist on a definition before a demonstration. Stripped of marketing, an agentic system is one that pursues a goal by taking a sequence of actions in the world, choosing those actions itself, rather than producing an answer and stopping. The difference from a conversational assistant is not intelligence. It is agency — the system does things, and the things it does have consequences that persist after the session ends.
In most enterprise contexts that distinction is a matter of engineering risk. In cross-border trade it is also a matter of law, because the actions available to an agent in this domain — transmitting a declaration, lodging a notification, submitting a claim document — are acts that regulation attributes to a specific accountable party. An agent that merely drafts has produced a document; an agent that transmits has performed a regulated act in someone's name.
The market has already discovered this, quietly. Enterprises are adopting agents broadly and autonomy narrowly, and the gap between those two numbers is the most informative datum available about where this technology actually stands.
What the Adoption Data Shows
Enterprise appetite for agents is high; enterprise appetite for unsupervised agents is not. Gartner's survey of 360 IT application leaders at organisations of at least 250 employees across North America, Europe and Asia/Pacific, conducted in May and June 2025, found that 75% were piloting, deploying, or had deployed some form of AI agent — while only 15% were considering, piloting or deploying fully autonomous agents, which Gartner defines precisely as "goal driven AI tools that do not require human oversight."
The reasons given are not about model quality. Only 19% of respondents reported high or complete trust in their vendor's ability to provide adequate hallucination protection; 74% considered AI agents a new attack vector into the organisation; and only 13% strongly agreed that they had the right governance structures in place to manage them. Gartner's own summary attributes the gap to "concerns around governance, maturity and agent sprawl."
Read against a regulated workflow, that 60-point gap between agents and autonomous agents is not timidity — it is an accurate assessment of where liability sits. An organisation deploying an agent into customer service is risking a bad interaction. An organisation deploying an agent into customs filing is risking a penalty proceeding against a named person.
The Legal Ceiling on Autonomy
Autonomy in trade compliance is capped by law before it is capped by capability, and the cap is explicit in the European framework. Article 14(1) of the EU AI Act requires that high-risk AI systems be designed and developed so that they can be "effectively overseen by natural persons" during the period in which they are in use. This is a design obligation, not a deployment preference: a system architecture that makes meaningful human oversight impractical is non-conforming irrespective of how well it performs.
The Act then anticipates the specific way oversight degrades in practice. Article 14(4)(b) requires that oversight enable the responsible person to remain aware of the tendency to rely or over-rely on system output — automation bias — and it flags this risk particularly where the system produces information or recommendations for decisions taken by people. That is a regulator describing the exact failure mode of an approval queue that has become a formality, and requiring the design to work against it.
Underneath sits the older allocation. Article 15(2) of the Union Customs Code makes the person lodging a declaration responsible for its accuracy and completeness, without any knowledge requirement. Combine the two and the ceiling is clear: an agent may prepare the declaration to any level of sophistication, but the lodging remains an act with a named owner, and that owner must have been in a position to actually exercise judgement.
Prepare Versus Act: The Line Worth Designing Around
The workable distinction for a trade-compliance deployment is not autonomous versus supervised — too coarse to build with — but prepare versus act. A preparing agent may read every document in a shipment file, reconcile them against each other, locate the governing rule, compute deadlines, draft the declaration, and present a complete determination with its evidence and its open questions. None of that touches a regulated endpoint, and none of it is attributable to anyone but the firm running it.
An acting agent transmits. It lodges, files, submits, or binds — and at that moment the output stops being a document and becomes a legal act performed in the name of a licensed party. Everything upstream of transmission is a design choice; transmission itself is a jurisdictional question. Systems built with that boundary explicit tend to survive audits and vendor due-diligence reviews. Systems that blur it tend to discover the boundary during an investigation, which is the expensive way to learn it.
What to Ask Before Buying "Agentic"
Three questions separate a considered architecture from a repositioned chatbot, and none require technical depth to ask. First: what actions can this system take without a human, and against which endpoints? A vendor unable to enumerate the action set has not bounded it. Second: when the agent is uncertain, what does it do? Systems that always produce an answer are systems that cannot signal doubt, and a compliance system that never says "unresolved" is not confident, it is uninstrumented — the discipline behind four-state readiness, where missing evidence resolves to UNCLEAR rather than silently to a pass. Third: what does the record show afterwards? If the agent's reasoning, sources and human interventions are not reconstructable months later, the oversight the EU AI Act requires cannot be evidenced even where it occurred.
The honest position on agentic AI in trade compliance is neither dismissal nor enthusiasm. The preparation layer is genuinely transformable and the gains there are large and immediate. The action layer is bounded by statutes that were not written with software in mind and will not bend to accommodate it. TradeWatch is built on that division: machine preparation at full depth, a reviewer-of-record at the decision, and a record that shows which was which. Kanan Labs prepares a readiness packet. It does not file Shipping Bills and holds no customs credentials — your licensed CHA files.