Operational Note AI in Trade Compliance

Nine Questions to Ask an AI Trade-Compliance Vendor

Last reviewed:

Demos test fluency, not defensibility. Nine questions on citations, abstention, logs and liability that separate an evidence system from a fast drafter.

Chapter 1

Buying for the Query, Not the Demonstration

Most evaluations of trade-compliance software test the wrong property. A demonstration shows how the system behaves on a clean shipment with complete documents, which is the case that was never difficult. The case that matters is the one where a document is missing, a field disagrees with another field, or a customs officer asks why a determination was made — and the answer has to be produced from the record months later.

The gap is measurable in the market's own confidence. In Gartner's survey of 360 IT application leaders, only 19% reported high or complete trust in their vendor's ability to provide adequate hallucination protection, and only 13% strongly agreed they had the right governance structures in place. Buyers are not short of enthusiasm; they are short of evidence that the systems behave well when they are wrong.

These nine questions test for that. None require technical depth to ask, and each has a wrong answer that is easy to recognise.

1.1

Questions 1–3: Can the Output Be Traced?

A determination that cannot be traced to a rule is an opinion, however well formatted. The first three questions establish whether the system produces evidence or merely output.

Get the details
1.2

Questions 4–6: What Happens When the System Does Not Know?

A compliance system's honesty is measured by what it does with missing evidence, not by what it does with complete evidence. This is the cluster of questions vendors are least prepared for.

See more about
Get the details
1.3

Questions 8–9: Where Does the Liability Actually Land?

8. Who is named on the output? If the determination carries no identified reviewer, the file cannot evidence what Article 26(2) requires — oversight assigned to natural persons with the necessary competence, training and authority. "The system checked it" answers nothing; "the platform is certified" answers nothing either. A named person, with a timestamp, is the only artefact that satisfies the question.

Get the details
1.4

The Pattern in the Answers

Taken together the nine questions separate two architectures that look identical in a demonstration. One produces plausible documents quickly. The other produces determinations that carry their rule, name their gaps, identify their reviewer, and survive being examined by someone adversarial. Only the second is worth deploying into a workflow where a regulator can ask questions.

Get the details
Sources
  1. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act)
    EURetrieved August 7, 2026
  2. Gartner Survey Finds Just 15% of IT Application Leaders Are Considering, Piloting, or Deploying Fully Autonomous AI Agents (30 September 2025)
    Other sourceRetrieved August 7, 2026
  3. Regulation (EU) No 952/2013 laying down the Union Customs Code
    EURetrieved August 7, 2026
Current layer Overview