What AI Can and Cannot Decide in Trade Compliance
AI reads and drafts in trade compliance; it cannot hold the licence or the liability. CBLR 2018 and the EU AI Act both put a named human at the decision.
The Question Businesses Are Actually Asking
The question put to vendors is what can your AI do. The question that decides whether a deployment survives its first audit is narrower and better: what is this system permitted to decide, and who answers when it is wrong? Those are not the same question, and the gap between them is where most trade-compliance AI programmes fail — not because the technology underperforms, but because the legal architecture around a cross-border declaration was never built to receive a decision from an unlicensed party.
The honest answer has two halves that must be held together. Artificial intelligence is genuinely good at the work that consumes most compliance hours: reading documents in inconsistent formats, extracting fields, reconciling one document against another, tracking amendments across gazettes, drafting, and flagging divergence. That is a large share of the labour and it is compressible. What AI cannot do is take on the legal position of the party that signs — because in both Indian and European law that position is defined by reference to a licensed or lodging person, and it carries consequences that only a person can bear.
This distinction is often presented as caution. It is better understood as jurisdiction: the boundary is drawn by statute, not by product philosophy, and it does not move when the model improves.
The Indian Position: A Licence Held by a Person
India assigns the customs broker's obligations to a natural or juridical person holding a licence, and states them as personal duties rather than outcomes. Under Regulation 10 of the Customs Brokers Licensing Regulations, 2018, notified by Notification 41/2018-Customs (N.T.), the operative stem is unambiguous — "A Customs Broker shall —" — and among the seventeen clauses that follow, two carry directly onto any question of machine assistance.
Regulation 10(e) requires the broker to "exercise due diligence to ascertain the correctness of any information which he imparts to a client". Diligence is a standard of conduct owed by the person who owes it; software can supply the evidence on which diligence is exercised, but it cannot be the party exercising it. Regulation 10(d) creates a duty to advise the client to comply and, where the client does not, to report the matter to the Deputy or Assistant Commissioner — a duty of judgement and escalation, not of output. Notably, neither clause is a duty to achieve compliance; both are duties to conduct oneself in a particular way, which is precisely the kind of obligation that cannot be discharged by a system that has no standing before the authority.
The penalty provisions complete the picture. Section 114AA of the Customs Act, 1962 reaches a person who "knowingly or intentionally makes, signs or uses, or causes to be made, signed or used, any declaration … false or incorrect in any material particular." The section is built on a mental state. A model has no mental state to interrogate, which means it can never be the respondent in a section 114AA proceeding — the proceeding will find the person who signed.
The European Position: Liability Follows the Act of Lodging
The European Union reaches the same destination by a different route, and its route is stricter. Article 15(2) of the Union Customs Code provides that the person lodging a customs declaration, notification or application is responsible for the accuracy and completeness of the information given in it, for the authenticity and validity of the supporting documents, and for compliance with the obligations of the procedure. There is no knowledge requirement and no diligence defence on the face of the provision: responsibility is allocated by the act of lodging itself.
Read against a system that prepares declarations, the implication is direct. An error introduced by an automated step does not become a lesser error because it was automated — it arrives at the authority as the lodging party's error, in the lodging party's name. Whatever internal recourse exists against a vendor is a commercial matter between two private parties; it does not travel to the customs authority, and it does not answer the demand.
Europe then adds a second layer aimed at the system rather than the declaration. Under Article 14(1) of the AI Act, Regulation (EU) 2024/1689, high-risk AI systems must be designed so they can be "effectively overseen by natural persons" while in use. The Act goes further than requiring a human in the room: Article 14(4)(b) requires that oversight enable the person to remain aware of the tendency to over-rely on system output — automation bias — specifically where the system produces information or recommendations for decisions taken by people. That is a regulator naming the exact failure mode of rubber-stamped review, and requiring the design to work against it.
Where the Two Regimes Converge
Indian and European law arrive at a common structure from opposite starting points. India attaches duties to a licensed person and conditions its heaviest penalty on a mental state; the EU attaches strict responsibility to the act of lodging and separately regulates the oversight design of the system itself. Neither framework contains a mechanism for transferring the position to a non-person, and neither contemplates one. For a business operating both lanes, the practical consequence is a single design rule that satisfies both: the machine's output must arrive at a named human in a form that makes real review possible, and the record must show that review happened. A system that cannot evidence its own oversight fails the European test on design and leaves the Indian broker with nothing to show for diligence.
What This Leaves for AI to Do — Which Is Most of the Work
Reading the limit as a ceiling on usefulness misreads where the cost sits. The expensive part of trade compliance is rarely the final determination; it is the assembly, comparison and maintenance work underneath it — checking that an invoice, packing list, transport document and declaration agree; noticing that a rate changed last Tuesday; finding the clause that governs; reconstructing a file months later for an auditor. That work is voluminous, repetitive, and exactly what machines are good at.
The productive framing is therefore not how much can be automated but how completely the decision can be prepared. A determination that arrives with the governing rule cited, the supporting documents attached, the divergences flagged and the unresolved questions stated as unresolved is a determination a licensed professional can make in minutes rather than hours — and can defend three years later. A determination that arrives as a confident output with no underlying record is faster to accept and impossible to defend, which is the worse trade in every respect that matters.
This is the architecture TradeWatch is built on: machines prepare and cite, evidence is validated-at-source rather than inferred, and a named reviewer-of-record signs before anything is relied upon. Kanan Labs prepares a readiness packet. It does not file Shipping Bills and holds no customs credentials — your licensed CHA files. Final HS classification requires human review.
- Customs Brokers Licensing Regulations, 2018 (Notification No. 41/2018-Customs (N.T.), 14.05.2018, G.S.R. 451(E))
- Section 114AA, Customs Act, 1962 — penalty for use of false and incorrect material
- Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act)
- Regulation (EU) No 952/2013 laying down the Union Customs Code