For most of the past decade, artificial intelligence in customs lived mostly in pilot programmes, conference panels and vendor roadmaps. That changed quickly through 2025 and 2026. The World Customs Organization has spent the past eighteen months publishing formal governance frameworks for AI and machine learning adoption across its membership. National administrations from India to Nigeria are running production risk-scoring models that are directly credited with real drug and contraband seizures. The European Union has written AI-ready, real-time risk management into the next generation of its customs code. And on the commercial side, agentic systems that read unstructured documents and draft submission-ready entries have moved from demo to deployment.
None of this is speculative, and none of it is confined to a single country or a single part of the customs process. Entry-writing, valuation, fraud detection and document handling are all being reshaped at once. This article takes each strand in turn — what is already in production, which dates actually matter, and where the technology genuinely helps — before asking the only question that counts for anyone filing declarations day to day: what changes for me?
Read as separate news items, each of the developments below looks incremental. Lined up in sequence, they show a single eighteen-month shift — from frameworks and pilots to production systems with budgets, contracts and statutory deadlines behind them.
“Agentic” is doing a lot of work in customs marketing at the moment, so it is worth being precise about what it means here. It does not mean a chatbot bolted onto a filing screen. It means a system that takes a stream of unstructured inputs — shipment data, emails, EDI feeds, PDFs — and works through them to produce a submission-ready output, checking and auditing its own work along the way rather than waiting for a human to drive each step.
Customs brokerage has remained a stubbornly manual profession. Licensed brokers classify goods, determine country of origin, calculate duties, check free trade agreement eligibility and secure the permits regulators require — and every one of those decisions is a judgement applied to data that first has to be read out of a commercial document. Those documents rarely arrive in a consistent format, or a consistent order, or from a consistent sender.
That is why the document-reading layer, not the judgement layer, is where automation lands first. Automate the reading and the judgement gets faster; leave the reading manual and every downstream improvement is throttled by a person retyping an invoice. It is also why the honest version of this technology does not claim to remove the declarant. Liability for what is submitted does not move, and neither does the responsibility for the calls a model cannot make on its own.
While the commercial side races ahead, the World Customs Organization has been building the institutional scaffolding underneath it. Its Smart Customs Project, financed by the Customs Cooperation Fund of China, published a Detailed Report on the Adoption of Artificial Intelligence and Machine Learning in Customs in March 2025, covering technical frameworks, governance, risk management, data protection, MLOps capability-building — and the ethical questions of fairness, transparency and accountability that come with automated decisions at the border.
The report grew out of a Global Smart Customs Survey run between April and June 2024, which found AI/ML, blockchain and cloud computing to be the three technologies of greatest interest to WCO member administrations. What follows is what that interest looks like once it reaches production.
Indian Customs offers one of the most detailed public accounts anywhere of a customs AI programme in live use. As described in WCO News in October 2025, the Central Board of Indirect Taxes and Customs built an unsupervised machine learning pipeline that turns messy free-text declaration data into structured, linked records — and then puts that structure to work.
Three downstream models sit on top of that foundation, and it is worth being precise about what each one does, because they are the shape most administrations are moving toward:
A related Post-Seizure Analysis Tool and Offence Database closes the loop, feeding registered fraud cases back into the risk profiles of the entities involved — so a known bad actor’s future declarations are scored accordingly. Indian Customs credits these models with contributing to concrete enforcement outcomes:
Reporting by Baker McKenzie shows the same pattern spreading across administrations of very different sizes and budgets.
| Administration | What is being deployed | Stated aim |
|---|---|---|
| India (CBIC) | Entity and goods codification, predictive targeting, network analytics, real-time valuation model | In production — credited with major seizures |
| Morocco | AI-powered risk-management channel built with the WCO and Switzerland’s State Secretariat for Economic Affairs | Risk-based channelling of consignments |
| Nigeria | AI-driven container profiling on the B’Odogwu clearance platform | 50% reduction in physical cargo checks |
| South Africa (SARS) | Customs modernisation programme built around AI-enabled risk detection | Modernised, risk-led clearance |
Valuation, fraud and routing analysis, in other words, is no longer a feature only the largest or wealthiest customs administrations can build.
The same data infrastructure that speeds up trusted, compliant trade is also the infrastructure used to find the trade that is not compliant. A valuation model that clears a consistent, well-evidenced line in real time is the same model that escalates one that is out of line with history. Clean, consistent declarations are what earn the first outcome rather than the second.
The same direction of travel is being written directly into EU law. On 26 March 2026, the Council and the European Parliament reached political agreement on the most significant reform of the Union Customs Code framework in years. At its centre is a new EU Customs Data Hub — a single online environment through which businesses submit customs and product data once, rather than separately to as many as 27 national customs authorities — and a new EU Customs Authority, headquartered in Lille, which will run EU-wide risk management and analytics from that pooled, real-time data.
Customs and product data is submitted separately, authority by authority, in national systems that do not share a live view of the same movement.
Authority 1Authority 2Authority 3…up to 27
Data is submitted once into a pooled, real-time environment; the EU Customs Authority runs EU-wide risk management and analytics across it.
One submission→EU-wide risk analytics
According to the Council of the EU, the rollout will be phased: the Data Hub becomes operational for e-commerce goods from 1 July 2028, extending to cover all goods movements by 1 March 2034. The scale the system is being designed for is worth sitting with, because it explains why real-time, AI-assisted risk management is not optional at the other end.
The UK picture, which we covered in detail in our recent look at HMRC’s Transformation Roadmap, is moving the same way, if more cautiously — steady adoption rather than a step change.
HMRC’s “Ask HMRC” assistant recorded over 6.3 million interactions in 2025 to 2026 and is being made more conversational.
The Online Trade Tariff is being redesigned into a smarter, journey-based service with integrated AI tools — framed as a 2026 to 2027 workstream rather than a finished feature.
The Microsoft Copilot rollout had reached over 28,000 staff licences by March 2026, on the way to a target of 50,000 during the year.
All of it supporting a Customs Declaration Service that facilitated more than £1 trillion of trade in goods in 2024 to 2025.
Across every example above, the pattern is the same: the constraint is shifting away from how fast a person can key data, and toward how clean and structured that data is at the point it first enters a system.
Brokers and traders should expect AI to keep compressing the low-value, repetitive layer of customs work, while pushing human review toward the judgement calls that genuinely need it. That is a different shape of job, not necessarily a smaller one.
✓ Extracting fields from invoices and shipping documents
✓ Matching and de-duplicating supplier records
✓ Flagging a declared value that looks out of line with history
✓ Repetitive line-by-line data entry across a multi-line entry
● A conflicting document that two sources describe differently
● A borderline classification decision
● An anomaly a model surfaces but cannot explain
● Final responsibility for what is actually submitted
For UK and EU traders specifically: the EU Customs Data Hub goes live for e-commerce goods on 1 July 2028, and HMRC’s AI-assisted Trade Tariff redesign is a live 2026 to 2027 workstream. Businesses that already have clean, structured, validated data flowing through their own systems will experience both as a change of interface. Businesses still re-keying paper documents into legacy tools by hand will find that every automated cross-check against pooled, real-time data compounds the rework rather than removing it.
None of this is abstract for CDUK customers, because the document-to-draft-declaration layer described above already exists inside the platform — and it is live and free across the declaration types below.
CDS ImportCDS ExportGB ENSEU ICS2NCTS Transit
Where the industry’s more ambitious agentic systems are built to resolve conflicts automatically, IDP is deliberately built the other way. When documents disagree with each other, it flags the conflicting fields for the operator to confirm rather than making the call itself — keeping the declarant in control of what actually gets submitted to HMRC.
Building a long declaration is repetitive in a very specific way: the same handful of answers, restated line after line. Automated Commodity Requirements Mapping groups those common requirements across every applicable commodity line, so operators complete them once instead of line by line — while still being able to vary the response for individual items that genuinely differ.
Cross-declaration generation takes the same principle across declaration types. The platform can generate an ENS directly from an existing import declaration, or an ICS2 entry summary declaration from an existing export declaration, in a single action — eliminating duplicate entry entirely. It happens inside one guided platform the declarant uses directly, with the data already validated at the point it was first captured.
The direction is consistent with everything above: less time spent keying data a machine can read reliably, and more of the operator’s attention spent on the judgement calls that still need a human behind them.
The WCO’s governance framework, Indian Customs’ valuation and fraud models, the wave of AI-driven risk programmes across Africa, and the EU’s Customs Data Hub are different responses to the same underlying pressure: trade volumes and document complexity have outgrown what manual review can keep up with. None of them remove the customs officer or the broker from the process; they compress the manual layer and surface risk, and opportunity, earlier.
Customs Declarations UK’s own Intelligent Document Processing and Automated Commodity Requirements Mapping sit on that same trajectory — already doing the document-to-draft-declaration work live and free across CDS import and export, ENS, ICS2 and NCTS transit, while keeping the final decision with the person filing.
AI has moved from customs pilot programmes into production risk scoring, real-time valuation and statutory EU infrastructure — and the balance that defines the next few years will be faster first-pass preparation with the declarant still in control of submission.
Sources: World Customs Organization, Smart Customs Project releases a detailed Report on the Adoption of AI/ML in Customs (28 March 2025) and WCO News 108, Issue 3/2025; Baker McKenzie, South Africa: Customs authorities accelerate modernisation through AI (April 2026); Council of the European Union, EU customs: Council and Parliament agree on landmark reform (26 March 2026); HM Revenue & Customs, HMRC Transformation Roadmap: update 2026.