AI Classification Is Raising the Bar. Is Your Process Keeping Up?

AI classification technology has become remarkably capable. Give a dedicated classification AI system a product description and it can produce a proposed tariff classification, supported by a written analysis of the General Rules of Interpretation, relevant legal notes, and potentially applicable rulings, in a matter of minutes.

For anyone who has spent years managing classification processes, that is an impressive development. Even for companies already using other forms of classification automation.

Over the past year, we've seen companies focus their evaluation of AI classification on accuracy, usability and time saved per classification. Those are sensible tests. But measuring time saved assumes that classification will continue to be performed as it is today, just faster.

I think that misses the bigger opportunity. AI can change the standard of work a classification team is able to deliver, including the depth of its analysis and the evidence behind its decisions. Capturing that value means rethinking the process, not simply replacing manual research or traditional automation with an AI tool. Otherwise, you risk automating an existing process rather than taking advantage of a fundamentally different capability.

And that redesign matters practically and commercially. Until companies understand how AI adds new value to the classification process, they cannot realistically judge the feasibility, cost or success of an AI classification implementation.

From on-time to defensible classification

In many organizations, the classification process is built around throughput. A request arrives, a specialist researches the product, assigns a code, and records it in an ERP, trade compliance system, or spreadsheet. A written rationale is desirable, but the combination of deadlines and volume often leaves little room for one. In my experience, rationales are frequently minimal, if documented at all.

The trade-off is often seen as speed vs. accuracy, where in many organizations the true (if unspoken) KPI is avoiding missing classifications holding up shipments.

AI changes that trade-off. A fully documented, detailed analysis of the General Rules of Interpretation, legal notes, and relevant rulings can now be prepared in minutes. We previously accepted that classification rationales were not properly documented, since the alternative of writing a full memo manually was simply operationally unfeasible.

AI now creates the opportunity to raise the standard from ‘did we classify everything in time?’ to ‘can we actually defend these classifications when challenged?’

It also changes the expert's work. Reviewing a reasoned memorandum means checking the product facts, testing the interpretation of the tariff provisions, and resolving assumptions. That takes more judgment than confirming that a suggested code looks plausible. And it may require asking your engineering team more questions up front.

The time saved in drafting an analysis should create room for stronger judgment, not merely a higher volume of approvals. Review intensity should reflect the complexity and risk of the decision, but organizations need to design that review deliberately, rather than assume AI output can be approved with a quick glance. That may require new review criteria, escalation paths for difficult questions, and a more deliberate division of work between experienced classifiers and design or engineering teams.

Customs is raising its game, too

Why should companies invest in more thorough classification analysis when their existing process is getting the job done? Because customs authorities are changing what it means to be compliant, too.

Just as AI allows companies to analyze and document classifications at a scale previously considered impractical, customs authorities are adopting similar technologies to scrutinize the declarations they receive.

In the US, CBP is increasingly using data mining and analytics to identify inconsistencies in classification, origin and valuation. Rather than examining individual declarations in isolation, authorities can identify patterns across similar products and shipments, potentially triggering much broader investigations.

The EU is moving in the same direction. The new Union Customs Code explicitly provides for primarily automated risk analysis, supported by the future EU Customs Data Hub. Implementation will be phased in, but the direction is clear: more connected data, more sophisticated analysis and greater ability to identify inconsistencies.

This changes the equation for importers. A classification process that has historically been considered good enough because it rarely attracted questions may no longer be sufficient. When customs challenges a classification, having an HS code in your ERP is not the same as being able to explain and substantiate why it is correct.

And that's precisely why the opportunity presented by AI matters now. We have the technology to make thorough, documented classification analysis operationally feasible, just as authorities are developing the ability to challenge our decisions more systematically.

For me, this makes defensibility a much more meaningful measure of success. The ambition should be to cover more of the product portfolio with validated, consistent and defensible classifications, without proportionally increasing the workload of the trade team. Speed still matters, but it is no longer the whole measure.

Defensible should not mean more work

All of this is not to say that you need to write extensive classification rationale memos for every classification in your system, AI or not.

For example, a company with 100,000 SKUs does not necessarily need 100,000 independent classification analyses. We see in practice that product variants and successor models share the characteristics that determine their classification. Where that is legally justified, one reviewed analysis can support an entire group of SKUs, avoiding excessive workloads to review AI generated classification analysis.

Experienced classifiers often recognize these similarities instinctively. At scale, however, that knowledge needs to be made explicit. Teams must define which product characteristics justify grouping, which differences require separate decisions, and whether a group remains valid across jurisdictions. The goal is not to combine products merely to reduce AI usage; it is to reuse sound reasoning without overlooking legally relevant distinctions. A group also needs a maintenance rule: when a new variant is added or a technical specification changes, who checks whether the original rationale still applies? Without that control, grouping can simply spread an incorrect decision across more products.

This changes the unit of work from the individual SKU to the classification decision, potentially covering many SKUs.

It also changes the business case. The number of AI analyses, the amount of expert review, and the additional work required for different jurisdictions all depend on how groups are defined and governed. A price per AI request multiplied by a SKU count may be a convenient estimate, but it is a weak basis for assessing operational feasibility or total cost before those choices are made.

AI can close product information gaps we have learned to live with

Grouping is one way to make more thorough classification analysis manageable at scale. Another is to rethink how we gather and use product information. AI creates an opportunity not just to improve the quality of our decisions, but also to reduce the effort needed to establish the facts behind them. Again, capturing that benefit depends on how we organize the work.

Classification teams rarely receive product data prepared specifically for customs purposes. Engineering and PLM records describe what is needed to design, manufacture, and manage a product. Tariff provisions may turn on material composition, function, or technical parameters that are not standard attributes in those systems.

Every company finds different ways to address this. Trade teams become product experts over time, understanding core product design, though most of it is undocumented, colloquial knowledge. Engineering teams are asked to fill out extensive product questionnaires for the trade team, indiscriminately asking product questions that may not even be relevant. Or anything in between.

AI can help expose and close this gap. Starting from available documentation, it reasons through to a suggested classification. When AI also has access to managed company context, such as previously validated product characteristics, technical documentation and earlier classification decisions, it can sometimes fill in missing details itself, using information the company already holds elsewhere.

And where the necessary information genuinely is not available, analysis may identify two possible headings and explain that the answer depends on a characteristic missing from the source material. Not only does this give you targeted questions to ask engineering, it also gives you the opportunity to systemically document the missing information.

That is a useful outcome, but it leaves work to be done to harness the opportunity. Who asks engineering for clarification? Can technical experts respond within the classification workflow, or will the discussion disappear into email? Where is the answer retained so it can be applied to related products?

The same design choice arises at the front end of the process. Should the trade team select the relevant specifications and drawings, or should every file attached to a SKU in the PLM be sent to AI? More documentation is not necessarily better. It may be contradictory, outdated, or irrelevant. The process must establish which product facts can be relied on and which remain assumptions.

These are not secondary implementation details. They determine both the quality of AI-assisted decisions and the workload created for the people supplying and reviewing the information. Done well, the process also builds a reusable record of customs-relevant product characteristics that the company did not previously capture. Over time, this can reduce repeated questions to engineering and improve the inputs available to the next classification, without needing to change the company's core PLM structure.

AI governance must be designed alongside the workflow

An enterprise operating model also needs to answer where product information may be processed, which AI providers are permitted, how outputs and source documents are retained, and who may review, approve, or share decisions. A standard cloud service and a customer-controlled deployment can have very different implications for feasibility and cost.

To my point above, deciding if all PLM documents on a product should be shared into the AI analysis by default – or whether the trade team makes a selection, is as much a question about usability and reliability as it is about data security, AI governance and operating cost.

In our experience, questions about roles, data and architecture are part of designing the solution, not an administrative step to leave until procurement. The answers may affect which people need access, which documents can be supplied to an AI provider, and whether the proposed workflow can be deployed at all.

Evaluate a future way of working, not just an AI engine

Before evaluating AI classification technology, I would encourage companies to think about what a genuinely better classification process should achieve. If success is still measured primarily by time saved per SKU, you risk buying a faster version of today's process. But what if success meant having a much larger share of your product portfolio covered by consistent, fully reasoned and defensible classifications, without requiring proportionally more expert effort? That ambition changes how you organize the work, involve other teams and ultimately evaluate the technology and its cost.

The best way to explore that opportunity is to combine some focused process design with hands-on testing. Let your classification specialists work with real products and documentation, explore grouping, review the AI's reasoning, and involve engineering when information is missing. Find out where the process needs to change, rather than simply measuring how quickly the system produces an answer. A quick software trial can demonstrate what the technology is capable of. Testing it within a deliberately designed workflow will tell you far more about what your organization could achieve with it.

AI has made a higher standard of classification achievable. The real opportunity is to rethink how we work to reach that standard.