Classify products: what software can take over and what still needs a classifier

Software and AI can gather product attributes and propose tariff and export control codes. A trained classifier must still decide new or disputed items and answer for the result. Neither works until product master data says plainly what each item is made of and what it does.

What the work actually involves

Classifying a product means assigning the codes that govern how it crosses a border. That usually covers a tariff code under the Harmonized System and its national extensions. It also covers an export control classification, and often a country of origin determination. Each code drives the duty owed and whether a licence is needed. It also decides which preference programmes apply. A wrong code is a compliance failure. Customs will treat it that way whether the error came from a person or a model.

In practice the work starts when a new item is created or an existing one changes. Someone reads the specification and compares it against the legal text and the explanatory notes. Past rulings get checked, a code is picked, and the reasoning is recorded. The record matters as much as the code, because an auditor will ask for the rationale long after the decision.

Where software and AI can carry the load

Collecting the inputs is the easiest step to hand over. A system can pull material composition, technical drawings and supplier declarations into one view. That spares the classifier from chasing engineers by email.

Suggesting a code is also realistic today. Tools trained on past classifications and on the tariff text can propose a likely heading with a confidence indication. For repeat items, colour variants and simple goods, that suggestion is often right. It should still be handled as a draft for review.

Monitoring is another good fit. When the tariff schedule is revised or a control list is updated, software can find every affected item and queue it for attention. It can also spot conflicts, such as near identical parts carrying different codes, or a code that does not exist in the destination country's schedule.

Record keeping can be fully automated. Every decision should be captured along with its evidence and the name of whoever approved it, without anyone retyping it.

Where a person has to decide

Genuinely new products need a human. So do goods that sit between headings, kits and sets, and anything where the interpretive rules must be applied in sequence. A model can explain its choice. It cannot carry legal responsibility for it.

Export control decisions deserve extra caution. Whether an item meets a technical threshold on a control list often depends on performance data that only engineering can confirm. Getting that wrong has consequences well beyond a duty adjustment.

A person also has to own requests for binding rulings, responses to customs queries and any decision to correct past declarations. These involve judgement about risk and disclosure that should not be delegated.

What has to be true about the data first

The item master must describe the physical product, with composition and intended use held in structured fields. Marketing descriptions and internal nicknames are useless to a classifier, human or machine.

Each good should map to a single record. Duplicate part numbers for the same thing guarantee inconsistent codes.

Historical classifications need their rationale attached. A model trained on bare codes learns past mistakes as readily as past good calls. If earlier decisions were never documented, clean a sample by hand before training anything on it.

Codes must be stored per country. A single global field invites the wrong national extension to travel with the item.

Every attribute needs a clear owner. When engineering changes a material, the classification team should hear about it automatically.

Questions to ask the people who run it

What gets looked at first when a new item arrives, and where does that information live?

Which products are classified from memory, and which get researched every time?

When engineering changes a design, how does the classification team find out? Does it find out at all?

Who gets called when the specification is unclear, and how complete is the answer that comes back?

Are there items where brokers or freight forwarders apply a different code from the internal one? Who reconciles that?

Where is the reason for a code written down? Could a colleague find it?

Which workarounds exist because the system will not hold the data the team needs?

What would be most worrying about a tool proposing codes?

Where changes tend to go wrong

Teams often automate suggestion before fixing the item master, then blame the tool for poor output. Others let reviewers approve machine proposals in bulk, which quietly turns a control into a formality. Review thresholds belong at the level of product risk. Keep a sample of automated decisions under periodic human audit, and make sure the auditors still classify enough items themselves to keep their judgement sharp.

APQC's Process Classification Framework® (PCF) is an open standard developed by APQC, a nonprofit that promotes benchmarking and best practices worldwide. To download the full PCF or to view definitions and measures, please visit www.apqc.org/pcf.