Why classification risk is a control problem, not a lookup problem
Most HS classification failures are not wrong answers. They are right answers that nobody checked again after the one person who knew them left, a tariff changed, or a new free trade agreement opened a window nobody was watching. Here is why classification is a control problem before it is a lookup problem, and what a working control actually looks like.

By Saurabh Goyal, Founder & CEO of Phlo Systems. Published 20 November 2026.
Ask an SME importer or exporter how confident they are in their commodity codes, and most will say they are fine: someone experienced set them up, and nothing has gone wrong yet. That answer describes a snapshot, not a control. The actual failure mode in classification is rarely "we picked the wrong code on day one." It is "the code was right on day one, and nobody has looked at it since."
The 30-second answer:
- Classification is usually not wrong because of bad judgement. It is wrong because the judgement was never repeated after the first time.
- The two most common triggers for a stale classification are a tariff change the business did not track, and a new free trade agreement eligibility window it did not know to check.
- When classification lives in one person's head or a spreadsheet, the business loses the classification the moment that person is unavailable, not just when they leave.
- A recent industry figure puts this in context: Thomson Reuters' own launch material for its generative-AI classification tool states that only 7% of organisations currently use technology to manage tariff changes, and just 5% use it for classification changes, despite AI interest in the space having grown sevenfold.
- The fix is not a better one-off lookup. It is a standing control: who reviews what, on what trigger, with what record.
Why a correct classification goes wrong later, not on day one
A commodity code assigned by someone who knew the product and the tariff schedule at the time is usually a correct answer to the question as it stood then. The problem is that the question does not stay still. Tariff schedules are amended. Trade agreements add, remove or reshape eligibility. A product's bill of materials or country of manufacture changes as sourcing shifts. None of these events announce themselves to the business that holds the classification; they simply make a previously correct code wrong, silently, until a shipment is stopped or a duty bill does not match expectation.
For a large enterprise with a dedicated trade compliance function, this is a known risk with a known answer: a standing team that monitors tariff and agreement changes and re-tests classifications against them. For the SME importer or exporter without that team, the same risk exists with no equivalent control, and it tends to surface in one of two ways: a shipment held at the border on a classification query, or a post-clearance audit finding years of a wrong rate applied retroactively.
The adoption gap is the real story
It would be easy to assume this is a solved problem now that AI tools can read a product description and suggest a code. The adoption data says otherwise. Thomson Reuters, introducing its own generative-AI classification product, states plainly that only 7% of organisations currently use technology to manage tariff changes, and just 5% use it for classification changes, even as interest in applying AI to the problem has grown roughly sevenfold. That is not a tooling gap caused by AI being new; it is a gap between a capability that now clearly exists and a process that, for the overwhelming majority of businesses doing their own classification, still does not.
For an SME without a dedicated tariff team, that gap usually shows up as exactly the failure mode above: classification done once, correctly, by whoever was available at the time, with no mechanism for anyone to be told when a tariff line or an agreement eligibility rule underneath that code changes.
What a working classification control actually looks like
- A record of who classified what, when, and on what basis, not just the code itself. If the reasoning and the date are not recorded, nobody can tell later whether a code needs re-testing or is simply old.
- A trigger for re-review that is not "whenever someone remembers." A tariff schedule update, a new or amended free trade agreement, or a change in supplier or country of origin should each independently prompt a re-check of the affected codes, rather than waiting for the next audit or shipment problem to surface it.
- Continuity that survives a person leaving. If the classification logic for a product line exists only in one person's head, the business has a single point of failure it may not notice until that person is gone and a new shipment needs a decision nobody else can make with confidence.
- An audit trail that answers a customs query without reconstruction. When an authority asks why a particular code was used, the answer should already exist in a record, not need to be rebuilt from memory or old emails under time pressure.
None of this requires AI to be the answer by itself. It requires classification to be treated as a standing control with an owner and a trigger, the same way a business would treat any other compliance obligation that does not go away after the first correct answer.
Frequently Asked Questions
Is HS classification a one-time task or an ongoing one?
Ongoing. A classification that was correct when first assigned can become wrong later without anyone changing the product, because tariff schedules, trade agreement eligibility, and sourcing can all change independently of the product itself. Treating classification as done once a code is assigned is the single most common cause of a classification that quietly goes stale.
Why do AI classification tools exist if adoption is still so low?
The tools solve the lookup problem: reading a product description and proposing a plausible commodity code quickly. Adoption has lagged not because the tools do not work, but because most businesses, especially SMEs without a dedicated tariff team, have not built the surrounding process that decides when a classification needs to be looked at again, which is a separate, harder problem than the initial lookup.
What is the actual business risk of a stale classification?
Two distinct risks: an import held or queried at the border on a classification question, which is disruptive but usually resolved quickly, and a post-clearance audit finding that a wrong rate was applied across a run of past shipments, which can mean duty reassessed retroactively plus potential penalties, discovered long after the shipments in question have cleared.
How is this different from a free trade agreement eligibility check?
Classification assigns the commodity code; FTA eligibility is a separate test, applied after classification, asking whether the product qualifies for a preferential rate under a specific agreement's own rules. Both need the same kind of standing control, because both can go stale independently of whether the underlying product has changed.
How Phlo Systems helps
tradePhlo Intelligence keeps classification and FTA eligibility as a standing, auditable record rather than a one-time lookup: every code carries its reasoning and the date it was set, and tariff or agreement changes affecting a business's own product lines surface automatically rather than depending on someone remembering to check. See tradePhlo Intelligence.
Related reading:
- What is FTA origin determination, and why is it harder than HS classification?
- What is landed cost, and why is my customs broker's number always wrong?
Saurabh Goyal is the Founder & CEO of Phlo Systems. He spent 12 years building CTRM and ERP systems for global commodity trading houses before founding Phlo in 2016.
Want to learn more about Phlo Systems?
See how our platform digitises international trade for commodity traders, importers, and exporters.
Get Started
