We Let the AI Classify Alone – No Safety Net

TL;DR: We pulled every human hand off the wheel and let a single AI formula, =TWHSHINT, classify 7,344 real products on its own. It put the correct six-digit HS code in its shortlist 98.4% of the time, never once left the right chapter, and corrected mistakes in the answer sheet we were grading it against. And yet 98.4% is a number we would never put in front of a customs officer. Here is the whole story, and why it changes the economics of global trade.

There’s a moment every customs broker knows. A shipment is sitting at the port. A six-digit code is in dispute. And the difference between a clean release and a five-figure penalty comes down to a single distinction the product description never bothered to spell out: was the yarn single or folded? Was the truck 6 tonnes or 7? Was the fish fresh, chilled or frozen?

Harmonized System classification is one of the most brutal and least visible jobs in global trade. Over five thousand subheadings. Hairline splits between “knitted” and “woven,” between a finished article and a part of one, between a battery-electric car and a plug-in hybrid. The logic is beautiful and unforgiving. Get it right and nobody notices. Get it wrong and it follows your company around for years, as reassessments, audits and a compliance file with your name on it.

That one fact shapes everything we build: in tariff classification, there is no acceptable error rate. Not 5%, not 1%, not even 0.1%. We don’t treat HS classification as a clever feature. We treat it as a compliance obligation, and we take it seriously.

So when it came time to tell you what our engine can actually do, we refused to write a brochure. We built a courtroom instead. We took the answer key away, removed every human from the loop, and let the raw AI classify 7,344 real products entirely on its own. Then we graded it, line by line, against the actual HS-2022 nomenclature.

This article is the verdict.

What exactly did we test?

We tested one TariffWolf spreadsheet formula, =TWHSHINT, on its own. TariffWolf lives where your trade team already works, in the spreadsheet, and the toolkit ships as four spreadsheet formulas. This audit is about exactly one of them.

The name is doing honest work. =TWHSHINT returns a Hint: a tight, ranked shortlist of candidate HS-2022 codes for a product, generated entirely by AI from the title and description. It gives the codes most likely to be correct, so a classifier never starts from an empty cell; it is neither a verdict nor a stamped declaration. Read every figure below with that in mind. We are stress-testing the quality of a hint, alone, with nothing and no one assisting it.

We took the human out of the loop on purpose

The rules we set ourselves were deliberately merciless:

• 7,344 real product lines, spanning all 21 HS Sections and 96 chapters, from a live breeding stallion to a battery-electric truck.
• Zero human intervention. No analyst reviewed, nudged or corrected the engine. What you’re scoring is the raw, unassisted output of =TWHSHINT.
• No answer key. Every line was independently re-classified to the full six-digit level under the General Interpretative Rules (GRI 1–6), the legal rules for choosing a code, and only then compared with what the AI had proposed.
• Total enumeration, not a flattering sample: 100% of the qualifying population, so no easy categories got cherry-picked and no average had anywhere to hide.

This is as close as an audit gets to watching the machine think with its hands tied behind its back.

How accurate was the AI on its own?

On its own, =TWHSHINT had the correct HS-6 code in its shortlist for 98.4% of the 7,344 lines:

The correct six-digit code sat in the hint 98.4% of the time, and it was the very first candidate in roughly four of every five cases. Every one of the 7,344 lines landed in the correct heading and chapter. And across the entire run, the engine never once reached for a retired, pre-2022 or otherwise invalid code.

The most revealing result sits outside the table. When the engine did miss, it missed by one neighbouring subheading inside the correct heading. It never wandered off and filed a refrigerator under livestock. That containment is the whole game.

How does the engine reason toward a code?

The engine reasons from the product’s family down to the six digits, and the error pattern proves it. You might assume a 98.4% comes from sophisticated keyword matching. It doesn’t.

A keyword matcher fails randomly: a stray word drags “building block” into construction one minute and “apple” into fruit the next, scattering mistakes across the tariff. This engine fails structurally. Its candidate codes cluster, with 100% agreement, inside the correct four-digit heading and two-digit chapter. It first decides what the product fundamentally is (a garment, a vehicle, a machine), locks onto the right family, and only then reasons toward the terminal six digits. That is GRI reasoning. (For more on how the rules work, see reading the General Interpretative Rules correctly.)

You can see the same thing in how current it is. The HS-2022 edition introduced codes for things that barely existed a decade ago, and the engine reaches for them confidently and correctly: 8485 for additive-manufacturing (3-D printing) machinery, the 8704.4x–5x hybrid and 8704.60 electric goods-vehicle bands, 8703.80 for battery-electric cars, 8517.13 for smartphones, 0309 for fish flours fit for human consumption, 8541.51 for semiconductor-based transducers. It classifies against the tariff as it stands today.

And it knows when it’s sure. 95.8% of its determinations rest on a decisive attribute or direct nomenclature validation: a stated power rating, a fuel type and gross weight, a species and state. The remaining 4.2% are honest “I’m working from a thin description” calls that resolve to a heading’s residual Other subheading. The engine is calibrated about its own confidence.

How much faster is =TWHSHINT than a human classifier?

=TWHSHINT is comfortably a 20× speed-up on the classification process across a real catalog, and the spreadsheet is the reason. =TWHSHINT is an add-on function that calls the engine over an API straight from a cell. You point it at your product-description column, drag it down, and the candidate codes come back in bulk. The work that swallows a skilled classifier’s entire day (researching, cross-referencing, keying one SKU at a time) collapses into the time it takes to fill a column.

Judgement stays with your experts. The speed comes from deleting its slowest, most repetitive layer: the cold-start research on every single line. Your experts stop hunting for candidates and start doing the only thing that truly needs a human, which is choosing between them.

The plot twist: it corrected the answer key

This part we did not expect. On several lines, our own test sheet marked the AI “wrong,” and when a human went back to the nomenclature, the AI was right and the sheet was wrong. Frozen salmon (0304.81, not 0304.89). Live lobster (0306.32). Fish meal (2301.20). Liquid eggs (0408.99). Water heaters (8516.10). Engineered oak flooring (4418.75). A compound microscope (9011.20). A 100 kVA generating set (8502.12). In each case the correct code was already sitting inside the hint, waiting for someone to notice the cheat sheet had a typo.

An engine that occasionally out-reasons its own ground truth is doing classification.

Where does the AI get HS codes wrong?

All 121 misses cluster in the places that torment expert human classifiers too: the engineering-heavy chapters (machinery, electronics, vehicles, precision instruments, aircraft), where a single heading splits on one hard number the description often omits. A 7.5-tonne refrigerated van gets mis-banded across the goods-vehicle weight thresholds. A drone lands in the wrong weight class. The rest are fine biological or material splits, such as live tilapia read toward a carp code instead of 0301.99, or knitted-textile-upper footwear offered only as rubber/leather variants instead of 6404.19, plus the occasional semantic trap, like a plastic toy “building block” tugged toward construction materials.

We’re publishing the weak spots because they are a map. Every one tells you precisely where a thin product description needs one more attribute to be airtight. Sixteen of twenty-one Sections scored a flat 100%. Consumer-facing trade (garments, footwear, jewellery, food, cosmetics) was essentially flawless.

So why would we never ship 98.4%?

We would never ship 98.4% because a hint that’s right 98.4% of the time is a phenomenal starting line and a completely unacceptable finish line.

Return to the first principle: in compliance there is no acceptable error rate. =TWHSHINT was never designed to be the final word. It is the fast, edition-current, well-reasoned first draft that a trained classifier then reviews and confirms. And once you put that human back into the loop (the loop we deliberately removed for this audit), the number stops behaving like a probability at all.

It becomes 100%.

Not 99.9%, 99.99% or even 99.999%: one hundred. Because “almost always correct” is a phrase we are not willing to set beside a customer’s customs declaration. The AI does the heavy lifting at machine speed; the human guarantees the outcome. That pairing is the only configuration we are willing to call delivered.

Why this changes the economics of trade compliance

For decades, trade compliance has had to pick one of two bad options. Slow and safe: armies of experts classifying by hand, accurate but expensive and impossible to scale to a catalog of fifty thousand SKUs. Or fast and risky: crude keyword tools and stale code lists that move quickly and plant errors that surface years later in an audit.

=TWHSHINT removes that trade-off. It is fast like a machine and structured like an expert: edition-current, GRI-grounded, calibrated about its own confidence, and contained enough that even its rare misses stay inside the right heading. It runs natively in the spreadsheet your team already uses, scales to an entire catalog in a single column, and it is verifiable: you can audit it exactly the way we just did. When the bottleneck of global trade becomes 20× cheaper to move through without loosening the compliance standard, the economics of an entire industry shift.

Frequently asked questions

What is =TWHSHINT?

=TWHSHINT is one of TariffWolf’s four spreadsheet formulas. It calls the TariffWolf engine over an API from a cell and returns a ranked shortlist of candidate HS-2022 codes, generated by AI from a product’s title and description. It is a hint for a classifier to review and confirm before anything is declared.

How accurate is AI HS classification without human review?

In TariffWolf’s internal audit of 7,344 product lines with no human intervention, =TWHSHINT had the correct HS-6 code in its shortlist 98.4% of the time (7,223 of 7,344) and ranked it first 81.1% of the time. It landed in the right four-digit heading and two-digit chapter on 100% of lines and emitted no deleted or invalid HS-2022 codes. There were 121 genuine misses (1.6%).

Which products does AI find hardest to classify?

In this audit, all 121 misses were in engineering-heavy chapters (machinery, electronics, vehicles, precision instruments, aircraft) or in fine biological and material splits. Typical cases were a 7.5-tonne refrigerated van mis-banded across goods-vehicle weight thresholds, a drone in the wrong weight class, and live tilapia read toward a carp code instead of 0301.99. Each miss marks a place where a thin product description needed one more attribute.

Was the TariffWolf accuracy audit independent?

No. TariffWolf ran the audit internally, with its own engine and methodology. Every line was classified independently of the engine’s output under GRI 1–6 before comparison, and every figure recomputes from the raw data TariffWolf provides with this post. The suggested check is to run the same audit on your own catalog against your own known-good codes.

Don’t trust us. Test us.

One last piece of honesty: we ran this audit ourselves. It’s an internal review, with our engine, our methodology and our hands on the wheel. We were careful, we classified independently of the engine’s output, and every figure recomputes straight from the raw data we’re handing you. But we’re the home team grading our own exam, and you’d be entirely right to keep one eyebrow raised.

Good. Keep it there, then settle it the honest way.

Drop a column of your products into =TWHSHINT and run the exact audit we ran, on your own catalog, against your own known-good codes. Watch how often the right answer is already in the hint. Time how much faster your team moves. Find your own 1.6% before it ever costs you a clearance. That is the entire point of a hint you can verify.

Keep the percentage at arm’s length. We are handing you the red pen.