The Four Words That Will Change How Classification Gets Done

Why we built two engines, four formulas, and a 20x speed-up — all of it living in the cell next to your product description.

— Surajit Roy, Co-Founder, TariffWolf

There is a particular kind of tired that only a classifier knows.

It is the tiredness of staring at row 4,217 of a product catalog at 6:47 p.m. on a Thursday, knowing that rows 4,218 through 9,800 are still waiting. It is the tiredness of having three browser tabs, two desktop applications, and a PDF of the tariff schedule all open at the same time, and somehow none of them are telling you what you actually need to know about this product.

I lived in that tiredness for the better part of a decade. I am writing this so that you do not have to.

The Tools Were Never the Problem. The Switching Was.

Every classifier I have ever worked with has the same muscle memory.

You read a product description. You think you know the chapter. You alt-tab to the classification tool. You type in a search. You scroll. You alt-tab back to the spreadsheet to check whether you already classified something similar last month. You alt-tab to the tariff schedule PDF to read the chapter note. You alt-tab to a browser tab where you have the WCO Explanatory Notes open. You alt-tab back to the tool to drill into the heading. You alt-tab back to the spreadsheet. You paste the code. You move to the next row.

That entire dance for one line item.

Now multiply that by the 800-row supplier catalog that landed in your inbox this morning. Multiply that by the broker’s manifest your colleague needs by tomorrow. Multiply that by the audit your compliance lead is preparing for next week. Multiply that by the new SKUs the sourcing team just added on Slack.

The bottleneck was never the thinking part. Classifiers are good at thinking. The bottleneck was the switching part — the constant, exhausting context-switching between the tool, the schedule, the spreadsheet, the notes, the browser, the email thread, and back to the tool again.

We measured it once on a real project: a senior classifier was spending 73% of her working time on screen navigation and 27% on actual classification judgment.

Read that number again. Seventy-three percent of an expert’s day, gone, to the gap between windows.

That is what we set out to close.

https://youtu.be/Vj9JcXimCpQ/

So We Did Something That Sounds Obvious in Hindsight

If classifiers live in the spreadsheet, the classifier’s tool should live in the spreadsheet too.

Not “exports to” the spreadsheet. Not “integrates with” the spreadsheet. Not “syncs to” the spreadsheet with a button you have to remember to press.

In the spreadsheet. As a formula. Next to the product description. With the answer appearing in the cell to the right.

That is TariffWolf.

But the obvious idea is the easy part. The hard part is making it actually work — and making it work in a way that matches how classifiers really think.

So we built four formulas, not one. And we built them on top of two different engines, each tuned for a different moment in the classifier’s day.

Let me tell you about them the way I would tell a colleague over coffee.

Two Engines, Because Classification Is Not One Job

When you are deep in the work, you know that “classifying a product” is not a single action. It is at least two different actions that happen at different speeds, for different reasons.

There is the scouting action — the one you do when a fresh supplier catalog lands and you need to know, in the next ten minutes, whether these 600 SKUs are mostly textiles, mostly electronics, or a mess of both. You do not need the full 10-digit HTS code at this stage. You need to know the chapter family. You need to bucket. You need to triage.

And then there is the commitment action — the one you do when a shipment is going out on Friday and the broker needs the full national tariff code that will print on the customs entry. Here, “close enough” is not good enough. You need the right code, fully formed, ready to file, and you need to be able to defend it if customs asks why.

These are two different jobs. We built two different engines for them.

tw-hs-sprint - the engine for scouting

Sprint is fast, cheap, and country-independent. You feed it a product description. It hands back up to five candidate HS-6 codes — the international six-digit prefix that is the same everywhere in the world.

You use sprint when you are still figuring out what the product is, before you commit to where it is going. You use it to bucket a 600-row supplier catalog into chapter families in a single afternoon. You use it to give the sourcing team a same-day answer when they ask “is this going to be a problem to import?” You use it when you want to spend half a credit instead of two.

Sprint costs 0.75 credit per unique product description. Cached descriptions are free.

tw-hs-intelligence - the engine for production

Intelligence is the engine that goes all the way. You feed it a product description and a destination country, and it gives you the full national tariff code — the 10-digit HTSUS for the United States, the 10-digit CN code for the European Union, the 8-digit AHTN code for ASEAN, and so on, depending on the country.

But here is what I love most about intelligence: it knows when it is unsure.

If the description you gave it is ambiguous — if the engine cannot tell whether your “table lamp” is mains-powered or battery-powered, whether it is for industrial use or consumer use — it does not silently guess. It pauses. It comes back with two or three short questions. You answer them in the next cell. It finishes the classification.

This is the part that took us the longest to build, and it is the part I am proudest of. Because it is what an experienced human classifier does. When you do not have enough information, you do not invent it. You ask.

Intelligence costs up to 2.00 credits per product. And — this is the part I want you to hear — those 2.00 credits are an envelope, not a per-formula charge. Everything you do with the same product after that first classification (asking for the justification paragraph, submitting clarification answers, re-querying the cached result) is free.

Four Formulas, Each One Doing One Job Well

Now the formulas. There are four of them. You only need to remember four words.

=TWHSHINT — "give me a hint"

This is the scouting formula. It runs on sprint. You type:

=TWHSHINT("BRD-001", "wooden chopping board, 35x25 cm, with steel handle")

And the cell fills with:

4419.90, 4419.20, 4421.90

Three candidate HS-6 codes, comma-separated, ready to scan. Drag the formula down 600 rows and you have a triaged catalog in the time it takes to drink a coffee.

This is the formula sourcing teams will use. Category managers will use it. Anyone who needs to bucket fast, before committing.

=TWHSCODE — "give me the real code"

This is the workhorse. It runs on intelligence. You type:

=TWHSCODE("BRD-001", "wooden chopping board, 35x25 cm, with steel handle", "US")

And the cell fills with:

4419.90.91.00

The full 10-digit US HTS code. Ready to put on the entry summary. Ready to send to your broker.

If the description is too thin to commit on, the cell fills instead with a short list of questions — written in plain English, the way a senior classifier would ask them. Which is to say: brief, specific, and decisive. You answer them in the next cell with the third formula.

=TWSUBANS — "here are my answers"

When =TWHSCODE comes back asking for clarification, you do not start over. You do not lose your place. You just answer.

=TWSUBANS("BRD-001", "stainless steel handle", "domestic kitchen", "no coating")

The cell fills with the final code, and your row is done. The clarification round costs nothing extra — it is included in the 2.00-credit envelope you already paid when =TWHSCODE paused.

This is multi-turn classification, the way humans actually do it, brought into the spreadsheet.

=TWCLJUST — "tell me why"

The fourth formula is the one customs officers love and the one auditors thank you for.

=TWCLJUST("BRD-001")

It returns a written paragraph — about 150 words — explaining why the chosen code is the correct one. It cites the applicable General Rule of Interpretation. It references the chapter note. It walks through the reasoning the way a compliance memo would.

Free, on demand, generated only when you ask for it. Included in the same 2.00-credit envelope.

You can leave the column blank for 99% of your catalog and only fire =TWCLJUST on the rows that compliance flags for review. Or you can fill the column on every row of a shipment manifest and hand the whole sheet to your customs broker as a self-documenting audit trail. Either way, it is sitting in the spreadsheet, in the cell, where you can read it.

What 20x Actually Looks Like

I do not like throwing numbers around without showing my work. So let me show you.

The classifier I mentioned earlier — the senior one I measured at 73% screen-navigation overhead — averaged about 8 fully-classified line items per hour on a complex catalog. That is roughly 7.5 minutes per item, of which 5.5 minutes was navigation and 2 minutes was actual judgment.

With TariffWolf, the navigation collapses to zero. You write the formula. The code appears. If the engine needs clarification, you answer in the next cell — still inside the spreadsheet, still no alt-tab. Your only remaining cost is the seconds spent reading the description and the few seconds spent reviewing the answer.

We have measured this in real catalogs across real customer teams. The honest number is between 150 and 200 fully-classified line items per hour, depending on how clean the descriptions are coming in.

That is 20x. Sometimes 25x. On a clean catalog where most rows do not need clarification, we have seen 30x.

But the speed number is not even the part that matters most.

The part that matters most is this: classifiers stop being interrupted.

They stay in the spreadsheet. They stay in flow. They stay in the headspace where they can spot the strange row, the edge case, the one shipment that does not look right. Because they are not spending 73% of their day swimming between windows, they have the cognitive room to actually catch the things that matter.

That is what we built. That is why we built it.

What This Means for the Industry

I have classified across e-commerce, automotive, freight forwarding, internet companies, and consumer electronics. I have used the big enterprise platforms. I have used the proprietary in-house tools. I have built classification workflows from scratch more than once.

In every single one of those environments, the classifier had a spreadsheet open on the side.

That spreadsheet was where the real work happened. Everything else was where the data lived.

TariffWolf collapses that distinction. The work and the data are now in the same place. The classifier’s judgment and the engine’s intelligence meet in the cell, side by side, in the column right next to the product description.

I think this changes the industry for three reasons.

First, it removes the onboarding tax. A new classifier joining your team does not need to learn a new platform. They need to know =SUM, =VLOOKUP, and now four more words: =TWHSHINT, =TWHSCODE, =TWSUBANS, =TWCLJUST. That is the entire learning curve. People who change jobs every three or four years — which is most of our industry — finally get a tool that travels with their skill set instead of resetting it.

Second, it democratizes good classification. The senior classifier on your team and the junior one on your team now have the same engine in the cell next to them. The junior gets the candidate codes, the questions when ambiguity arises, and the written justification — three things that used to require a senior to provide. The senior is freed up to handle the genuinely hard cases, the audits, the strategy. The team’s overall quality goes up and its throughput goes up at the same time, which almost never happens with a single tool change.

Third, it makes compliance auditable by default. Every classification carries a request ID. Every classification can have a written justification attached on demand. Every classification is reproducible from the original product description plus the destination country. Three years from now, when customs asks why a 2026 shipment was classified the way it was, you will not be reconstructing it from memory. The answer is in the spreadsheet, in the column next to the code, where it always was.

Where to Start

If you have read this far, you already know whether this is for you.

If you are a customs broker, freight forwarder, sourcing manager, or compliance lead — and if you have ever kept a spreadsheet open next to your “real” classification tool — then TariffWolf is for you.

The setup takes about 60 seconds. You paste a single Apps Script file into your Google Sheet. You set your API key. You reload. The four formulas become available in any cell of any tab. Your existing spreadsheets — the ones you have been using for years — light up with classification intelligence the moment you type =.

Email us at api@tariffwolf.com to request an API key. Tell us which tier you want. We will mint a key, send you the welcome email with the script attached, and you will be classifying inside the spreadsheet within the hour.

That is it. That is the whole pitch.

We built TariffWolf because the spreadsheet refused to leave the classifier’s desk. So instead of fighting it, we moved in.

Welcome to classification intelligence that finally lives where you do.

— Surajit Roy, Co-Founder, TariffWolf

Want to see it live? Book a 30-minute walkthrough and we will run the four formulas against a sample from your own catalog. No slides. No demo data. Your products, classified in front of you.