AI classification tools for customs brokers: what they cost and how to back-test one

Every few weeks another AI tool promises to classify goods for you. Some are cheap, some are expensive, and almost all publish accuracy claims without saying how they were measured. This post puts published prices side by side, explains the one test that tells you whether a tool works on your goods, and gives a simple way to decide whether it's worth paying for.

What the tools cost

Published prices as of 2026. All are vendor figures, and some are reported by a competitor, so check the vendor's own page before you budget.

Two things stand out. Entry-level prices are low, so the subscription is rarely the real cost; your broker's review time is. And if you already run CargoWise or Magaya, test the AI features in the system you own before adding another subscription.

Why vendor accuracy numbers don't settle it

Vendor accuracy figures usually come without a method: which goods, which digit level, compared against what. Your mix of goods is not their test set. A tool that does well on consumer electronics may struggle with your client's machine parts or chemicals.

The legal side raises the stakes. Under CBP ruling H350722, 8- and 10-digit classification has to be directed by a licensed broker (ST&R Trade Report, Mar 2026). So the question is never "can it replace review?" It's "does it make review faster without making it worse?"

The back-test: measure it on your own entries

You can answer that in a week with entries you've already filed.

1. Pull a sample

Take 200 or so invoice lines from entries your broker has already classified and filed. Mix clients and product types. Include the hard ones.

2. Hide the answers

Strip out the final codes. Keep the product descriptions as they arrived, messy wording included, because that's what the tool will see in real life.

3. Run the tool

Record the tool's suggested code, its confidence score if it gives one, and its reasoning.

4. Score at three levels

For each line, compare the suggestion with the filed code at 6, 8 and 10 digits. A tool that's right at 6 digits but wrong at 10 can still save research time, but it can't be trusted to finish the job.

5. Look at where it's wrong

Sort the misses by duty impact. One wrong code on a high-duty line costs more than ten wrong codes on duty-free lines. Check whether the tool's low-confidence flags actually caught its mistakes; a tool that's confidently wrong is the dangerous kind.

6. Time the review

Have an entry writer work a batch with the tool's drafts and a batch without. The question you're answering is whether review time per line actually drops.

7. Write it down

Keep the results with the tool name and version. If anyone later asks how you checked the tool, this is your answer. Under the 2022 Part 111 rules, a brokerage needs "a sufficient number of licensed brokers" exercising responsible supervision and control (GHY International, Oct 2022), and a documented test shows how that supervision works.

Re-run a smaller back-test whenever the tool updates or tariff rules shift.

Doing the math

Small brokerages earn per entry, so time per entry is the unit that matters. One published fee guide lists formal entries at $150 to $250, complex entries at $250 to $400, and entry amendments at $50 to $150 (Suaid Global, a brokerage services seller, Sep 2026). Before the 2025 tariffs, brokers charged about $4 to $7 per HTS code (Reuters via AJOT, Jun 2025).

A simple worksheet, using your own numbers:

1. Lines classified per month.
2. Minutes of review per line today, from step 6 of the back-test.
3. Minutes per line with the tool's drafts.
4. Minutes saved per month, times the loaded hourly cost of the person doing it.
5. Minus the subscription and the time spent on set-up and re-testing.

If line 4 doesn't clearly beat line 5, don't buy. If the back-test showed the tool is often confidently wrong on your goods, don't buy at any price, because the time it saves on easy lines gets spent catching its mistakes.

A short decision guide

• Your system of record already has AI classification: back-test that first.
• Low volume, varied goods: a pay-per-lookup tier is enough. Use it as a research aid.
• High volume, repeat goods: the bigger gain often comes from a reviewed parts or product catalog with approved codes, with AI handling only new items.
• Any tool, any volume: your licensed broker still decides every 8- and 10-digit code, and you keep the draft and the decision in the entry file for at least the 5-year retention period in 19 CFR 111.23).

Related reading

• Where AI fits in a small brokerage
• The H350722 review checklist
• Drafting CAPE refund declarations

Lunan is not a law firm or a licensed broker, and nothing here is legal advice. Lunan doesn't sell a classification tool; we help brokerages set up and back-test the one they pick. If you'd like a hand running this test on your own entries, start a free conversation.