Hello, this is Ryuta Hamamoto from TIMEWELL.
The short answer first. When you choose an AI tool for export classification, what matters more than an accuracy number is whether your reviewer can check its conclusion and whether you can try it on your own products. AI speeds up the groundwork: reading specification sheets, narrowing down which entries to check, and organizing the rationale. The final classification belongs to the exporter that bears legal responsibility under Japan's Foreign Exchange and Foreign Trade Act (FEFTA), meaning your export control officer. No amount of tool progress changes that split.
This article covers how to choose an AI tool and how to roll it out. The legal treatment of feeding drawings and specifications into generative AI (the deemed-export question), hallucination safeguards, and the design of an AI-assisted flow are covered in 20 FAQs on generative AI and export control. I leave those to that article and stay with what to compare and how to start. Before you read on, a three-minute export-control readiness check helps you map the gaps against your own program.
Working sheet for the EAR path: A fill-in sheet that walks "subject to the EAR? → ECCN on the CCL? → EAR99?" plus embargoed destinations, end-use and end-user traps, de minimis and foreign direct product rules, counterparty screening, and a two-track cross-check with Japan's FEFTA / Appended Table 1. Record basis and outcome in one place for internal audit or customer requests. → Download the EAR Classification Flow & EAR99 Practical Checklist (2026) (Free. Company name and work email required.)
Then and Now: What Has Changed Since the September Edition
| Item | September edition | Now (November 6, 2026) |
|---|---|---|
| Japan's ordinance revisions | Reflected the revisions effective February 14, 2026 | Unchanged. METI's matrix table page still shows the February 14, 2026 version as of November 6, 20261 |
| Retention of classification records | "Seven years or more" | Seven years as a rule; five years for goods and technologies in entries 5 to 16 of the appended tables2 |
| U.S. BIS 50% rule (relevant to ownership checks on counterparties) | Status as of early October | Stayed on the Federal Register through November 9, 2026. A law-firm report as of late October says no rule implementing an extension had been published3 |
| Product descriptions | Phrases such as same-day reflection of each country's rules | Narrowed to claims I can support. See "Where TRAFEED Fits" below |
What Classification Is: Two Kinds of Control
Classification means deciding whether goods you plan to export, or technology you plan to provide, fall under export-control regulation, and documenting why. Japan's system has two pillars.
The first is list control. Goods and technologies in entries 1 to 15 of Appended Table 1 of the Export Trade Control Order need a METI license depending on the destination. The second is the complementary export control (catch-all), under which an item that is not on any list can still need a license because of its end use or end user. It corresponds to entry 16. How to read the entries is in a guide to Appended Table 1, and the requirements for catch-all are in what the catch-all control is.
The lists change often; the Goods and Technologies Ordinance, among others, was revised with effect from February 14, 2026. METI's matrix table, the usual starting point for checking, is replaced each time the version changes, so the premise is that you never keep using an old copy. How to use it is explained in the matrix table: how to use it and where it trips people up. For readers who work mainly with the U.S. regime, the parallel path is in What Is an ECCN?.
Replace siloed classification work with AI.
METI's FY2024 data shows 52% of foreign exchange law violations stem from classification errors. Download the TRAFEED product catalog covering features and rollout.
Why Classification Gets Heavy, and Where AI Fits
What consumes time in practice is the groundwork, not the decision itself. You collect specifications from engineering, translate product specs into the vocabulary of the law, look for the entries that might apply, and compare against the numerical thresholds in the text. One item can take hours to days. If the knowledge sits with one person, the process stalls when that person is away. Records vary too: forms and supporting documents differ from case to case.
AI fits into that groundwork: extracting numbers from specification sheets, narrowing down candidate entries, citing the supporting provisions, and drafting the certificate. What stays with people is judgment where interpretation splits, the final controlled or non-controlled call, and approval. Even when the AI says "not applicable," the reviewer should check the cited provision against the original text. Skip that step and an error flows straight into the certificate.
What to Check When Choosing an AI Classification Tool
Start with the work your team needs to improve: collecting specifications, finding relevant control-list entries, documenting reasoning, or routing approvals. Ask for a demonstration using a representative specification sheet so you can assess the steps between input and human review. A polished demo alone does not establish whether the tool fits your workflow.
Check whether reviewers can inspect the sources, assumptions, and missing information behind an answer. Ask what happens when specifications are incomplete: does the tool request clarification, or return a conclusion anyway? Define who reviews the output and which records your team needs to retain.
Compare costs using your expected workload. For a points-based plan, confirm points consumed per investigation, monthly allowances, additional usage charges, and setup fees. An effective unit price calculated at full utilization is different from your cost at lower usage. Request examples for both a typical month and a busy month, separating routine checks from deeper investigations.
Before adopting a tool, confirm where technical information is stored, how access is controlled, and how the service fits existing approvals and records. Verify that the specific features you need are available under the proposed contract, including their delivery dates.
Also ask which countries' rules the tool actually classifies under, and where classification ends and list screening begins. Does it classify under Japan's rules, handle other countries' classifications such as U.S. ECCNs, or only match counterparty names against lists? A tool that is vague on this line produces a gap between expectation and reality. If you look up ECCNs often, keep the ECCN cheat sheet nearby.
Rolling It Out: Start with a Small Trial
I think rollout goes better if you do not start with every product. The order I recommend:
- Set the scope of the trial. Pick a few dozen past cases you have already classified, and include some where interpretation was contested, not only easy ones.
- Compare the AI output with your existing determinations. Do not only count matches; for each mismatch, have a reviewer decide whose rationale held up.
- Decide the review steps and the records. Who checks the cited provisions against the original text, and who approves? In what form do you keep the certificate, the supporting documents, and the AI output? Retention is seven years as a rule, and five for entries 5 to 16.2
- Widen the scope. Add frequently seen product families in stages, and set a date about six months out to review.
- Connect the flow to the paperwork others ask for. If a classification comes out non-controlled, how to write the certificate for customers is covered in how to write a certificate of non-applicability.
After rollout, check regularly that your records would stand up to a METI on-site inspection or a customer audit. The export control audit preparation checklist works for that. Other ways to streamline the classification work itself are in five steps to streamline classification.
Where TRAFEED Fits
TRAFEED is TIMEWELL's export-control AI agent (TRAFEED catalog (PDF)). In Japan's security export control field, it is the world's first export-control-specific AI agent as of March 2026, by our own research.
The claims I can support are these. We ran a joint validation with Okayama University for Japanese security export control classification. The mechanism that supports classification is covered by a patent (Japanese Patent No. 7862062). As of July 2026 it is used at more than 20 organizations across companies and universities. That accuracy figure refers to Japanese classification; it is not a figure for U.S. ECCNs or other countries' classifications.
The functions are first-pass analysis of Japanese classification under FEFTA (candidate entries with cited provisions), counterparty screening against published lists from various countries, and draft certificates with record keeping. It does not classify under foreign law such as the U.S., EU or Chinese regimes. Counterparty screening covers published lists from the U.S., EU, UN and others, and lists published by the Chinese government are screened as reference information. The final classification is made by your export control officer.
On security, TIMEWELL holds ISO/IEC 27001 certification for the registered scope of AI SaaS product planning, development, and provision. Main processing runs on AWS in Japan, but confirm where data resides, including the use of external services, against your own security questionnaire.
If you are considering it, review the pricing and service information, then discuss your workflow. Bring the types of products you handle, your monthly review volume, and the documents you use today. These details make the discussion about your actual process more useful.
Summary
- AI classification tools support reading specifications, narrowing candidate entries, citing rationale and drafting certificates. The final decision and approval stay with the export control officer
- Compare tools on whether reviewers can verify conclusions, whether you can test on your own products, and whether the line between classification and screening is clear, rather than on accuracy figures
- Start rollout with a trial against a few dozen past cases, set the review steps and records, then widen
- Retain classification records for seven years as a rule (five for entries 5 to 16). No further ordinance update was visible as of this check, after the February 14, 2026 revisions
- Price on your own volume, with a typical month and a busy month, not list price
Whichever tool you pick, I think the first month is for practicing doubt, not trust. Look closely at the cases where the tool and your people disagree, and you will often find the weak spot is in your own procedure.
References
- BIS, Commerce Control List and EAR resources
- eCFR, 15 CFR Parts 730–774
- METI, Security Export Controls
- CISTEC
- e-Gov, Export Trade Control Order
Footnotes
-
METI, "About the Matrix Tables for Goods and Technologies" (last updated February 18, 2026; checked November 6, 2026) ↩
-
METI, "Compliance items under the FEFTA" (records management; Article 1, item 2(ri) of the Exporter Compliance Standards Ordinance; Japanese) ↩ ↩2
-
Dechert LLP, "BIS Affiliates Rule: November 10 Reimposition Remains Scheduled Despite Expected Extension" (October 28, 2026) ↩






