TorokuSentry

1. One-liner

Watches every Japanese supplier’s invoice registration for silent lapses, and flags the ledger note that saves the deduction.

2. Trend signal — why now?

Three things land on the same date, and they compound.

1 October 2026 — the transitional deduction drops from 80% to 70%. Buy from a non-registered supplier and you can no longer credit the full consumption tax. The FY Reiwa 8 tax reform restaged the whole run-off: 70% from Oct 2026, 50% from Oct 2028, 30% from Oct 2030, 0% from Oct 2031. NTA’s own reform page states it plainly: 「控除可能割合は令和8年10月から2年間70％、令和10年10月から2年間50％、令和12年10月から1年間30％となり、令和13年10月以降は0％となる。」

Worth flagging, because it nearly cost me this idea: most English-language sources are still publishing the repealed 80→50 schedule. Stripe’s transitional-measures page and several Big-4-adjacent explainers describe a jump to 50% on 1 Oct 2026 that the reform cancelled. I built the first version of this thesis on that number and had to tear it up at the primary source. Foreign-owned SMEs in Japan are reading the same stale English pages — which is itself a demand signal, and a content wedge.

30 September 2026 — the 2割特例 dies. That was the sweetener that got hundreds of thousands of formerly tax-exempt micro-businesses to register in the first place: pay 20% of your sales tax, done. It ends with the tax period containing 30 Sept 2026. The replacement 3割特例 covers individual proprietors only for 2027–2028 — corporations get nothing. So a large cohort of suppliers who registered purely for the relief now has a live reason to walk away.

And deregistration is silent. This is the part that makes it a product. NTA’s own deregistration guidance sets the mechanics: file the 取消届出書 by the 15th day before the start of the next tax period, and 「登録の効力は、提出日の属する課税期間の翌課税期間の初日に失われます。」 Miss that window and it slips to the period after. Nowhere in that form is there a duty to tell your customers. The buyer discovers it at audit.

Practitioners are already writing the warning by hand. One tax note’s checklist puts it exactly: 「一度確認した登録番号も、取引先が廃業や取消申請をすれば失効する」 — and then names the failure mode I’d have missed: 「長期取引の相手先は見落としやすいため要注意」. The suppliers you’ve billed with for eight years are the ones nobody re-checks.

A vendor blog aimed at accounting teams titles the whole problem 「取得して終わりじゃない！？」 — “collecting the number isn’t the end” — and describes current practice as 手作業・目視で確認: manual, eyeball verification. Microsoft ships a free Excel template for tracking supplier registration status. That is the state of the art for the segment.

Provenance:

• Signal 1 (demand): Practitioner checklists warning that confirmed registration numbers lapse silently and long-term suppliers are the ones most often missed — 「長期取引の相手先は見落としやすいため要注意」 — https://note.com/fmtax/n/n7523e7433a6e — observed 2026-09-03
• Signal 2 (feasibility): NTA’s 適格請求書発行事業者公表システム Web-API, live since Oct 2021, returns per-number lookups AND period-based difference (差分) data in JSON — https://www.invoice-kohyo.nta.go.jp/web-api/index.html — observed 2026-09-03
• Signal 3 (economic): FY Reiwa 8 reform sets 70% deduction from 1 Oct 2026 while the 2割特例 ends 30 Sept 2026 with no corporate successor — raising buyer cost and supplier deregistration incentive on the same date — https://www.nta.go.jp/taxes/shiraberu/zeimokubetsu/shohi/keigenzeiritsu/invoice-review/index.htm — observed 2026-09-03 Category: Regulatory arbitrage

3. The opportunity

The gap isn’t “look up a registration number.” That’s solved and free — NTA runs a public search site and a public API, freee and MoneyForward both check numbers at the point you save a supplier record, and a consultancy will sell you a one-off 名寄せ batch match.

The gap is the watch between checks, and the ledger entry that survives the audit.

Every incumbent treats registration as a property you validate on entry. It’s actually a status with an expiry you’re never notified about. Three things break in that gap:

1. The lapse arrives with no message. The supplier files a form with their tax office. Their invoice template may still carry the old T-number for months. Nothing in your accounting software changes.
2. The lapse is retroactive to a period boundary, not to the day you notice. Effect lands on the first day of the next tax period. By the time you spot it, you may have booked two quarters of full deductions you weren’t entitled to.
3. Catching it is only half the work. Claiming the 70% transitional relief requires the ledger to say so. Practitioners are explicit that 「帳簿に『経過措置適用』『80%控除』の旨を記載することが要件」 — miss the annotation and 「経過措置の適用が否認され、追徴課税・加算税のリスクがある」. You lose the partial credit and eat penalty tax, having done nothing wrong except fail to write a note.

There is one genuine competitor doing continuous monitoring — TransAcc’s easy Invoice Check, which does alert on status changes. I take it seriously. But it is quote-only (「取引先数に応じてお見積りいたします」, no published price), and its reference customer manages roughly 100,000 suppliers. Its documented limitation is that sole proprietors can’t be verified by batch file at all — only via direct survey response. Which is precisely the population that deregisters when the 2割特例 ends.

So: the tooling aims at 10万社 buyers. The exposure lands on the 30-to-500-supplier buyer who has no procurement team, checks in Excel, and has never re-checked a vendor they onboarded in 2023.

4. Target market

• Primary customer: The 経理担当者 — usually one person, sometimes the owner — at a Japanese company with 20–300 employees and 30–500 active suppliers. Construction subcontracting, design/creative agencies, IT services, logistics, light manufacturing. Anywhere the supplier base includes a meaningful tail of 一人親方 and freelancers. Also: foreign-owned subsidiaries in Japan, who read the stale English guidance and are budgeting against the wrong number.
• Why they buy: Because the ledger is theirs to defend. In an audit the examiner checks whether the invoice came from a registered issuer, and disallows the credit if not — the supplier’s paperwork failure becomes the buyer’s assessment. One consultancy states the audit consequence directly: 「税務調査において登録事業者だと思っていた取引先等が未登録事業者だったことがわかった場合、消費税の修正申告や差額の追加納税を求められる可能性がある」.
• Rough TAM reasoning: Japan has roughly 3.4M SMEs. The addressable slice is buyers with enough suppliers that Excel breaks but too few to warrant an enterprise contract — call it the 30–500 supplier band. Even a conservative read puts that in the low hundreds of thousands of companies. I need 400 of them.
• Why now for them: The cost of a missed lapse is rising on a published staircase, and the supplier-side incentive to quit registration peaks in the exact period the buyer’s cost goes up. Before Oct 2026 an unnoticed lapse cost 20% of the tax. After, 30%. In 2031, all of it.

5. Product sketch (MVP)

• Upload the supplier master (CSV, or the export from freee / MoneyForward / 弥生) — we match and attach a registration status to every row
• Daily poll of NTA’s difference feed; any supplier of yours that lapses, gets revoked, or newly registers triggers an alert before the next payment run
• A dated evidence record per supplier — status on each check, with the NTA response retained, so the audit answer is “here is what the register said on the day we paid”
• Boundary calculator for the Sept/Oct 2026 split: services spanning the change need 9月分 and 10月分 apportioned at different rates, which is the error practitioners expect most
• Auto-generated ledger annotation text (「経過措置適用 70%控除」) per affected transaction, so the relief isn’t denied on a missing note
• Exposure forecast: given your current unregistered suppliers and spend, what the 70/50/30/0 staircase costs you each year — the number to take into a price negotiation
• 屋号-only supplier resolver: for sole proprietors trading under a business name, guided workflow to obtain and confirm the number, since the public register won’t link 屋号 to 氏名 unless the proprietor opted in
• Japanese-first UI, with an English mode for foreign-owned subsidiaries

6. AI angle — what’s load-bearing

Take the AI out and this degrades into a cron job plus a spreadsheet — so I want to be honest about where it’s genuinely doing work, and where it isn’t.

Not load-bearing: the NTA lookup. That’s a REST call. Anyone claiming AI there is decorating.

Load-bearing: entity resolution. The supplier master says 「(株)山田製作所」, the register says 「株式会社山田製作所」, the invoice PDF says 「ヤマダ製作所」, and the sole proprietor bills as 「アトリエ木漏れ日」 while registering under a personal name the register won’t link publicly. Japanese business-name matching — 新字体/旧字体, katakana/kanji variants, 全角/半角, 前株/後株 — is the actual hard problem, and it’s why the enterprise vendors sell 名寄せ as a distinct product line. An LLM with a deterministic normalisation pass in front of it handles the messy tail that fuzzy string matching alone won’t.

Second load-bearing use: reading the invoice PDFs and receipts as they arrive, pulling the T-number and the transaction date, and checking that number against the register’s status on that date — not today’s status. That’s document extraction across wildly inconsistent Japanese invoice layouts, which is exactly what current vision models became good enough for.

7. Localization angle

This is a Japan-only product by construction — the regulation is the market. Everything about it is local: the NTA API, the register’s disclosure rules for sole proprietors, the tax-period-boundary effective dates, and the 帳簿 annotation requirement.

Two localisation notes that matter commercially. First, pricing: ¥9,800/mo reads as a rounding error to a Japanese SME back office and is well inside a 経理 manager’s discretionary spend, where an enterprise quote requires 稟議 and three signatures. Second, distribution runs through 税理士 (licensed tax accountants). Japanese SMEs don’t buy compliance tooling from a website; their 顧問税理士 tells them what to use. That’s a channel, and it’s covered in section 9.

The English mode is not an afterthought — foreign-owned subsidiaries are systematically misinformed right now because the English-language sources still describe the cancelled 80→50 schedule.

8. Business model — path to $1M–$5M ARR

• Pricing: ¥9,800/mo up to 100 suppliers; ¥24,800/mo up to 500; ¥49,800/mo up to 2,000. Annual prepay standard in Japan — offer 2 months free, which also crushes churn.
• ACV: ¥240,000 ($1,600) blended, assuming most land on the middle tier.
• Rough math to $1M ARR: ~950 customers at the ¥9,800 tier, or ~400 at ¥24,800. Realistically a mix: ~600 customers gets there.
• Rough math to $5M ARR: ~2,500–3,000 customers, which means winning through 税理士 firms rather than one at a time — a single mid-size firm with 200 client companies is a channel, not a customer. Plus the 2028 and 2030 steps down the staircase, each of which re-opens the conversation with everyone who ignored it in 2026.
• Expansion path: supplier-count tiers grow naturally. Then: the annotation engine becomes a filing-prep add-on; the 屋号 resolver becomes an outreach service (we contact your unregistered suppliers on your behalf); and the exposure forecast becomes a negotiation pack for the price discussions the reform forces — carefully, since one-sided price cuts on tax-exempt suppliers raise 独占禁止法/下請法 exposure, which is a real constraint and a real reason buyers want documented reasoning.

9. Go-to-market wedge — first 100 customers

1. Free exposure scan, timed to the deadline. Upload your supplier list, get back: how many are unregistered, how many lapsed since you last checked, and what the 70% step costs you annually. It’s cheap for me to run and it produces a specific yen number. Every buyer facing 1 Oct has to produce that number anyway. The scan finds the lapses they didn’t know about — that’s the demo.
2. Correct the English-language record. Publish a properly sourced English page on the actual FY Reiwa 8 staircase, citing NTA, against the stale 80→50 pages currently ranking. Foreign-owned subsidiaries in Japan are a searchable, high-ACV, underserved slice, and right now they’re being told the wrong number by every English source I checked.
3. 税理士 firms as the channel. Japan has roughly 80,000 registered tax accountants, many running practices with 50–300 SME clients, and they are contractually on the hook for advising on exactly this. Their pain is that they must tell every client to re-check suppliers and have no tool to do it at scale. Offer a firm dashboard across all client companies, revenue-share or flat per-client. Land 5 firms, get a few hundred companies. This is the path to $5M and it’s why I tag the idea sales-heavy.
4. The lapse list as cold outreach. NTA’s difference feed publishes revocations. That’s a public, dated list of companies whose customers are now silently exposed. I can’t see who buys from whom — but I can publish a monthly “N registrations lapsed this month, here’s how to check if any are yours” and let the anxiety do the routing. Post it where SME 経理 and 税理士 actually read: note, X, and 税理士ドットコム.
5. Bookkeeping and 記帳代行 shops. They process the invoices already, they carry the blame when a deduction is disallowed, and they serve dozens of SMEs each. Same multi-tenant dashboard as the 税理士 play.

10. Build complexity — justification

Low. The regulated data source is a free, documented, JSON-returning government REST API with a difference endpoint — no scraping, no auth theatre, no reverse engineering. The staircase math is arithmetic. Alerting is a scheduler.

The two pieces of genuine work are Japanese entity resolution (the 名寄せ problem, which is why incumbents charge for it) and invoice document extraction. Both are off-the-shelf model calls plus a normalisation layer and a correction UI for the tail. A competent solo builder ships the monitor and exposure scan in 5–6 weeks; the extraction and annotation engine adds 4–6 more. Native or fluent Japanese is non-negotiable — this is not a market you enter through machine translation.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Consumes a public NTA API under its published terms; no regulated advice given
Ethical — no harm / dark patterns	✅	One caution: exposure data could fuel coercive price cuts on tax-exempt suppliers. 独占禁止法/下請法 warnings belong in the product, not the fine print
Market exists (evidence above)	✅	Free Microsoft template, manual Excel practice, paid 名寄せ services, one enterprise monitoring vendor
1–5 person team can build this	✅	Solo builder with Japanese fluency; ~10 weeks to full v1
Launchable with <$50K / ₹40L	✅	API is free; cost is one person’s time plus model inference

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	16/20	Real money, audit-backed, and silent — but it’s a per-transaction leak (30% of the tax on affected spend), not a business-ending event. Felt at every payment run; rarely screamed about
Demand evidence	15	12/15	Strong indirect evidence: manual practice documented, free templates, paid batch services, one funded enterprise vendor doing exactly this monitoring. Docked because I found no SME saying “I would pay for this” — the complaints are written by accountants, not buyers
Build feasibility	15	13/15	Free documented government API with a delta feed; the hard part is Japanese name matching, which is solved-ish, not novel
Distribution clarity	15	11/15	The 税理士 channel is real and specific but slow to land; the free scan and the English-language correction are fast and cheap. No named scrapeable list of buyers with exposure — that’s the weak link
Revenue mechanics	15	11/15	Pricing sits comfortably under Japanese SME discretionary spend and far under an enterprise quote. Risk: the value is loss-avoidance, which prices lower than revenue generation, and ¥9,800 needs volume
Time to first revenue	10	8/10	A hard deadline 4 weeks out is the best closing mechanism there is. Free scan → paid conversion should be weeks, not quarters
Defensibility	10	5/10	The API is public and freee could ship lapse alerts in a quarter. Real moats: the accumulated dated evidence record (switching means losing your audit trail), the 名寄せ mapping table you build per customer, and the 税理士 relationships. Execution-and-speed moat, not a structural one
Total	100	76/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · domain-expertise-required — and functionally sales-heavy if you chase the 税理士 channel, which you must to get past $1M. Native Japanese is a hard requirement.

Key assumptions to validate (3–5)

1. Assumption: Lapses among SME supplier bases are frequent enough to be alarming — say >1% of a typical 200-supplier base per year, rising after Oct 2026. How to test: pull NTA’s difference feed for the last 12 months, count revocations, and cross-check against a few real supplier lists obtained from friendly SMEs. This is measurable before writing product code, and it is the assumption the whole idea rests on.
2. Assumption: SME 経理 will pay ¥9,800/mo rather than accept the free Excel template. How to test: run the free scan for 30 companies; measure how many convert when shown a real lapse in their own supplier list versus how many say “we’ll just check manually.”
3. Assumption: 税理士 firms will resell or recommend rather than build a spreadsheet macro. How to test: 15 structured conversations with firms of 50+ clients; ask what they’re currently telling clients to do about Oct 2026.
4. Assumption: TransAcc and freee stay aimed up-market / at point-of-entry checks through 2027. How to test: track their release notes and pricing pages monthly.

Risk flags

1. Platform dependency: the entire product consumes one government API. If NTA changes terms, rate-limits, or ships its own alerting, the core is gone. Mitigate by owning the evidence record and the annotation workflow, not just the lookup.
2. Incumbent absorption: freee and MoneyForward already hold the supplier master. Lapse alerting is a feature they could add, and the Oct 2026 deadline gives them the same prompt it gives me. This is the most likely way the idea dies.
3. Market timing, cutting both ways: the deadline is 4 weeks out, which is superb for urgency and bad for build time. Arriving in Nov 2026 means selling into remorse rather than fear — workable, since the next step is 2028, but a weaker sale.
4. Ethical/legal edge: the exposure forecast is one short step from a coercive price-cut tool. 下請法 and 独占禁止法 constrain one-sided reductions on tax-exempt suppliers. Build the guardrails in.
5. Sourcing risk I hit personally: the English-language record on this reform is currently wrong. Any competitor — or any customer — working from those sources will dispute your numbers. Cite NTA in-product.

14. Structured verdict

Score:                  76/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Japanese-fluent solo technical founder with a 税理士 or SME-finance
                        advisor; comfortable selling through professional-services channels
Time to revenue:        6–10 weeks (free scan → paid, against a 1 Oct deadline)
Capital to launch:      ¥1.5–2.5M (~$10–17K) — one builder's time plus inference costs
Top 3 assumptions to validate first:
  1. Lapse frequency is materially >1%/yr per supplier base — measure directly from
     NTA's difference feed before building anything
  2. SME 経理 converts from free scan to ¥9,800/mo when shown a real lapse in their own
     data — 30 free scans, measure conversion
  3. 税理士 firms with 50+ clients will recommend rather than DIY — 15 interviews
Kill criteria:
  - Abandon if NTA's difference feed shows <0.5% annual lapse rate across a
    representative SME supplier base — the fear isn't real and the alerts are noise
  - Abandon if freee or MoneyForward ships lapse monitoring in their standard SME tier
    before your v1 — they own the supplier master and you don't
  - Abandon if <3 of 30 free-scan companies convert to paid within 60 days

15. Next step — 1-week validation sprint

• Day 1–2: Pull 12 months of NTA difference-feed data. Count revocations and lapses; segment by corporate vs sole proprietor. This is the falsifiable core — if the lapse rate is trivial, there is no product, and I’ll know before writing a line of application code.
• Day 3: Get real supplier master lists from 5 SMEs (via one friendly 税理士). Match against the register. Count how many suppliers are currently unregistered or lapsed and the company doesn’t know.
• Day 4–5: Take those 5 results back to the companies with a yen figure for their Oct 2026 exposure. Ask one question: would you pay ¥9,800/mo to be told the moment this changes? Then call 10 税理士 firms and ask what they’re currently advising for 1 October.
• Decide on: ≥1% annual lapse rate in the feed AND ≥2 of 5 companies discovering at least one supplier they didn’t know had lapsed AND ≥4 of 10 税理士 saying they have no tool for this. Anything less and the fear is theoretical — shelve it and revisit before the Oct 2028 step to 50%.