FuyoCheck

1. One-liner

Judges whether a foreign worker’s overseas-dependant paperwork will actually survive a Japanese withholding audit.

2. Trend signal — why now?

Three things moved at once, and they move in the same direction.

The population exploded. Foreign workers in Japan hit 2.6 million as of October 2025, up 11.7% year on year — a record for the thirteenth consecutive year. Foreign residents overall crossed 4,125,395 at end-2025, up 9.5%. The government is targeting 1.23 million intake under labour-migration programmes. Roughly 500,000 technical trainees and a fast-growing 特定技能 cohort sit in exactly the demographic that sends money home to parents and children abroad.

The rules got strict, and stayed strict. Since January 2023, a 30–69-year-old overseas relative only qualifies as a dependant if they are a student abroad, disabled, or received ¥380,000+ in the year. Every claim requires 親族関係書類 (proof of relationship) and 送金関係書類 (proof of remittance), including Japanese translations where the originals are foreign-language. The NTA reissued its guidance pamphlet as recently as June 2025 (令和7年6月改訂) — this is live, actively-maintained enforcement territory, not a dormant rule.

The people who used to absorb this are gone. 71.6% of registered tax accountants are 50 or older; only 10.9% are in their 20s–30s. The effective jobs-to-applicants ratio in the profession is ~2.3x. Firms are carrying more clients per head, and the first thing that gets rushed is the fiddly annual work — precisely the 年末調整 evidence review.

Meanwhile the obvious adjacent fix is already commoditized: SmartHR gives away 8-language + やさしい日本語 年末調整 to all users for free (Jan 2025), and OBC’s 奉行クラウドEdge added English in Nov 2024. Translation is not the gap. Nobody adjudicates the evidence.

Provenance:
  - Signal 1 (demand): Foreign workers in Japan reach record 2.57–2.6M, up 11.7% YoY; foreign population passes 4M — https://www.japantimes.co.jp/news/2026/01/30/japan/foreign-workers-record-high/ and https://www.nippon.com/en/japan-data/h02750/ — Jan 2026
  - Signal 2 (feasibility/regulatory): NTA overseas-dependant rules require 親族関係書類 + 送金関係書類 with Japanese translation; 30–69 bracket needs ¥380k remittance proof; pamphlet revised June 2025 — https://www.nta.go.jp/publication/pamph/pdf/0022009-107_02.pdf — Jun 2025
  - Signal 3 (economic): Tax-accountant capacity crunch — 71.6% aged 50+, 10.9% under 40, ~2.3x jobs-to-applicants ratio — https://note.com/zeirishi_ai_lab/n/ne8ba58dd03b2 — 2026
  Category: Regulatory arbitrage

3. The opportunity

Every 年末調整 system on the market — SmartHR, 奉行Edge, jinjer, MoneyForward — solves collection. They render the form in Vietnamese, chase the employee, and store the upload. Then they stop. The uploaded artefact is treated as a checkbox: a file exists, therefore the field is complete.

But the tax office does not audit whether a file exists. It audits whether the file proves the claim. And that is a judgment call across a messy set of inputs:

• Is this Vietnamese household-registration extract actually a 親族関係書類, or is it a document that names the worker but never establishes the relationship to this specific dependant?
• Do these twelve Wise transfers total ¥380,000 for this one named relative — or were they lump-summed to a brother who redistributed them? The NTA is explicit: a lump sum to one representative is remittance proof for that representative only, not the other relatives.
• Is the dependant 29 or 31? Because at 30 the entire evidentiary bar changes.
• Is there a Japanese translation attached, or just the original?

That judgment currently happens in a 総務 person’s head in late November, at speed, for a document they cannot read, in a year when their tax accountant has 20% more clients than last year. The failure is silent — nothing breaks at the time. It surfaces two years later as a 源泉所得税 assessment with 不納付加算税 (10% base rate) plus 延滞税, landing on the employer as withholding agent, not the employee.

That’s the gap: a judgment layer, not a collection layer. It’s exactly what an LLM with vision is now good at and what a payroll form-renderer structurally is not.

4. Target market

• Primary customer: HR/総務 manager or outsourced 社労士/税理士 office serving Japanese SMEs with 20–300 employees of whom 5–80 are foreign nationals — manufacturing, food processing, construction, nursing care, logistics, agriculture, hotels. Also 登録支援機関 and 監理団体, who support 特定技能 workers and 技能実習生 as their core business and already own the client relationship.
• Why they buy: The pain is a specific, dated, recurring dread. November–December, a stack of foreign-language documents, a hard deadline, no ability to read them, and personal accountability if the tax office disagrees later. One 総務 manager’s job is to not be the reason the company gets assessed.
• Rough TAM reasoning: 2.6M foreign workers. Even assuming only a minority claim overseas dependants, the employers are concentrated in industries that hire foreign labour in clusters. There are thousands of 登録支援機関 registered with the Immigration Services Agency, each with a book of client firms — a natural aggregation point.
• Why now for them: The cohort is growing double-digits annually, the ¥380k/age-30 rules are strict and unfamiliar, the tax accountants who used to catch this are stretched thin, and free multilingual 年末調整 has trained everyone to collect documents without ever teaching them to evaluate documents.

5. Product sketch (MVP)

• Drop the stack, get a verdict. Upload whatever the employee handed over — photos, PDFs, screenshots of a remittance app — and get a per-dependant ruling: Sufficient / Insufficient / Needs human review, each with a plain-Japanese reason.
• Relationship-proof reader. Identifies whether a foreign document (birth certificate, household register, family book) actually establishes the claimed relationship, and flags when it names the worker but not the dependant.
• Remittance math per named person. Parses transfer records, attributes each transfer to a specific dependant, sums the year, and shows the running total against the ¥380,000 line — with an explicit warning when transfers appear lump-summed to one representative.
• Age-bracket gate. Auto-applies the 30–69 rule and tells you which of the three escape hatches (student / disabled / ¥380k) the claim is relying on.
• Translation attachment. Generates the Japanese translation required for foreign-language originals and attaches it to the evidence bundle.
• The “ask the employee” message. When something is missing, drafts the request in the worker’s language and やさしい日本語 — specific about which document, not “please resubmit”.
• Audit bundle export. One organised, indexed file per employee per year, with the reasoning recorded — the thing you hand the auditor two years later.

6. AI angle — what’s load-bearing

Remove the AI and this product cannot exist. The core operation is: read an arbitrary document, in an arbitrary language, in an arbitrary layout, photographed badly, and decide whether it satisfies a specific legal evidentiary test. That is not OCR-plus-rules — a Vietnamese household register, an Indonesian family card, a Filipino PSA birth certificate and a Nepali migration certificate share no template. Vision-capable models reading semi-structured foreign documents and reasoning against a codified rule set is the entire product. The rest is a file uploader.

The economics also only just landed. This is a seasonal, bursty workload — thousands of documents in a six-week window — and at current inference prices a full evidence review costs cents per employee against a service that bills in thousands of yen.

7. Localization angle

This is the localization play, and it runs in both directions. The product is Japan-specific on the rule side (NTA evidentiary standards, 年末調整 calendar, 源泉徴収義務者 liability) and multi-country on the document side (Vietnam, Indonesia, Philippines, Nepal, Myanmar, China — the actual source countries of Japan’s foreign workforce). That dual-sidedness is the defensible bit: a generic global HR tool has neither end.

Pricing must be yen-native and per-employee, matching how Japanese payroll outsourcing is already priced (year-end adjustment is billed as an option at ¥1,000–2,000 per person, on top of ¥400–1,500/person/month for payroll). The market has an existing, well-understood per-head mental model for exactly this task — sit inside it.

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

• Pricing: ¥1,500 per foreign employee reviewed per year, sold as a seasonal package, plus a ¥30,000–80,000/yr base for the workspace. For 登録支援機関/税理士 offices, a partner tier at ¥800–1,000/head with volume bands.
• ACV: A 30-foreign-employee manufacturer ≈ ¥95,000/yr. A 登録支援機関 with 400 workers across client firms ≈ ¥400,000/yr. Partner accounts carry this business.
• Rough math to $1M ARR (~¥150M): ~200 partner accounts averaging ¥600,000, or a blend of ~150 partners plus ~400 direct SMEs. Reachable, because the partners aggregate.
• Rough math to $5M ARR: Requires expanding beyond the dependant question into the adjacent foreign-employment compliance calendar — residence-status expiry tracking, 在留カード validity, the mandatory periodic reporting that 登録支援機関 already owe the Immigration Services Agency. Same buyer, same season-agnostic relationship, much larger surface.
• Expansion path: Headcount growth is automatic in this market (11.7%/yr). Then per-head → per-workflow → year-round retainer.

Honest weakness: it’s seasonal. Revenue concentrates Oct–Jan. The expansion path above isn’t optional garnish — it’s how this becomes a real business rather than a Q4 spike.

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

• The registry is public. The Immigration Services Agency publishes the list of registered 登録支援機関 (moj.go.jp/isa). That is a downloadable, addressed, phone-numbered list of organisations whose entire job is handling paperwork for exactly these workers. Work it directly — this is the single highest-leverage list in the plan, because each one carries dozens of employer clients.
• Sell the horror story, in September. Build a one-page worked example: a real ¥380k claim that fails because transfers were lump-summed to a brother, with the resulting assessment math. Mail it to 監理団体 and 社労士 offices in the eight prefectures with the densest 技能実習生 populations, timed for early autumn when 年末調整 planning starts.
• Free “one employee” audit as the hook. Let a prospect upload a single real case and get the verdict free. The moment a 総務 manager sees a document they’d have accepted come back marked Insufficient — no relationship established to named dependant, the sale is made. This converts because it demonstrates a gap they didn’t know they had.
• Ride the 社労士/税理士 channel. These offices are capacity-constrained (see §2) and actively hunting for tools that let existing staff carry more clients. Revenue-share or white-label; they bring the book.
• Industry associations. JAC (construction) and the sector-specific 特定技能 bodies run member communications into precisely this employer base.

10. Build complexity — justification

Low. Off-the-shelf vision-capable LLM APIs do the document reading and adjudication; the rule set (age brackets, ¥380k threshold, document categories, translation requirement) is small, published, and stable enough to encode by hand. No payroll integration needed for v1 — this sits beside SmartHR rather than inside it, consuming uploads and emitting a bundle. Realistic v1 for one or two people: 8–10 weeks. The hard work is not engineering, it’s assembling a correct, defensible rule interpretation and a golden test set of real documents from the top 5–6 source countries.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Decision-support for the employer. Ship as evidence-sufficiency review, not tax advice — the 税理士法 line matters, see risks.
Ethical — no harm / dark patterns	✅	Net-positive both ways: workers get legitimate deductions they’d otherwise forfeit for want of a document; employers avoid assessments.
Market exists (evidence above)	✅	2.6M workers, strict rules, priced adjacent services (¥1,000–2,000/head for 年末調整).
1–5 person team can build this	✅	8–10 weeks, no custom models.
Launchable with <$50K / ₹40L	✅	Inference + a landing page + one Japanese-fluent domain advisor.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	16/20	Real money, personal accountability, hard deadline — but annual, not daily. A once-a-year terror scores lower than a weekly one.
Demand evidence	15	11/15	Population, rules and pricing benchmarks are all documented. What’s not documented: direct verbatim complaints about this specific evidence-review pain. Inferred from adjacent evidence, not observed.
Build feasibility	15	13/15	Off-the-shelf vision LLM + small rule set. Golden test set is the only real work.
Distribution clarity	15	11/15	The 登録支援機関 registry is a genuinely excellent named list. Conversion rate through it is untested.
Revenue mechanics	15	11/15	Per-head pricing benchmarked against real payroll-outsourcing rates. Seasonality is a genuine drag on the multiple.
Time to first revenue	10	8/10	Strong seasonal forcing function — but you must hit the Sept–Nov window or wait a full year. Timing risk is real.
Defensibility	10	6/10	Rule knowledge + accumulated document-pattern library across source countries is a soft moat. SmartHR could add this; whether they’d bother for a foreign-worker edge case is the bet.
Total	100	76/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · domain-expertise-required — needs business-level Japanese and a 税理士 or 社労士 advisor on the rule interpretation. A non-Japanese-speaking founder cannot sell this.

Key assumptions to validate

1. Assumption: A meaningful share of overseas-dependant claims currently submitted would fail an audit. How to test: Get 3 sympathetic 社労士 offices to run 20 real anonymised prior-year claims through a manual version. If the defect rate is under 10%, the fear isn’t justified and the pitch collapses.
2. Assumption: Employers feel this as their risk, not the employee’s. How to test: 20 interviews with 総務 managers at foreign-hiring SMEs. Ask who eats the assessment. If they say “the employee”, the urgency evaporates.
3. Assumption: 登録支援機関 will resell rather than treat it as a threat to billable support hours. How to test: 15 conversations from the public registry, with pricing on the table.
4. Assumption: ¥1,500/head clears procurement without a committee. How to test: Put the price in front of 10 targets and watch for the “I’d need approval” reflex.

Risk flags

1. Platform dependency / incumbent squeeze: SmartHR already owns the 年末調整 workflow and gives multilingual away free. If they ship evidence adjudication, the wedge narrows fast. The counter-bet is that foreign-dependant edge cases are unattractive to a horizontal player — but it is a bet.
2. Regulatory boundary (税理士法): Japan restricts who may give tax advice. The product must present as evidence-sufficiency review and document preparation, with a 税理士 in the loop for anything resembling a determination. Get this wrong and the business is illegal rather than merely unprofitable.
3. Seasonality: Revenue clusters into one quarter. Cash management and the year-round expansion path are load-bearing, not optional.
4. Accuracy liability: A Sufficient verdict that later fails an audit is a trust-destroying event and a possible claims exposure. The Needs human review band must be used generously, and the tool must be positioned as a check on humans rather than a replacement for them.
5. Demand evidence is inferential: Every structural fact here is sourced. The specific pain is reasoned from those facts, not from customers saying it out loud. That is the single biggest reason this is a 76 and not an 85 — and the first thing to fix.

14. Structured verdict

Score:                  76/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Japan-based or Japanese-fluent technical founder with a 社労士/税理士 advisor
Time to revenue:        8–14 weeks if launched into the Sept–Nov window; otherwise a year's wait
Capital to launch:      ¥1.5–3M ($10–20K)
Top 3 assumptions to validate first:
  1. Real defect rate in prior-year claims — 20 anonymised cases via 3 friendly 社労士 offices
  2. Employer (not employee) owns the perceived risk — 20 総務 manager interviews
  3. 登録支援機関 will resell — 15 calls from the public ISA registry with pricing on the table
Kill criteria:
  - Abandon if the defect rate in real prior-year claims is <10% — the fear isn't real
  - Abandon if <3 of 15 登録支援機関 will take a paid pilot
  - Abandon if SmartHR ships overseas-dependant evidence checking before v1 lands

15. Next step — 1-week validation sprint

• Day 1–2: Pull the 登録支援機関 registry from the ISA site. Build the list. In parallel, codify the NTA rule set from the June 2025 pamphlet into a one-page decision tree and have a 税理士 red-line it.
• Day 3–4: Hand-run 20 real anonymised prior-year dependant claims — begged from 2–3 friendly 社労士 offices — through the decision tree manually. This is the core experiment. Count how many would fail.
• Day 5: Call 15 organisations from the registry. Lead with the measured defect rate from Day 3–4, not with the product.

Falsifiable outcome: if the measured defect rate across 20 real claims is ≥10% and ≥3 of 15 registry calls agree to a paid pilot at ¥1,500/head, build it. If the defect rate is below 10%, kill it regardless of how enthusiastic the calls are — enthusiasm without defects means they’re being polite about a problem they don’t have.