Tachiai

1. One-liner

Compares pickup and return walkarounds, proves which scratch is new, drafts the claim or the all-clear on the spot.

2. Trend signal — why now?

The Japanese word for the moment this product lives in is 立会い (tachiai): the attended inspection at the counter when the car comes back. Every discount rent-a-car store in Japan does it with a paper checksheet, a biro, and whichever staff member is free. The customer does it with a phone. When the two records disagree, the store either eats the repair plus the ノンオペレーションチャージ (NOC, the ¥20,000–¥50,000 downtime charge) or bills the customer and takes its chances with the consumer centre and X.

Three things changed in the last eighteen months.

The regulator named the problem. On 28 March 2025 the National Consumer Affairs Center (国民生活センター) published 「レンタカーでの傷トラブルにご用心 −記録、確認、連絡！−」. Its two lead cases are the whole product. Case one: a renter photographed the car before departure, confirmed with staff at return that there was no damage, and was contacted days later about a side-mirror scratch. Case two: a renter bought full coverage and was still charged ¥20,000 NOC for a small scratch. The centre’s earlier release (25 March 2021) puts the volume at 「ここ数年、相談件数は年間500件程度で推移しています」 and names the dominant type: 「返却時もしくは返却後に身に覚えのない傷の修理代金を請求された」. Consumer-centre consultations are the tip; most disputes never get that far.

A franchise got burned in public. In March 2026 a Niconico Rent-a-Car store billed ¥88,000 for two 1–2 mm stone chips on a windscreen. The friend of the renter posted it on 17 March; by 19 March 弁護士ドットコム was quoting the replies (「支払う必要無いでしょ」「これが通ったらレンタカー借りれません」) and the company’s apology (「この度は車両の傷の請求に関する件で、多くの方にご心配をおかけしておりますこと、深くお詫び申し上げます。」). On 21 March former prosecutor 前田恒彦 wrote it up on Yahoo News: the claim was withdrawn, and the company’s own line was 「お客様にご負担をお願いすることが適切であったかについては慎重な判断を要する事案」. The legal position, per lawyer 荒川香遥 in the same coverage, is that stone chips are still 「原則として『利用者の責めに帰すべき事由（落ち度）』とみなされ、修理費用等の負担を求められるケースがほとんど」. So the store was legally right and commercially destroyed. That is the exact shape of pain a paid tool fixes: the charge was defensible, the evidence was not presentable.

Reading two videos became almost free. On 1 September 2026 Google shipped agentic video understanding for Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. The model navigates the timeline and inspects only the segments it needs, at “standard Gemini API token pricing with no additional feature fee”, with “up to 66%” lower analysis cost and “up to 88%” fewer tokens than fixed-frame processing. Gemini 3.7 Flash launched 13 August 2026 at $0.75 per million input tokens (introductory to 31 December 2026). Two 90-second handheld walkarounds cost a few yen to compare. In 2025 this was a computer-vision project; now it is a prompt, a body-panel map and a checksheet template.

Two more signals make the timing sharper. Hertz’s UVeye scanner gates, rolled out across roughly 100 US airport locations by end-2025, were reported in Japan by president.jp (23 September 2025) as 「肉眼では確認が難しいほどの微細な『損傷』で、日本円で約3万円もの請求を突きつけられた」. Japanese renters now arrive expecting machine evidence and distrusting it; a store that shows a before-and-after clip wins the argument, a store that shows a biro sketch loses it. And the customer mix is shifting toward people who cannot read the checksheet: Nikkei reported in July 2026 that foreign renters were in 212 accidents in 2025, above the 2019 figure of 208 and 4.5 times the count five years earlier, prompting a Japan Tourism Agency survey. Local TV (ytv, 24 August 2026) has Jnet Rent-a-Car’s Yamanashi area manager explaining stop signs to overseas drivers word by word. The inspection at return now has to be understood by a customer who speaks no Japanese, at a store with no English.

Provenance:

• Signal 1 (demand): 国民生活センター alert 「レンタカーでの傷トラブルにご用心」, lead cases of post-return scratch claims and NOC despite full cover — https://www.kokusen.go.jp/news/data/n-20250328_1.html — 2025-03-28; volume 「年間500件程度」 from the 2021 release — https://www.kokusen.go.jp/pdf/n-20210325_1.pdf — 2021-03-25
• Signal 2 (demand / economic): Niconico Rent-a-Car ¥88,000 stone-chip claim withdrawn after SNS backlash — https://www.ben54.jp/news/3319 — 2026-03-19; https://news.yahoo.co.jp/expert/articles/4a7a75f17f07607f0d4f69e9bf847d81dbb7dfa2 — 2026-03-21
• Signal 3 (feasibility): Agentic video understanding for Gemini Flash models, no feature fee — https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-agentic-video-in-gemini/ — 2026-09-01
• Signal 4 (economic / geographic arbitrage): Hertz UVeye AI scanner gates at ~100 US airport sites, $195 charge for a barely visible dent — https://president.jp/articles/-/102619?page=1 — 2025-09-23
• Signal 5 (why-now): Foreign-renter accidents 212 in 2025, above pre-COVID, 4.5× in five years; JTA survey — https://www.nikkei.com/article/DGXZQOUA234NS0T20C26A7000000/ — 2026-07
• Signal 6 (customer voice): itmedia blog, discount rental store admits 「該当箇所を見落としてしまい、チェックシートへ記載されないままサインをいただき」 — https://blogs.itmedia.co.jp/agile/2026/09/rentacar_kizu_shoko_ai_evidence.html — 2026-09-03 Category: Tech-unlock (agentic video made a pickup-versus-return diff cost yen, not a vision team) + Geographic arbitrage (US majors bought scanner gates; Japan’s 10,000-plus licensed operators have a paper checksheet)

3. The opportunity

The incumbents split cleanly by price band and none of them do the job.

• UVeye, ProovStation, Ravin AI. Drive-through gantries or enterprise phone apps sold to Hertz, Sixt, Avis and insurers. Ravin holds a Japanese patent but sells to fleets and claims departments. None of them will quote a 12-car store in Naha.
• Japanese reservation systems. ITトレンド’s 2026 round-up lists nine (リザエン, れんサポ, グランドレンタカー, レンタカー侍 from ¥8,800/month, and so on) at 「月額料金：5,000円〜３万円」. The most advanced feature on the damage side is an e-signature on a digital 車両キズチェックシート. That is a PDF of the biro sketch. None compare pickup against return.
• ゼロダッシュ「カーチェックシート」. The one purpose-built tool: staff send four photos (front, back, left, right) which are “managed via email” with a 10 MB limit and a CSV export that does not include the images. No comparison, no verdict.
• Kaflix Cloud. The DX vendor for mid-size operators (JR West Rent-a-Car and Gulliver logos on its site). Its own blog describes the store floor: 「口頭・LINE・メモで情報を共有しているため、伝え忘れや勘違いが起きる」 and 「出発対応と返却対応が同時に発生します。そのため「ちょっと確認します」の一言が何度も積み重なります」. Its return app records 「走行距離・燃料残量・車体状態」 as fields, not as evidence.
• The adjacent 2026 launch is not a competitor. Aioi Nissay Dowa and Navitime announced (9 April 2026) an inbound-driver accident-reduction app for rental operators, due in FY2026: English voice navigation, safe-driving diagnosis, traffic-rule quiz. Nothing on damage.

So the gap is the decision, not the capture. Every party already records: the store has its checksheet, the customer has fifteen photos (the itmedia author had exactly that and won in minutes), and the guides all say 「出発前に車の周りを1周だけ撮る。できれば動画で。5分です」. What nobody produces is a document that says, for this rental, on this car, these three marks were present at 10:02 and this one was not, here are both frames, here is the tariff line and the NOC line separated as the consumer centre asks, signed by both sides before the customer walks. Tachiai produces that document, or the all-clear, in the language the customer reads.

The reason a focused team beats Kaflix or レンタカー侍 to it: their product is the reservation calendar and the store’s relationship with them is about bookings. The scratch fight is a five-minute event at the counter that their sales team never sees. The store owner does see it, every day of Golden Week and Obon.

4. Target market

• Primary customer: the owner or store manager of a discount or independent rent-a-car store in Japan with 5–60 cars. Three sub-types: (a) franchisees of Niconico (1,439 stores as of June 2026, per MIC’s PR Times release) and Gutsu (targeting 500 stores, roughly 500 franchisee companies), typically a petrol station or used-car dealer running rentals on the side; (b) resort-area independents in Okinawa, Hokkaido, Ishigaki and Miyako, where inbound renters dominate and NOC補償 is sold at the counter for ¥500–¥1,080 a day; (c) 業務レンタカー (weekly and monthly commercial rental) operators, whose customers are tradesmen driving hard.
• Why they buy: in their words. The store in the itmedia case, by email: 「貸渡時には傷の確認を行いましたが、該当箇所を見落としてしまい、チェックシートへ記載されないままサインをいただき」. The 業務レンタカー operator blog (田川英紀, updated 31 January 2026): 「レンタカーの傷には明確な許容範囲はありませんので」 and, on the customer side, 「補償が使えずに高額請求となる可能性があります」. The industry’s own stance, from a rental employee on Yahoo知恵袋 (old, 2009, but unchanged): 「貸渡契約期間内は、全て契約者様に管理責任があります」. They know they are entitled to charge. What they lack is proof that survives a phone photo and a tweet.
• Rough TAM reasoning: MLIT’s own trend chart shows roughly 10,500–12,400 licensed rental operators across the recent years it plots; the All Japan Rent-a-Car Association counts 1,168,522 rental vehicles as of end-March 2025. Strip the majors (Toyota Rent-a-Car alone is about 11% of the fleet) and airport chains, and there are on the order of 8,000 stores in the 5–60 car band. At ¥12,800–¥24,800 a month that is a ¥1.5–2.4 billion (about $10–16M) annual pool, plus a per-rental line for multilingual all-clears. Enough for a $2–5M business, not a venture case.
• Why now for them: the Niconico case taught every franchisee that a legally sound claim can cost more than the repair. Foreign renters are up, accidents are up, and the checksheet is in Japanese. Agentic video priced the fix at yen per rental this month. IT導入補助金 already lists rental operators as eligible, so year-one cost can be subsidised.

5. Product sketch (MVP)

• Two walkarounds, guided. Staff (or the customer, at unattended pickups) film a 60–90 second loop following on-screen prompts: four corners, each panel, wheels, windscreen, mirrors, the interior. The app refuses to finish until every zone has been covered and lit.
• Pickup record in 30 seconds. Existing damage is listed by panel with a still, a size estimate and a severity, laid onto the store’s own checksheet diagram. Customer signs on the phone. Both parties get the record on LINE or email with a timestamp and a hash.
• Return diff. The return video is compared against the pickup record. Output is one page: unchanged marks (greyed), new marks (paired before/after frames), and “cannot confirm” items with a prompt to re-shoot that panel while the customer is still there.
• Claim draft or all-clear. For new damage: a Japanese 請求書 draft with repair estimate line, NOC line and coverage applied, separated the way the consumer centre says they must be, plus a plain-language summary in the customer’s language (English, Chinese, Korean at launch). For no new damage: a signed 返却確認書 the customer can keep.
• Stone-chip and wear rules. Store-level policy toggles: minimum size to charge, windscreen chip policy, wheel-scuff policy. The draft cites the policy so staff stop improvising at the counter.
• Dispute pack. If the customer contests later, one button exports both videos, the diff page and the signed records as a folder for the insurer, the franchise HQ’s 事故トラブル対応 desk, or the consumer centre.
• Fleet memory. Every car accumulates its damage history across rentals, so the third renter of the week is not blamed for the first renter’s scuff that the second inspection missed.

6. AI angle — what’s load-bearing

Remove the model and the product is ゼロダッシュ: photos in an email. The load-bearing work is (1) matching two handheld, differently lit, differently framed videos of the same car panel by panel, (2) deciding which marks are new versus pre-existing versus dirt, reflection or rain, and (3) drafting a bilingual document that separates repair, NOC and coverage in the format the consumer centre and the franchise HQ accept. Until 1 September 2026, step 1 meant sampling frames at a fixed rate and paying for all of them, or training a damage detector on your own labelled set. Agentic video lets the model seek to the segments where a given panel is in frame and compare only those. Step 3 is standard LLM drafting with the store’s policy as context. The honest risk is step 2 in bad conditions (night returns, rain, dark paint); that is what the validation sprint measures.

7. Localization angle (if any)

Japan-first by design, not by translation.

• The document is the product. 貸渡証, 車両キズチェックシート, 返却確認書 and the NOC/免責 breakdown are Japanese artefacts with expected layouts. The output must drop onto the store’s existing forms, including the e-signed versions reservation systems already produce.
• LINE, not email. Customer records and all-clears go out on LINE; staff notifications too. Email is the fallback.
• Inbound languages. English, Traditional Chinese, Simplified Chinese and Korean summaries at launch, matching the inbound mix in Okinawa, Hokkaido and Kansai. The Japanese record remains the legal one; the translation is explanatory.
• Consumer-contract law framing. After the Niconico case, the draft claim must avoid clauses a court could read as 「消費者の利益を不当に害する条項」. The tool’s job is to make a defensible claim presentable, not to make an indefensible one look official. Store policy toggles default to charging only for damage that meets the store’s own published standard.
• Subsidy. Register as an IT導入支援事業者 so the store can claim IT導入補助金 on the subscription. Kaflix already markets this route to rental operators.
• Pricing in yen, monthly, no annual lock. Discount stores are seasonal; they will not sign a 12-month contract in November.

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

• Pricing: ¥12,800/month per store up to 30 cars (includes 300 returns); ¥24,800/month for 31–100 cars (1,000 returns); ¥40 per return beyond the allowance. Multilingual all-clear included. Franchise HQ tier priced per store with a shared dispute-pack inbox.
• ACV: about ¥200,000 (roughly $1,350) per store, blended; ¥300,000-plus for resort stores that run over allowance in season.
• Rough math to $1M ARR: 750 stores × ¥200,000 = ¥150M ≈ $1.0M. That is under 10% of the 5–60 car band, or one franchise HQ deal plus Okinawa and Hokkaido independents.
• Rough math to $5M ARR: 2,500 stores on subscription (¥500M) plus two of: a franchise-wide contract with Niconico or Gutsu at HQ level; a white-label to one reservation system as its キズチェック module; a per-claim fee from NOC補償 underwriters who want the evidence pack; car-sharing operators (Times, Orix) for member-shot returns. Any two get to ¥750M ≈ $5M.
• Expansion path: per-return overage in season; add-on for 出張引取 (delivery to the customer’s door, where pickup is unattended); dispute-pack for franchise HQs; damage-history export at vehicle disposal (a used-car dealer running rentals wants the log when the car goes to auction).

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

• Okinawa first, in person. Okinawa has the densest cluster of independent and franchise rental stores in Japan and the highest inbound share. Take the 沖縄県レンタカー協会 member list and the row of rental lots around Naha Airport, visit 100 stores in two weeks with a tablet, and run the diff on the store’s own car in the lot. Ask for ¥12,800 a month on the spot with the subsidy form pre-filled. Target 15 signed.
• MLIT’s own operator list as the cold list. The 2025 LINKS Veda release of 貸渡実績報告書 data carries operator name, transport-bureau jurisdiction, office count and fleet by category across 8,310 Form-1 records. Filter to 5–60 cars, drop the majors, and you have roughly 5,000 named targets with a fleet size. Send a one-page fax and a LINE-friendly QR code (fax still works in this trade) with the Niconico ¥88,000 story on one side and a before/after diff page on the other. Expect 2% replies, 30 demos, 10 stores.
• Franchise HQ pilots. MIC (Niconico) advertises 「事故トラブル対応」 as HQ support to franchisees; that desk fielded the March 2026 backlash. Pitch a 20-store pilot in Okinawa and Hokkaido on the argument that HQ’s next viral case is a franchisee with a paper checksheet. Same pitch to Gutsu. One HQ pilot is 20 stores in a month.
• Reservation-system partners. レンタカー侍, れんサポ and グランドレンタカー each sell into the same band and have no damage feature. Offer Tachiai as their キズチェック add-on at a revenue share; their sales calls become yours. Even one partner listing on ITトレンド’s comparison page puts the product in front of every store shopping for a system.
• The blog that already exists. The itmedia author and the 業務レンタカー operator blog both rank for 「レンタカー 傷」. Pitch each a follow-up: “here is what the store should have had”. The renter-side traffic converts to nothing, but store owners read the same posts when they get burned.

10. Build complexity — justification

Medium. The capture app (guided walkaround, zone coverage checks, LINE delivery, e-signature) is standard mobile work. The comparison layer is prompt engineering on agentic video plus a panel map per body style and a per-store damage policy; no model training. The document layer is templating onto Japanese checksheet and 請求書 formats with a bilingual summary. The hard part is robustness: consistent results on dark paint, night returns and rain, and refusing to issue a verdict when coverage is poor rather than guessing. Two people, roughly 12 weeks to a v1 that two Okinawa stores use daily; a further month for reservation-system exports and the franchise dispute inbox.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Recording a vehicle’s exterior at handover with customer consent is ordinary practice; personal data (plates, faces in frame) is handled under APPI with notice at signature; no change to 貸渡証 rules.
Ethical — no harm / dark patterns	✅	The tool makes charges evidence-based in both directions; a clean return produces a signed all-clear the customer keeps. Policy toggles default conservative after the Niconico case.
Market exists (evidence above)	✅	国民生活センター alerts (2021, 2025, ~500 consultations a year), a franchise’s public reversal in March 2026, a dated operator admission in September 2026, and US majors paying for scanner gates.
1–5 person team can build this	✅	Two people, ~12 weeks to a usable v1 on Gemini Flash agentic video.
Launchable with <$50K / ₹40L	✅	Two phones, an Okinawa trip, model spend in the hundreds of dollars a month at 100 stores.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	15/20	Every missed scratch is ¥20,000–¥100,000 eaten or a fight; in season it is daily. But disputes hit a minority of rentals and stores have lived with the paper sheet for decades. Real, regular, not yet hair-on-fire for the store owner until a case goes public.
Demand evidence	15	10/15	Regulator alerts with counts, a dated public reversal, operator admissions, and majors paying UVeye. Missing: a store owner saying “I would pay ¥12,800 for this” and any Japanese vendor selling the diff (the absence cuts both ways).
Build feasibility	15	9/15	Off-the-shelf model, standard app, but robustness on dark paint, rain and night returns is an engineering grind and the product must refuse to guess. 12 weeks with discipline.
Distribution clarity	15	10/15	Named lists exist (MLIT operator data with fleet sizes, prefectural associations, two franchise HQs, three reservation vendors with no damage feature). Conversion of a cash-tight, seasonal, fax-era store is the unknown.
Revenue mechanics	15	9/15	Priced against ¥5,000–¥30,000 reservation systems and against one NOC; 750 stores for $1M is plausible. Churn in the off-season and the low ACV mean $5M needs an HQ or partner deal, not just stores.
Time to first revenue	10	7/10	An Okinawa store can be closed at the lot with a live demo; subsidy paperwork and a two-week trial push first cash to 6–10 weeks.
Defensibility	10	4/10	Copyable by Kaflix or a reservation vendor within a year. Soft moats: per-model damage priors, franchise HQ contracts, the reservation-system integrations, and being the name in the next viral case.
Total	100	64/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · sales-heavy · a Japanese-speaking founder or first hire is non-negotiable; store owners will not onboard in English.

Key assumptions to validate (3–5)

1. Assumption: agentic video on two handheld walkarounds detects new damage of 1 cm or larger with under one false positive per ten returns across paint colours and lighting. How to test: 40 pickup/return pairs at two stores, ten with planted new marks, shot by staff not by us; score detection and false positives by condition.
2. Assumption: a discount store owner will pay ¥12,800 a month rather than keep the paper sheet. How to test: 20 in-person demos in Naha on the store’s own car; count signed monthly agreements, not “interesting”.
3. Assumption: the bilingual all-clear is worth something to inbound-heavy stores. How to test: ask the Okinawa stores whether they would pay ¥40 per return for it alone; if fewer than a third say yes, drop it from the pitch and keep it as a feature.
4. Assumption: franchise HQs see the March 2026 case as a systemic risk, not a one-off. How to test: one meeting each with MIC and Gutsu; a 20-store pilot offer accepted or declined within 30 days.
5. Assumption: reservation vendors will list a third-party キズチェック module rather than build their own. How to test: pitch レンタカー侍 and れんサポ; a signed referral or listing agreement, or a stated build plan, within 60 days.

Risk flags

1. Platform dependency: the unit economics rest on Gemini Flash agentic video pricing; the introductory rate ends 31 December 2026 and doubles. Even at the standard rate a return costs yen, but keep a second model path.
2. Vendor build-out: Kaflix or a reservation system adds a photo-diff feature. Speed and the franchise HQ relationship are the only answers.
3. Robustness: a false “new damage” verdict at a counter is worse than a paper sheet; one such incident on X kills the brand. The product must show its uncertainty and ask for a re-shoot rather than guess.
4. Seasonality and churn: resort stores may pause in winter. Monthly pricing invites it; the per-car fleet memory is the retention hook.
5. Consumer backlash by association: the Hertz coverage taught Japanese renters that “AI found a scratch” means an inflated bill. Marketing must lead with the signed all-clear, not the claim.

14. Structured verdict

Score:                  64/100
Verdict:                VALIDATE
Confidence:             Medium
Best-fit builder:       Technical founder plus a Japanese-speaking seller who will spend spring in Okinawa rental lots
Time to revenue:        6–10 weeks
Capital to launch:      ¥1.5–3M ($10–20K): two phones, travel, model spend, IT導入支援事業者 registration
Top 3 assumptions to validate first:
  1. Detection quality — 40 walkaround pairs at two stores, ten planted marks, scored by paint colour and light
  2. Store willingness to pay — 20 Naha demos on the store's own car, count signed monthly agreements
  3. Franchise HQ appetite — one pilot meeting each with MIC (Niconico) and Gutsu inside 30 days
Kill criteria:
  - Abandon if false positives exceed 2 per 10 returns in any common condition (night, rain, black paint) after four weeks of prompt work
  - Abandon if fewer than 3 of 20 demoed Okinawa stores sign at ¥12,800/month
  - Abandon if a reservation vendor or Kaflix ships a pickup-versus-return diff before v1 lands in a store

15. Next step — 1-week validation sprint

• Day 1–2: two cooperating stores (one Naha, one Tokyo 業務レンタカー). Staff shoot 40 pickup/return pairs on their own phones, ten with planted marks (tape-removable chips, chalk scuffs, a real wheel scuff on a retired car). Run the diff; tabulate hits, misses and false positives by panel, paint and light.
• Day 3–4: print the best five diff pages and one signed all-clear. Walk the rental lots around Naha Airport and the 沖縄県レンタカー協会 list: 20 demos on the store’s own car. Offer ¥12,800 a month, subsidy form in hand. Log every objection verbatim.
• Day 5: go if detection is at or above 9 of 10 planted marks with at most 1 false positive per 10 returns, and at least 5 of 20 stores sign or commit to a paid trial. Otherwise write down which condition or which objection killed it and decide whether it is a prompt problem or a market problem.