Grounded in the 22-memo research library. Honest ROI math. Japan-first. Consultation-compatible by architecture.
Kashi is privacy-bounded meeting governance infrastructure for Japanese mid-market employers. It surfaces repeated interaction asymmetries (floor dominance, cut-off concentration, topic-credit gaps) for human review — not emotion or sentiment, not individual performance scoring.
MHLW パワハラ 6類型 + 2022 SME law make meeting-dynamic oversight effectively mandatory in Japan. No product today delivers it in a way that survives both legal (APPI, 就業規則) and trust (worker-rep) review. Kashi does.
Three views (Manager Mirror · Executive Brief · Evidence Vault) over the same structural detectors. Privacy enforced by the schema (RLS), not the UI. Deployment requires documented executive + legal + worker-rep consent.
Three-buyer framework: executive sponsor + legal/compliance + worker representative. All three required before pilot. Target: 50–500-employee JP mid-market, starting with companies already revising 就業規則 to reference パワハラ防止法. Expansion: Japan → Singapore → NL/UK → DE → US.
Starter free · Professional ¥800 per covered employee/mo · Enterprise contact-led. Pricing structure is coherent enough to test, not yet commercially validated. See pricing strategy.
Procedural-fairness spine (k-anonymity, per-speaker baselines, audit log visible to affected individual, no named subordinate telemetry by default). Not a feature moat — a deployment-posture moat. Competitors who try to copy this ship a product their worker reps reject.
Introduction to three roles at one 50–500-person Japanese company: executive sponsor, legal reviewer, worker representative. 90-day pilot with real data, real consultation, real outcome measurement. Not “intro me to the CEO.”
Kashi is a privacy-bounded meeting governance system that surfaces repeated interaction asymmetries — under explicit uncertainty — as contestable structural signals for human review, inside strict procedural and use limits.
We are not a harassment detector. We are not a coaching dashboard. We are not a culture-survey tool. We are the governance layer that makes repeated meeting-level patterns institutionally legible and procedurally contestable — so organizations can no longer plausibly claim they didn't see, and so the people inside them can no longer silently carry the cost.
Earlier decks led with "¥7.9M per leave case × N cases = ¥X saved." That math is too coarse. The rebuilt-ROI memo makes it precise: you must apply an A×C×I×E discount stack — Addressable share × Coverage × Intervention uptake × Effect size — to go from total loss pool to realistically captured loss. We rebuild the pitch accordingly.
| Bucket | What it captures | Evidence anchor |
|---|---|---|
| 1. Presenteeism (biggest) | People at their desks underperforming while carrying unresolved interpersonal weight. Invisible on the headcount dashboard. | Yokohama CU 2025: ¥7.3T of ¥7.6T loss is presenteeism, not absence. Deloitte UK 2024: £24B of £51B is presenteeism. |
| 2. Regrettable attrition | Replacing lost talent when the loss was avoidable. Manager-driven departures cost ~200% of salary to replace; technical ~80%, frontline ~40%. | Gallup 2025: managers account for 70% of team engagement variance. Global manager engagement fell to 22%. |
| 3. Formal-escalation premium | The 3× multiplier on cost once a conflict passes from informal to formal resolution. Legal fees, time-in-HR, distraction, reputation. | Acas (UK): formal procedures cost 3× informal resolution. £28.5B/yr UK conflict cost. |
| 4. Leave-case tail | The visible end-stage: 健康保険傷病手当 claims, temp backfill, training-replacement. Late and late-stage. | METI 2025: ¥9M+ per year-long leave at ¥6M salary; ~¥7.9M at ¥5.25M salary. But: this is tail risk, not base case. |
Assumes ¥5.25M average salary, 5% productivity drag attributable to harmful team dynamics, 30 employees on leave in a year.
| Scenario | Total annual loss pool | Notes |
|---|---|---|
| Strict / employer-cash floor | ¥20.2M | What even a hostile CFO must accept |
| Base case / operational-economic | ¥127M | Presenteeism + overtime spillover + coverage drag |
| Upper / aggressive | ¥205M | All buckets loaded |
Per the rebuilt-ROI memo: multiply total loss × A × C × I × E where A = addressable share of the pool we can touch, C = coverage we achieve, I = intervention uptake, E = effect size on those who acted. Realistic ranges for a mature deployment:
| Scenario | A × C × I × E | Total reduction | Captured savings |
|---|---|---|---|
| Strict | 15 × 60 × 30 × 20% | 0.54% | ¥0.69M/yr |
| Mid (plan) | 25 × 70 × 40 × 25% | 1.75% | ¥2.22M/yr |
| Upper | 40 × 80 × 50 × 30% | 4.80% | ¥6.10M/yr |
These are captured-loss numbers for a 30-leave-case/year company. Translate per-customer: a 500-person company paying ¥4.8M/year for Professional tier lands at 0.5–1.3× direct ROI on leave-linked savings alone — which is why leave is not the sell. The sell is the portfolio (presenteeism + attrition + escalation premium) where Kashi's value is 4–8× higher per saved case.
1.5–2M companies worldwide with 50+ employees in knowledge-work sectors and a labor regime that rewards preventive governance. Blended ACV ¥3M/yr → ¥4.5–6T/yr ceiling. We don't pitch this number. SAM is what's real.
JP enterprises 50–500 employees × knowledge-work filter = ~45,000 companies. Blended ACV ¥4.2M/yr (85/15 Pro/Enterprise mix) → ~¥190B/yr revenue ceiling in Japan alone.
Year-3 target: 500 paying customers. ~1.1% of SAM. Conservative for a governance-compliance category (typically saturates at 5-10%). Y3 ARR: ~¥2.5B.
The founder-narrative and labor-politics memos converge: Kashi has three buyers, each with different language needs. Treating the CEO as the only buyer is how this category's prior attempts died.
CEO / CFO / COO. Needs the portfolio math (hidden loss, regrettable-attrition exposure, formal-escalation premium). Signs the contract. Commits to the 3-lane accountability architecture.
HR / Legal / Compliance / CHRO. Needs the procedural-fairness spine (contestability, bounded context, retention-under-challenge, meaningful human review, rollback triggers). Clears the CAIQ/SIG-lite security review.
Worker representative / union delegate / employee-rep / works council. Needs the anti-capture + bounded-visibility story. Signs off on 就業規則 amendment. Their consent is architectural, not ceremonial.
| # | Channel | Share of pipeline | Notes |
|---|---|---|---|
| 1 | Direct outbound to JP CEOs / CFOs of mid-market knowledge-work | 30% | Target: post-IPO, post-recent 労基署 inspection, post-public resignation. Message: hidden-loss portfolio, not "surveillance." |
| 2 | Labor-law firm referral partnerships | 20% | Oh-Ebashi LPC, Mori Hamada, Anderson Mōri. 10% first-year rev share × 12mo. |
| 3 | Worker-representative + union engagement | 15% | NEW — Rengo, Zenroren affiliate unions, works-council networks. Co-credentials the tool; unblocks Buyer 3. |
| 4 | Academic partnership (NAQ-R validation) | 10% | National-affiliated workplace-harassment research labs or 慶應 productivity-health. Long-cycle credibility play. |
| 5 | JP VC portfolio intros | 10% | DNX, Globis, ANRI, WiL — 5-10 warm intros per seed check. |
| 6 | Content + research publishing | 5% direct, 100% ambient credibility | The 22-memo library is substrate for ~2 articles/month. 日経 Business, Tech In Asia, note. |
| 7 | Occupational-health channel (JOHAS / EAP partners) | 5% | NEW — reframe Kashi as pre-consultation visibility layer for occupational-health intake. |
| 8 | Ombuds + neutral-review associations | 5% | IOA (International Ombuds Association) JP chapter. Critical for future Ally/Observer Concern Path rollout (V2). |
| Tier | Cycle | Decision path | Pilot structure |
|---|---|---|---|
| Starter | Self-serve, 1-week activation | Team lead signs up, 20-employee cap | No pilot — free forever at that scale |
| Professional | 30–90 days | CEO + HR/Legal + worker-rep (3 signatures) | 30-day free trial → annual |
| Enterprise | 4–9 months | All 3 buyers + IT security review + CAIQ/SIG pack + works-council consultation | 90-day paid pilot (1/6th year-1 contract) → 3-year |
Pilot-to-paid conversion target: 60%. Three-role qualification pre-filters heavily, so pilot entry is already a strong buying signal. Floor 40%.
| Tier | Price (status) | Scope | Illustrative ACV |
|---|---|---|---|
| Starter | Free (published) | ≤20 employees | ¥0 |
| Professional | ¥800 per covered employee/mo (published, provisional) | ≤500 covered employees | ¥1.44M @ 150 covered employees |
| Enterprise | Contact-led (unpublished) | 500+ / multi-region / high-compliance | Internal hypothesis: ~¥19.6M @ 1,000 covered employees (¥10M base + ¥800 per covered employee/mo) |
| Tier | CAC (target) | ACV | Payback | LTV (4-yr retention) | LTV/CAC |
|---|---|---|---|---|---|
| Professional | ¥400k | ¥1.44M | 3.3mo | ¥4.9M | 12× |
| Enterprise | ¥4M | ¥19.6M | 2.4mo | ¥51M | 13× |
| Period | Pro | Ent | New ARR | Cumulative |
|---|---|---|---|---|
| Q1 Y1 | 2 (pilots) | 0 | ¥2M | ¥2M |
| Q4 Y1 | 8 | 2 | ¥40M | ¥52M |
| Q4 Y2 | 60 | 15 | ¥280M | ¥380M |
| Q4 Y3 | 400 | 100 | ¥1.3B | ¥2.5B |
The founder-narrative memo is direct: don't fabricate pedigree. The honest "why us" for Kashi is not resume-based; it's disciplined category design. The pattern of refusal — no affect inference, no content reading by default, no HR decisions, no health score — is the answer to "why are you the right team?" That's defensible because it's rare and it's observable in what we've already built.
We built Kashi because existing workplace-AI products fail a specific test: they look legitimate to whoever is paying, and illegitimate to whoever is being analyzed. The products that claim "we detect harassment" are the ones that should worry you most. The products that claim "we're just about productivity" quietly accumulate evidence that gets misused later.
We chose a different starting point. Start with the refusals. Make the refusals architecturally enforced, not contractually promised. Make the rhetoric match the architecture. Then build the smallest useful product inside those limits.
The 22-memo research library behind this plan is part of the answer to "why us": we did the reading before we wrote a line of code, and we wrote the code to survive the reading. The product at kashi-lilac.vercel.app is the proof that this discipline produces a shippable artifact, not just a whitepaper.
Per the founder-narrative memo, keep this epistemically honest:
The funding-pathways memo identified a track most founders miss: Japan has multiple non-dilutive programs that Kashi fits — not as "mental health" (we're not eligible) but as AI adoption and workstyle reform. Run both tracks in parallel. Equity pays for velocity; grants pay for customer-side procurement subsidy that makes the enterprise sale cheaper.
| Program | Country | Fit | Timing |
|---|---|---|---|
| Digitalisation & AI Adoption Subsidy 2026 | Japan | P1 — high | Frame as AI tool for workstyle / governance. Subsidizes customer purchase, making enterprise deals cheaper. |
| Workstyle Reform Promotion Support Subsidy 2026 | Japan | P1 — medium-high | Direct fit: Kashi supports preventive labor-governance under 労働施策総合推進法. |
| JOHAS occupational health routes | Japan | P1/P2 | Channel partnership with occupational-health providers; pre-consultation visibility layer. |
| PSG (Productivity Solutions Grant) | Singapore | P1 — high | Singapore is the first international market (per the funding memo), not EU. |
| EDG (Enterprise Development Grant) | Singapore | P1 | SME-facing co-funding for transformation projects. |
| DIGITAL / EDIHs / Apply AI Strategy | EU | P2 — medium | Pilot-legitimacy funding, not revenue. |
| OH / welfare-counselling tax routes | UK | P2 — medium | If Kashi bundles with EAP. |
The legal-procedural-fairness memo reframes what our moat is. Detectors are table stakes. The moat is the procedural architecture around the detectors — specifically, the contestability state machine, the bounded-context rule, retention-under-challenge, meaningful human review, downstream-use charter, and rollback triggers. Harder to copy than detectors. Much harder.
Every detector declares its lane at compile time in{" "} src/lib/pipeline/detector-registry.ts. Three structural detectors (intrusive-interruption, chilling-delta, floor-time Gini) touch no transcript text and default to employer-facing. Four hybrid detectors (unanswered-question, topic-credit, agreement-asymmetry, keigo) read transcript text and are gated behind a tenant feature flag that defaults OFF.
This replaces the blanket "metadata only, no content reading" claim (which was technically false given our embedding-based similarity detectors) with an honest, machine-readable taxonomy that procurement reviewers and legal counsel can audit. Adding a new detector requires updating the registry. Changing a detector's lane requires governance review.
The "Kashi will not do" list is the product. Viva can't copy it without dismantling Viva's value proposition. Archaic can't copy it without abandoning content-scanning. Adding our refusals to their products requires gutting their products.
Baseline stack = self-history + within-meeting + meeting-type + role + dyad. After 500 customers, we can tell a 150-person SaaS "your keigo-asymmetry is 78th percentile for your sector and size" — not replicable without matching customer density.
SOC 2 Type II, ISO 27001, JP 就業規則 template library, jurisdiction playbooks for NL/UK/DE. Each takes 12-18 months. Table-stakes by the time a competitor catches up; we have more reference chains.
The retaliation-risk memo surfaces a moat layer we almost missed. Employer must not be able to infer that an employee opened their pattern page, created a vault, marked confounds, began a draft, or triggered a review. Telemetry partitioning, protected routes, min-group-size suppression, batching/redaction/delay — all architectural, none rhetorical. Almost impossible for Viva-style products to bolt on because their analytics stack is designed to log exactly these signals.
Expanded from 6 to 11 risks based on the adversarial + anti-capture + retaliation + false-negative + attack-surface memos. Ranked by impact × probability. Each has an architectural mitigation — not just a policy.
Probability: medium. Impact: catastrophic.
Mitigation: The 3-lane accountability architecture IS the defense. If Lane-B governed remediation is contractually enforced + documented, the pattern's visibility supports duty-of-care rather than undermining it. Pre-pilot: signed legal review acknowledging remediation obligation. Audit trail designed to be discovery-friendly for the company.
Probability: medium-high in 24mo. Impact: large on Pro tier, small on Enterprise.
Mitigation: Retaliation-risk memo upgrades this: incumbents can't copy anti-inference architecture without cannibalizing their analytics business. Enterprise moat (Layers 1 + 4 + 5) is intact. Starter becomes lead-gen funnel; revenue migrates toward Enterprise over time.
Probability: high early. Impact: 6-month Y1 slip.
Mitigation: Target pilot customers with mature employee-rep infrastructure (post-IPO JP companies, regional manufacturers with established 労使関係). Free first 5 pilots in exchange for case-study rights.
Probability: high once product is socially known. Impact: undermines claim base.
Managers who know the detector surface route pressure around it: metric substitution (shift dominance to agenda control), channel displacement (move to 1:1s, async), hierarchical laundering (push decisions via proxies), symbolic compliance (cleaner surface, same dynamics).
Mitigation: "Adaptation-watch" product layer that flags suspiciously clean metric improvement as a signal, not success. Ban single-metric victory claims. Multi-metric corroboration required. Never market "interruption down X%" as proof of improvement.
Probability: medium. Impact: large — category-killing if exposed publicly.
An org buys Kashi, excludes executive leadership from scope, targets only ICs or middle managers. "Kashi surveils workers, protects leaders" becomes the story.
Mitigation: Gate 1 of 6 deployment gates (scope parity — leadership in scope). Non-negotiable before pilot launch. Contractually enforced. If Gate 1 fails, pilot doesn't launch.
Probability: medium without careful design. Impact: catastrophic in JP (MHLW prohibits disadvantageous treatment for consultation).
Employer infers that a worker opened their pattern page, created a vault, marked confounds, or began an escalation draft — even without seeing content. Directly actionable retaliation risk.
Mitigation: Telemetry partitioning (security vs analytics on separate surfaces). Protected routes (pattern-page opens, vault creation, draft state never exposed to business analytics). Small-team inference suppression. Batching + redaction + delay on any employee-facing event before it reaches analytics.
Probability: high if unmitigated. Impact: category-killing.
Employer cites clean Kashi dashboard as evidence-of-absence in a harassment complaint. "We ran the system, no signal, therefore no harm." EEOC base rate: 3 of 4 harassment targets never speak up.
Mitigation: 4-state dashboard (signal / no qualifying signal / insufficient observation / out-of-scope — never binary). Case-review rule explicitly prohibits complaint closure on "no signal." Scope warnings surfaced in UI, not buried in governance page. Contract prohibits citing Kashi output as exculpatory evidence.
Probability: known-present, continuous. Impact: medium-large if unmitigated.
NOTSOFAR-1 baseline speaker-attributed tcpWER: 32.4% multichannel / 46.8% single-channel. 54.9% on overlap-heavy sessions. Koenecke: 0.35 vs 0.19 WER racial disparity. Disfluency bias documented.
Mitigation: Input-quality gating before output is pilot-grade (overlap flags, L2 caution surfaces, subgroup audits). Baseline reset rules after reorg / manager change. Per-platform × language × feature support matrix published.
Probability: low-medium. Impact: large.
Mitigation: Architecture is already more restrictive than current rules. 6+ months of product-side lead time given deployment preconditions. Regulatory advisory-board seat monitors monthly.
Probability: low. Impact: catastrophic if public.
Mitigation: "No export to HR systems" is product-enforced, not just contractual (no API, no export format). Public commitments on deck + governance page. Incident response plan: public post-mortem within 30 days.
Probability: medium. Impact: soft — erodes outcome claims but doesn't sink product.
Even with only self-comparison (not public leaderboards), managers game the private mirror. Keigo-asymmetry and interruption signals improve structurally while dominance expresses in untouched channels.
Mitigation: Remediation outcomes + human recovery as mandatory measurement layers (see §11). Metric movement alone is never the victory claim.
Delaware C-corp + JP KK subsidiary. Enables US VC participation, preserves Series A optionality, supports JP-compliant billing via the KK.
The procurement / security-buyer memo flagged a concrete infrastructure problem: Vercel's primary processing facilities are in the United States. If Kashi makes a Japan-data-residency claim to a JP enterprise buyer and any regulated content flows through Vercel compute, the claim is technically false.
Fix (before first JP Enterprise deal):
| Stage | Headcount | Structure |
|---|---|---|
| Pre-seed (now) | 2-3 | Founder(s) + fractional labor-law advisor |
| Post-pre-seed (mo 6) | 4 | + 1 eng + 1 JP enterprise salesperson |
| Post-seed (mo 18) | 8 | + 1 security/DevSecOps + 1 CS + 1 product + 1 ops |
| Post-Series A (Y3) | ~35 | eng (12) + GTM (10) + CS (6) + ops (4) + founder-office (3) |
V1 of this plan said "EU first (DE/NL/UK), US second." The cross-cultural + labor-politics + funding-pathways memos converge: reverse this. Japan-first is the doctrine, not a fallback.
| Phase | Markets | Why |
|---|---|---|
| Y1-Y3 (harden) | Japan | Regulatory conditions most favorable. Channels established. Funding programs (AI Adoption, Workstyle Reform) fund customer acquisition. |
| Y3 selective | Singapore | PSG + EDG grants, English-capable workforce, procurement discipline similar to JP. First international. |
| Y4 | NL + UK | Procedurally lighter than DE. UK ICO + EU AI Act compliance already aligned with our architecture. |
| Y5 | DE | After works-council playbook matures. BetrVG §87(1)(6) requires co-determination on monitoring tech — heavier. |
| Y5+ | US | After SOC 2 Type II + ISO 27001. US market larger but saturated with HR-tech incumbents. |
JP Mothers listing at ~¥5B ARR (standard SaaS exchange threshold). 5-7 year horizon. Product improves with data (cross-industry baseline moat); category is a new SaaS segment we'd define. Don't optimize for acquisition. The value to society is highest if we become the JP governance standard — which requires independence to say no to acquirers whose business models conflict with our refusals.
The remediation-outcomes memo rewrites what success looks like. We cannot declare success on detection metrics alone ("caught 3 out of 3 seed patterns"). That's proof of mechanism, not proof of value. Traction must add two further outcome layers.
| Layer | What it measures | Example metric |
|---|---|---|
| 1. Adoption | Is anyone using the product? | Pattern-page monthly active users, manager feedforward-check rate |
| 2. Behavior | Did the manager's own pattern change? | Structural-signal movement + adaptation-watch flags |
| 3. Responsiveness | Did the org act? | Time from flag → review, % Lane-B transitions after persistence |
| 4. Strategic | Did the category of problem recur? | Recurrence rate at 30 / 90 / 180 days |
| 5. Remediation quality (NEW) | Was the handling actually fair? | Worker-rated fairness, clarity, follow-through at 30-day post-handling |
| 6. Human recovery (NEW) | Did the affected person recover? | Speaking-share recovery, answered-question recovery, chilling-delta reduction at 30/60/90 |
Paying 50-500-person JP companies actively using Kashi with (a) a signed 就業規則 amendment documenting consultation, and (b) at least one completed Lane-B remediation cycle where worker-rated fairness ≥ 4/5.
Why: it rewards the end-to-end outcome (not detection alone), penalizes vanity signups, and specifically resists Goodhart by requiring worker-side assessment.
Six of the research memos independently flag the same risk: rhetorical overreach is the fastest way this category gets killed. Here's the discipline, codified.
Everything in this plan traces to one or more of the memos below. They are investor-read artifacts; open source for diligence.
24 PDFs · all hosted at /research/*.pdf · license: internal research, freely shareable.
A separate body of 17 technical-dev memos + 1 README, explicitly marked as perspective-expanding material, not binding specifications. These memos directly shaped the new detector registry, evidence-grade types, meeting-type normalization, and procedural-fairness spine. Hosted for diligence:
42 total documents across both libraries · 24 business research + 18 technical-dev consideration · all hosted at{" "} /research/ and{" "} /research/ideas-wave3/.