Kashi 可視 · Pitch · 2026-04-21
01 / 10
The hook

A bounded governance instrument that helps organisations surface repeated communication asymmetries before they harden into late-stage people incidents.

¥7.6T
Japan's annual mental-health productivity loss · ~1.1% of GDP
Yokohama City University 2025
¥7.9M
Per mental-health leave case, total employer cost
METI 2025 benchmark (¥5.25M salary, 1 yr leave)
1 in 3
Japanese workers experienced in the last 3 years
MHLW
Full positioning statement

Kashi is privacy-bounded meeting governance and accountability infrastructure. A bounded review-support system for surfacing repeated interaction asymmetry in supported meeting contexts. Structural-first by default on the employer-facing lane; semantic-support detectors exist but are disabled by default, require tenant opt-in, and surface only to the affected individual or a neutral reviewer under audit.

Kashi 可視
02 / 10
The cost is a portfolio, not one hero number

¥7.3T of Japan's ¥7.6T loss is presenteeism. Absence is the tail.

Earlier pitches led with “¥7.9M per leave case.” That's tail risk, not base case. The defensible story is the four-bucket hidden-loss portfolio — and the A×C×I×E discount that separates “total pool” from “what Kashi realistically captures.”

Mid-plan · what Kashi realistically captures
¥2.22M/yr
1.75% reduction · 500-person co · ¥127M loss pool · 25×70×40×25%
Full 4-bucket loss portfolio
BucketWhat it capturesEvidence anchor
1. Presenteeism (biggest)Underperformance while at desk; invisible on headcount dashboardsYokohama CU 2025: ¥7.3T of ¥7.6T
2. Regrettable attritionManager-driven departures; replacement costs 200% of salaryGallup: managers = 70% of engagement variance
3. Formal-escalation premiumLegal fees, time-in-HR, reputation when informal → formalAcas: formal vs informal cost
4. Leave-case tailVisible end-stage: 傷病手当 claims, backfill, recruitingMETI 2025: ¥7.9M/case (tail, not base)

The key insight: leave-cost alone isn't the sell. The sell is the portfolio — presenteeism + attrition + escalation premium are 4–8× larger than the leave tail.

Strict & Upper A×C×I×E scenarios

Strict

15×60×30×20% → 0.54% reduction = ¥0.69M/yr

Upper

40×80×50×30% → 4.80% reduction = ¥6.10M/yr

Moral-weight evidence backing the budget

2.0× suicide-ideation, 2.67× suicidal-behavior odds (Leach 2023, Lancet Public Health). Laissez-faire managers produce 4.3× and 2.6× mental-health problems at 6mo (Japanese occupational-health research, 2023).

Kashi 可視
03 / 10
The problem

Four structural reasons existing tools fail.

01 · Burden on the victim

The person under pressure must recognize it, prove it, and raise it, alone. Power asymmetry makes all three hard. Gaslighting makes them doubt perception.

02 · Fragments, not patterns

Each moment can be explained away. The harm is in the pattern. Organizations only see isolated complaints, never the structure.

03 · Lagging indicators

HR sees attrition, leave, 自殺. By then the damage is done. By the time the metric moves, the bill has arrived.

04 · Single-constituency sales

Existing tools are sold to HR alone. A workplace-AI deployment that survives scrutiny requires executive sponsor + legal review + worker-representation alignment before launch. All three, or none.

Why existing competitors fall short

Archaic's ハラスメントチェックAI reads content and sells to HR alone.
FRONTEO KIBIT Eye, deployed at MUFG and Aeon, has gone quiet.
Content classifiers miss 70%+ of Japanese cases because is .

Kashi 可視
04 / 10
The insight

Point the mirror upward at power. Structural signals, not content.

Five empirically-backed structural signals. Each deterministic. Each explainable. Each maps to MHLW 3 (isolation) or 類型 5 (cold shoulder) — the transcript-visible harassment types.

SignalEvidence
Intrusive-interruption asymmetryAnderson & Leaper 1998 · meta-analysis 43 studies · d=0.33 · status beats gender
Speaking-time / floor-share inequalitySchmid Mast 2002 · Human Communication Research · dominance correlation
4 more structural signals + calibration method
SignalEvidence
Topic-credit exclusion (ignored → taken over)Sacks / Schegloff / Jefferson 1974 · foundational conversation analysis
Chilling delta (post-trigger participation drop)Morrison 2014 · organizational silence · Detert & Burris 2007
Response-latency asymmetryStivers et al. 2009 PNAS · delayed response = dispreference cross-linguistically
Keigo () peer-addressee asymmetry (JP-specific)Cook 2011 · Saito 2011 · Pizziconi · Ide wakimae framework

False-positive mitigation: every signal calibrated against each speaker's own 90-day baseline, not the team average. Defeats introversion, chair-role, L2, neurodivergence confounds.

Kashi 可視
05 / 10
Live walkthrough · drive with arrow keys

Watch the system work on one real meeting, end to end.

Six frames. Same meeting. Structural signals → classifier → 63-day pattern → three different views for three different roles. Real data, no mocks, no voiceover — you drive it live in front of the room.

Open the walkthrough → Fallback if wifi drops: /demo/mirror · /demo/ceo
What you'll see in each frame

Frame 01–03 · system

One 47-minute product-review meeting. Kimura cuts Nao mid-word at 00:01:05 (200ms overlap). Layer 1 extracts 1 interruption, chilling delta −93%, Gini 0.34. Layer 2b (keigo) classifier: Kimura scores 0.38 toward Nao, 0.88 toward Nakamura. 0.50 asymmetry gap. Same speaker, same meeting.

Frame 04 · pattern

Zoom out: this is not one meeting. Nao's speaking share dropped 19% → 6% sustained over 63 days. Kimura's interruptions toward her 4.7× higher than toward peers. Persistence + directionality + sustained drop = review-worthy event fires at Layer 5.

Frame 05–06 · money shot

Three windows side-by-side. Nao sees a private, observational pattern page. Kimura sees his own Manager Mirror with one concrete action. CEO sees the Executive Brief: Kimura flagged, ¥3–8M modeled impact. Three framings, one data source, RBAC-separated.

Kashi 可視
06 / 10
The system · one meeting, all 6 layers

Every number below is computed from the same 47-minute transcript.

m-product-w12 · 2026-04-15 · 89 turns · 5 participants. No abstractions. No “it would look like this.” Real.

LAYER 5
review-worthy event
Composite: severity × repetition × directionality × confidence = 0.87. Threshold 0.60. FIRES. Every component traces to specific turn IDs. Explainable. Reproducible. No LLM judgment in the scorer.
3 / 3
Patterns detected on seed
0
False positives on healthy control
Deterministic
Same input → same output replay
8
Detectors live · 4 structural, 3 hybrid, 1 text-deterministic (keigo)
All 6 layers feeding Layer 5 — same meeting
LAYER 1
deterministic
Turn-timing extraction. 1 intrusive interruption (Kimura→Nao, overlap 200ms). Speaking-share Kimura 42% / Nao 11%. Chilling event: Nao's 3 turns after the cut average 1.3s (vs her 18s personal baseline).
LAYER 2
structural wrappers
Dyadic continuity check: Kimura→Nao interruption pattern appears in 3 of last 3 meetings, span 28 days → persistence score 0.45. Speaker-baseline drift: Nao's 90-day share trend 19% → 6%, sustained below baseline for 63 days.
LAYER 2b
classifier
Per-addressee politeness-register index from surface grammar only. Kimura's keigo toward Nakamura: 0.88 (honorific), toward Yoshida: 0.50, toward Nao: 0.38 (plain form, ""). Asymmetry gap 0.50 — review-worthy at ≥0.30.
LAYER 3
meeting metrics
Floor-time Gini 0.34 (heavy skew). Interruption-asymmetry matrix: 82% of Kimura's cuts land on one target. Response-reciprocity: when Nao speaks, response-rate from Kimura is 0.18, from others 0.71.
LAYER 4
longitudinal
Rolling 30/90/180-day windows, per-speaker calibrated. Directionality ratio 4.7×. Sustained-drop 63 days. Keigo-asymmetry persistent across 4 consecutive meetings. This is where one meeting becomes a pattern.
LAYER 6
role-based view
RBAC + k-anonymity (k≥5) + differential privacy (ε≤1). Three different framings render: Nao's private pattern page (observational), Kimura's Manager Mirror (own behavior + action), CEO's Executive Brief (aggregate, no individual subordinate data). See frame 05 of the walkthrough.
Kashi 可視
07 / 10
Governance · 3-lane accountability model

Coaching is not the endpoint.

Manager Mirror is a bounded developmental lane inside a broader accountability system. Private self-correction first. If the pattern persists, it escalates. That is the anti-laundering rule: Kashi is not a device for saying “we told him” and doing nothing.

Lane A · Private self-correction

Manager sees their own pattern in Manager Mirror. Private, weekly, observational language, paired with feedforward commitments ("next 2 meetings, wait for sentence completion before redirecting"). ~80% of cases resolve here if the manager is coachable.

Lane B · Governed remediation

If the pattern persists after the protected self-correction window, it leaves the developmental lane. Moves to a documented remediation process with timeline, check-ins, and structured support. Not HR punishment. Not silent.

Lane C · Formal review

Only if Lane B fails. Strict entitlement rules, audit logging, and employee notice required. The organization can no longer plausibly claim it didn't know. The tool does not adjudicate — humans do, under process.

Architectural guarantees & what Kashi will not do

Architectural guarantees

  • k-anonymity (k≥5) on aggregates · differential privacy (ε≤1) on exec views
  • Per-speaker baseline — no cross-person norms
  • Sample-size suppression (<5 meetings / <30 days = no signal)
  • Audit trail visible to affected individual
  • No named subordinate telemetry on manager views by default

Will not do

  • Detect harassment, intent, or illegality
  • Infer emotion / affect (EU AI Act Art. 5)
  • Feed HR decisions (Annex III §4)
  • Read message content server-side / analyze audio
  • Show company-wide health score
  • Allow coaching to be the endpoint when asymmetry persists
Kashi 可視
08 / 10
Security · Board-level confidence

Where your data lives, who can read it, what happens if we're breached.

The six questions every CEO or board asks before signing off.

Japan-region residency + AES-256 / TLS 1.3 Tenant data stays in Tokyo/Osaka (ap-northeast), never crosses a border. Encryption at rest and in transit, quarterly key rotation, per-tenant keys on Enterprise.
Structural-only by default · no third-party content processing All 4 structural detectors + the keigo classifier run in our own infrastructure with zero transcript egress. The 3 hybrid-text detectors (topic-credit, agreement-asymmetry, unanswered-question) are disabled by default; tenants who enable the semantic lane process transcripts via the Anthropic API under a signed BAA — no OpenAI, no Google, no third-party training pipelines.
RLS enforced at the query layer — not the UI Supabase Row-Level Security. Managers cannot SELECT subordinate individual data even with direct DB access. The schema is the boundary.
Four-tier retention with hard TTLs Transcripts 14 days. Analytics 24 months. Review events 12 months. Legal-hold only when justified and notified. Automated, auditable deletion.
Audit log the affected employee can read Every drill-down writes an audit row. The person being observed sees who looked. No silent surveillance.
No admin content access + 24h breach SLA Kashi staff do not have routine access to customer transcripts. Break-glass is pre-committed, reason-coded, and audit-logged. Tenant notification within 24h; regulators per APPI/GDPR; public post-mortem within 30 days.
SOC 2 Type II / ISO 27001 on roadmap Year-1 SOC 2 Type II. Year-2 ISO 27001. Annual third-party pentest + continuous vulnerability scanning.
Works-council consultation built in Enterprise activation requires documented revision with written employee-representative opinion. GDPR Art. 88 + JP labor-law aligned. Real process, not a notice screen.

Deployment preconditions — Kashi is not a silent pilot

Before launch: executive sponsor + legal review + worker-rep alignment, all three documented. Every deployment ships with a labor-consultation packet (purpose statement, access matrix, retention justification, challenge workflow, anti-retaliation commitments, sunset rules). Every retention window and drill-down has a one-line necessity argument (UK ICO / CNIL / EDPB style).

Kashi 可視
09 / 10
Why now · why us

Nobody has shipped this yet. Six reasons it's finally possible.

01 · Transcription

JP ASR on multi-speaker meetings only became good enough in 2023–2024. Sortformer v2 DER on CALLHOME: 12.7%.

02 · Regulation

EU AI Act Art. 5 (Feb 2025) + Annex III §4 (Aug 2026) crystallized the defensible approach.

03 · New category

“Governance infrastructure” bridges engagement tools, compliance tools, productivity tools. Courage to cross.

04 · JP politics

Pointing the mirror upward at power is politically dangerous. It's our defining feature.

05 · Privacy cost

K-anon + DP + 4-tier retention + audit trails + 就業規則 consent is 6–12 months of infra most skip.

06 · Wrong buyer

Traditional HR buyers want defensive tools. CEO buyers want preventive. Different sale, higher value.

Competitor capability matrix · Kashi vs. 4 alternatives
Capability Kashi Archaic / FRONTEO Viva Glint Read.ai Wevox / Geppo
Structural pattern detection (not content)
Mirror points upward at powerrefused
Keigo (敬語) asymmetry (JP-specific)
Refuses company-wide health score
EU AI Act Art. 5 clean by design????
Published harassment-outcome evidencepending NAQ-R study
Kashi 可視
10 / 10
ROI · pricing · ask

¥13M–¥26M/year recoverable per 500-person company.

Starter
Free
up to 20 employees
Professional
¥800/emp/mo
up to 500 employees
Enterprise
Contact
500+ · multi-region · SOC 2

Pricing provisional · under validation.

The ask

Not “introduce me to the CEO.” A pilot that survives scrutiny needs three introductions at one 50–500-person JP company:

  • Executive sponsor — commits to the 3-lane architecture
  • Legal / compliance — APPI + labor-law sign-off
  • Worker representative — consults on the amendment

90-day pilot. Real data, real consultation, real outcome measurement.

How the ROI math works

500-person co × ¥5.25M avg salary × 5% productivity drag = ¥131M/yr at stake. Shifting 10–20% of cases from late- to early-stage = ¥13M–¥26M/yr recovered. Model-based plausibility check — not a proven return.

Full tier comparison · 8 dimensions
Starter Professional Enterprise
Scaleup to 20 employeesup to 500 employees500+ / multi-region
Detectors4 core structuralall 8 (incl. keigo & response-latency)all 8 + evidence vault (E2E)
ViewsManager Mirror + self-visibility+ Executive Brief + feedforward+ custom role matrix
Retention90-day hard delete24-month analyticsconfigurable
Data residencyshared JP regiondedicated JP regionchoice incl. on-prem (roadmap)
Labor-consultation packetdownloadable templatehands-on review
Compliance artifactsSOC 2 Type II + ISO 27001
Supportcommunityemail · 72h SLA24/7 · dedicated CSM
ROI multiples at 100- and 500-person scale
~8× modeled
100-person co at Pro: ¥960k/yr. 1 averted case ≈ ¥7.9M. Directional heuristic.
~3.3× modeled
500-person co at Pro: ¥4.8M/yr. 2 averted cases ≈ ¥15.8M. Assumption-sensitive.
Enterprise · scoped
Contact-led. Priced against deployment complexity + fixed-service scope.

ROI values above are model-based plausibility checks, not proven returns. Real validation comes from signed deals and discount behavior, not a slide.

4-phase product roadmap
  • Now: proof of signal, live at kashi-lilac.vercel.app
  • Phase 2: governance-ready, NAQ-R outcome validation, labor-consultation packet v1
  • Phase 3: SOC 2 Type II, E2E evidence vault, remediation workflow
  • Phase 4: formal-review interface, jurisdiction playbooks (JP / DE / NL / UK)