A consistent pattern across eight jurisdictions: practitioners already work with AI
FIRM DEPLOYMENT11–46%
Privilege and compliance risk blocks procurement — and the scissors are closing
0255075100%
World's first AI-related practice restriction on a lawyer (Australia, Sep 2025)
UK SRA warns privilege can be permanently lost; 42 AI-misuse reports in one year
PI insurers issue AI exclusion endorsements — the buying trigger has been pulled
Shadow AI: 78% of AI users bring their own AI to work · sensitive share of corporate data pasted into AI rose 34.8% → 39.7% · 20% of breach incidents involve unapproved AI use, 670,000 USD more per incident [F]
Demand isn't missing. The compliant channel is. Tacita is that channel.
[F] Compiled from regulatory & industry sources across eight jurisdictions; available in the data room
問題 PROBLEM
律師已在用 AI，機構未敢批准。
個人採用61–89%
八個法域趨勢一致：前線從業員已在日常工作中使用 AI
機構部署11–46%
合規與特權風險窒礙正式採購，剪刀差持續收窄
0255075100%
全球首宗律師 AI 執業限制處分（澳洲，2025-09）
英國 SRA 警示：特權可永久喪失；一年內 42 宗 AI 濫用報告
PI 保單增設 AI 除外條款——採購誘因已成事實
影子 AI：78% AI 用戶自攜工具工作 · 貼入 AI 的企業數據中敏感資料佔比由 34.8% 升至 39.7% · 20% 資料外洩事件涉及未經批准的 AI 使用，該等事件成本高出 670,000 USD [F]
問題不在需求，而在缺乏合規通道。Tacita 正是這條通道。
[F] 八個法域監管及行業資料彙編，盡職審查資料室可供查閱
TACITA · PITCH DECK02 / 18
CASE LAW
Sanctions are now facts.
2023-06
Mata v. Avianca (S.D.N.Y.) — Rule 11 sanctions US$5,000 — six ChatGPT-fabricated cases filed
2025-06
R (Ayinde) v Haringey; Al-Haroun [2025] EWHC 1383 — £1,500 wasted-costs order + referral to Bar Standards Board
2025-09
Victoria Supreme Court (Melbourne, ordered 2024-10) — practice-certificate conditions (barred as principal) — the world's first AI practice restriction
2025-08
Western Australia lawyer (Federal Court) — "using AI to check AI" still failed — referred to regulator
HK / SG
Hong Kong / Singapore — gap: no public AI sanction — the education window
The common sanction logic is "failure to verify," not "using AI" — a mandatory human gate plus an audit chain answers both risks.
THE REAL COST OF DIY
Manual redaction + free model: ≈26–96 USD per document [A], no audit trail — Personal at 399 USD/yr ≈ less than two DIY documents a month. Lost billable capacity: ≈21,800–41,000 USD/yr per lawyer [A].
判例線
處分已由概率變成事實。
2023-06
Mata v. Avianca（S.D.N.Y.）——Rule 11 罰款 US$5,000，提交 ChatGPT 虛構的 6 宗案例
2025-06
R (Ayinde) v Haringey；Al-Haroun [2025] EWHC 1383——大律師被處 £1,500 浪費訟費令，並轉介 Bar Standards Board
2025-09
維州最高法院案（墨爾本，2024-10 判令）——執業證書附帶條件（不得擔任 principal），全球首宗 AI 虛構引文執業限制處分
Read, mark up and supplement source documents — sensitive content never leaves the machine
pseudonymized text →
Pseudonymization engine × human confirm gate
← replies return · restored locally with diff review
NIGHT ZONE
Pseudonyms flow to models
Cloud LLMs only ever see pseudonymized text; BYOK — no cloud lock-in
Never leavesSensitive identifiers stay on the device
BYOKBring your own keys — no cloud lock-in
AuditableEnd-to-end evidence chain
Three-layer compliance framing: ① direct identifiers never leave the device — an architecture-level guarantee [F, git-verifiable] · ② pseudonymized content leaves only with explicit user authorization, with a local small-model "zero-egress" fallback · ③ legal characterization by a written Hong Kong counsel opinion. Desktop-first, no web client — Windows first-class from day one, macOS in parallel.
解決方案 SOLUTION
先緘默，後智能。
紙面區 PAPER ZONE
明文，只在本機
閱讀、標記及補充原文——所有敏感內容不離開設備
假名文本 →
假名化引擎 × 人手覆核閘
← 回覆回流 · 本地還原及比對
暗夜區 NIGHT ZONE
假名，流向模型
雲端 LLM 只見假名文本；BYOK——不被任何雲綁定
不出本機敏感標識符 永不離開設備
BYOK用戶自攜模型密鑰 免受雲端綁定
可審計全鏈路證據鏈 合規可自證
合規三層精確表述：① 直接標識符不出本機——架構層面保證 [F，git 可核證] · ② 假名化案情跨境傳輸須經用戶明示授權，並設本地小模型「零外傳」降級路徑 · ③ 法律定性由香港執業大律師意見書界定。桌面優先、不設網頁版——首發即完整支援 Windows，macOS 同步推出。
TACITA · PITCH DECK04 / 18
THE PRODUCT
Safety you can see.
Local pseudonymization core — a five-layer hybrid detection pipeline, fully offline
Judgment returned to the user — manual supplementation, de-selection, cross-file sync; nothing ships until confirmed
Two-zone preview — cleartext left, pseudonyms right: see exactly what the model will see
產品 PRODUCT
安全，清晰可見。
本地假名化引擎——五層混合式識別流程，可完全離線運作
裁量權交還用戶——人手補充、取消選取、跨文件同步，確認後方可出站
雙區預覽——左為原文、右為假名，出站前清楚顯示模型實際所見
TACITA · PITCH DECK05 / 18
THE PRODUCT
Interface as compliance statement.
Review workbench — 15 entity types highlighted by type, grouped entity list with counts, audit preview always on
Human review gate — no 100% algorithm-coverage promise; the final call is the user's; review decisions are audit-logged
Real product screenshots, synthetic demo documents — every redaction, restoration and export lands in an append-only hash-chained log (proof-lite audit)
High-risk entity recall — the compliance-critical metric for a redaction product
1500test cases all green
139merges, zero rollbacks
8-gateCI enforcement
≈70%to sellable-scope feature complete
100% publiccorpus, per-document ledgers
Voluntary disclosure: 0.9316 is engineer self-scored — external statements follow the 15-advisor blind review (not yet unblinded); a public blind test vs Microsoft Presidio is committed; "saleable quality bar" is defined by the market, not by us.
驗證 × 引擎
引擎質素走可核證路線。
通用基座——同一語料、同一評分準則
0.2778
凍結基準——第三方可重複驗證
0.9316
現役引擎
0.9658
候選版本——逐代登記
0.9701
00.250.500.751.00
高風險實體召回率——去識別化產品的合規關鍵指標
1500測試用例全綠
139次合併零回退
八重CI 強制關卡
約七成可銷售口徑研發進度
全公開來源語料 · 逐份合規清冊
主動披露：0.9316 為工程師自評數據——對外質素基準須以 15 位法律顧問盲評結果為準（尚未解盲）；並已承諾與 Microsoft Presidio 進行公開盲測；「可銷售質素線」的定義權在市場，而非我方。
TACITA · PITCH DECK07 / 18
WHY NOW
Three lines hardening in one window.
1
Regulation line — rules becoming procedural. Technical-safeguard duties move from outcome liability to process liability; process logging is the evidence form of that era — Tacita's product shape.
ABA Model Rule 1.6(c) · IESBA technology revisions (2024-12-15) · SRA warning (2026-08) · Singapore Courts / MinLaw guidance (2024-10 / 2026-03) [F]
2
Case-law line. Sanctions, warnings and insurance exclusions landed densely across 2023–2026 — confidentiality duty + professional responsibility + OCG + PI exclusions already form a de facto "quasi-export-control", independent of PDPO §33.
3
Supply line. Top players are all cloud; Harvey / Legora landed Singapore in 2026 — educating the market while structurally unable to enter the 399–2,999 USD band. ≈18-month category mindshare window [A].
The window in one line — regulatory pressure × individual adoption, converging in the same 24 months.
時機 WHY NOW
三條主線同步成形。
1
監管主線（規則成文化）。技術保障義務正由「結果責任」走向「過程責任」；過程責任時代的證據形態正是過程紀錄，亦即 Tacita 的產品形態。
ABA Model Rule 1.6(c) · IESBA 技術修訂（2024-12-15 全球生效）· SRA 警示（2026-08）· 新加坡法院 / MinLaw 指引（2024-10 / 2026-03）[F]
2
判例主線。處分、警示及保險除外條款於 2023–2026 年密集落實——保密義務、專業責任、OCG 及 PI 除外條款已構成事實上的「準跨境傳輸管制」，且不依賴 PDPO §33 生效。
FY3 ARR-anchored (≈30% of SAM) · static floor ≈0.46M USD
≈4.68M USD Singapore (= HK$36.47M)
1,107 firms / 6,512 practising lawyers [F]; PSG channel subsidizes up to 70%; the world's densest AI-governance guidance — the committed dual-entity springboard
1.4–4.6B USD legal-AI base · five calibers
Bottom-up cross-check: ~1.92–2.10M lawyers × 5–10% paid penetration × 800–2,000 USD ARPU ≈ 0.77–4.2B USD/yr [A] — Hong Kong T1 is 0.13–0.78% of the global pool
Hong Kong→Singapore→UK / AU→NA / EU
All amounts in USD; HKD converted at the 7.8 peg — case fines quoted in original currency.
Founding 100 — first 100 personal users lock at 299 USD/yr for two years
100% personal→Team credit — members' paid fees credit fully against Team's first year (capped at tier price)
15-run refund + demand gate — refund reasons feed product iteration and the WTP evidence base
Price coordinates: Harvey ≈1,200 USD/seat/mo (leaked, flagged) — Team entry 2,999 USD ≈ 2.5 months of one Harvey seat; Legora ≈3,000 USD/user/yr, min ACV ≈30,000 USD; Purview 12 USD/user/mo but needs an E3 base. The whole band sits "10–40× cheaper than cloud legal AI, cheaper than DLP, on par with general Copilot."
Practitioners feel the gain first; content + alumni network + front-office channels cold-start
02
Multi-seat use inside firms
"N lawyers at your firm already use it" — payment records become evidence to management
03
Firm-wide deployment (Team)
Personal fees credit 100% against year one — conversion friction approaches zero
04
Replication across firms
Once density forms, a pincer movement on institutions from below and above
Split: personal growth led by the partner (HKU alumni network + law-firm front-office channel), founder supports with content & product; institutional negotiation levers the audit evidence chain + insurer report — compliance material is the sales material.
Luminance the only on-prem exception — enterprise doc review, not a workbench
Top ten legal-AI vendors are all cloud-only [F]; Harvey (15.5B USD) and Legora (5.5B USD) both opened Singapore in 2026 — capital consensus that "Asia is the next battlefield," yet nobody serves a US$399/yr solo lawyer: minimum ACVs start at ~US$30K.
Engine layer is a replaceable interface with quarterly benchmark tournaments — local small models (Ollama ~176K stars, 7B at ~100 tok/s [F]) already clear the NER bar: model progress is a supplier-competition dividend, not a threat.
Differentiation = Hong Kong legal semantics + workbench + audit + BYOK — with a committed public blind test vs Microsoft Presidio.
Entity taxonomy, document formats, mixed-script rules, decision rubrics — freeze-anchor 0.9316 vs general baseline 0.2778 (engineer self-assessed; 15-advisor blind review underway); hard for foundation models or open source to shortcut
Verified
Engineering delivery discipline
Frozen baselines, eight-gate CI, 139 merges zero reverts — git-auditable; proves delivery capability, not a structural barrier — marked as such
NOW15 legal advisors in blind engine review · first-round firm field tests on record · 0 paying, 0 signed LOIs (honest baseline)
TARGETS · 6 MO POST-CLOSE
10paying pilots
≥50weekly-active lawyers
3–5paid LOIs
PERSONAL LEG · 399 USD
≈5.5×
LTV ÷ CAC = lifetime gross profit per acquisition dollar [A] · GM basis 62%
CAC · self-serve PLG≈64–192 USD
Annual churn · personal range30–40%
CAC payback≈6 months
Speed, density, institutional pipeline map
FIRM LEG · TEAM 5,999 USD
≈20–27×
LTV ÷ CAC = lifetime gross profit per acquisition dollar [A] · GM basis 70%
CAC · density-harvest talks≈1,025 USD
Annual churn · institutional range15–20%
CAC payback≈3 months
Revenue body, retention, audit assets
PQL: ≥3 personal users per firm triggers the institutional conversation ("N lawyers at your firm already use it — everything they paid is credited to you"); PQL conversion 25–30% vs MQL 5–10% [F/flagged] · year-end paying accounts 114 / 302 / 684 · pre-revenue unit economics are hypotheses, not results [A] — FY1 cohort's 12-month retention and NRR gate the next round · "we see density, not content"
Why this team — 15 legal advisors + chief hardware & software advisors.
Zhen Cao
FOUNDER / CEO
California College of the Arts, Industrial Design (2021); consumer-electronics design & strategy consulting — DJI, Giant Network, Baidu, Vanke (ESG), Deloitte, Fangda Partners, JunHe, SkyGuard. Leads product definition, technical architecture and the R&D system.
Edwin Yau
PARTNER · FUNDRAISING & EXTERNAL
University of Hong Kong, Environmental Science (Innovation & Technology minor); former business consultant at a Hong Kong startup incubator; former ESG audit consultant. Owns channels & customer success, fundraising and external affairs.
Yongzhong Peng
CHIEF HARDWARE ADVISOR
Cisco Principal Technical Marketing Engineer / Senior People Manager (San Jose); University of Wisconsin–Madison.
Zidong Zhang
CHIEF SOFTWARE ADVISOR
Software Development Engineer at Amazon (Seattle, full-time since Apr 2020); University of Wisconsin–Madison. Full-stack & large-scale distributed systems — Java / C# / TypeScript / React / Python / AWS.
Governance finalized: co-founder agreement, 4-year vesting + 1-year cliff, good/bad leaver, reserved matters, deadlock mechanics, key-person insurance, dual repo admins with key escrow, monthly 13-week rolling cash reporting — cap table and full terms in the data room. Advisor agreements (15 legal advisors; council of 2–3) countersigned before closing.
團隊
為什麼是這個團隊——15 位法律顧問 + 硬件 / 軟件雙首席。
Zhen Cao
創始人 / CEO
California College of the Arts 工業設計（2021 屆）；消費電子設計與戰略諮詢——DJI、巨人網絡、百度、萬科（ESG）、Deloitte、方達、君合、天空衞士。主導產品定義、技術架構與研發體系。
Cumulative 3-year operating outflow ≈-346K vs -1.24M USD under a traditional headcount model; founding team contributes zero cash (cash spent to date ≈770 USD); BYOK shifts inference-cost inflation to the customer. Gross margin 62% → 70% → 75%.
RUNWAY DISCIPLINE
The pure-angel ≈385,000 USD stress path survives throughout (no grants, no later rounds); pre-close gap narrowed to ≈-24K USD, covered by presales; later rounds are pure expansion fuel.
FY1 = the 12 months from first closing · full cost-to-cash reconciliation is re-computable year by year · scenario: revenue ×0.7 → FY3 ARR ≈526K USD (downgraded expansion plan)
