# Two crews, measured over a year. **Two crews, measured over a year.** Both studies include the numbers, the controls and the part that did not work. A case study with no friction in it reads as marketing, and converts like marketing. Case studies ## LawCrew document review at Ridgeway & Partners A 41-lawyer Singapore commercial practice, reviewing commercial tenancy and services agreements under Singapore law. Regulated output, permanent full human review, and a managing partner who had already rejected two AI tools. Matters / month 412 Up from 240 with the same team Reviewer minutes / matter 19 From 64 · −70% Cost per matter reviewed S$9.46 Platform and Pack, fully loaded Evaluation pass rate Not recorded Nothing in the catalogue records one **The situation** First-pass review of incoming commercial agreements was handled by four paralegals and a supervising associate. Average 64 minutes per agreement, a two-day queue in busy months, and no consistent record of which sources a reviewer had relied on. The firm had trialled two general-purpose AI tools and stopped both — not because the drafting was poor, but because nobody could answer the question a client eventually asks: on what basis did you conclude that? The partner's condition for a third attempt was explicit. Every clause-level conclusion had to be traceable to a source with a date, and a named person had to appear against every output that left the firm. The crew LawCrew Document Review from the Singapore Legal Practice Pack, configured to Singapore governing law with the firm's own precedent library attached as a Knowledge Pack. Jurisdiction router Evidence Research Pod · 4 Agents Drafting Adversarial Reviewer Regulatory Review Pod · deterministic gates Professional approval The controls - Evidence floor set at 0.78, higher than the Pack default of 0.72, at the firm's insistence. - Autonomy pinned to the **regulated floor** — 100% human review, permanently. No graduation path enabled. - Dual approval on any matter classified high-risk by the Regulatory Review Pod. - Per-matter access control; no client data leaves the Singapore region; nothing is used to train any model. - Audit retention configured at seven years to match the firm's own file retention policy. **Twelve-month trajectory** No evaluation pass rate is recorded for this deployment, at go-live or since. The case counts below say how many cases exist, not how a suite performed on them. - **Evaluation cases at go-live**\ 60, from the Pack - **Evaluation cases at month 12**\ 247 - **Of those, from Correction Records**\ 187 - **Abstention rate**\ 4.1% - **Review rate**\ 100%, by design **Commercials** - **Tier**\ Professional - **Subscription**\ S$2,400 / mo - **SG Legal Practice Pack**\ S$1,500 / mo - **Credits drawn, BYOK**\ 34,700 / mo - **Top-ups purchased**\ None - **Reviewer seats in use**\ 38 of 100 - **Payback on implementation**\ 1.6 months **What did not work** The first eight weeks were worse than the manual process. The firm's precedent library had been maintained as a shared drive with no dates on anything, so the Evidence Research Pod could not establish currency and abstained on roughly one matter in five. Reviewers read that as the tool failing. It was not. It was the tool refusing to guess about which version of a clause was current — which, once the library was dated and re-indexed, turned out to have been the correct behaviour on every single one of those matters. The firm now treats the abstention rate as a knowledge-quality metric rather than a product defect, but that reframing took a difficult month and two sessions with the supervising associate. > "I did not buy a drafting tool. I bought the ability to sit in front of a client, or the Law Society, and show them exactly which authorities a conclusion rested on and which of my people signed it. That is what makes it usable in a practice. The speed is a bonus." > > Sarah Rajah · Managing Partner, Ridgeway & Partners LLP · Singapore ## AquaMind advisory at Selat Aquafarms A barramundi and grouper operation across nine sea-cage sites in the Johor Strait, with 14 farm technicians and one aquaculture specialist covering all of them. Advisories / month 1,340 From 210 specialist-answered Median response time 2.4 min From 5.8 hours Cost per advisory S$1.94 Platform and Pack, fully loaded Abstention rate 11.3% Escalated to the specialist **The situation** A technician who sees a dissolved-oxygen reading drop at 5am needs an answer before the shift supervisor wakes up. The single specialist could answer roughly 210 questions a month within a working day; everything else was resolved by whoever had the most experience on site, which is how a confident wrong answer about water chemistry kills a pond. The operation is not regulated in the way a law firm is, but the failure mode is more expensive and much faster. The design brief was the inverse of Ridgeway's: speed first, but with a hard rule that the system must say when it does not know. The crew AquaMind Advisory from the Aquaculture Operations Pack, with the farm's telemetry feeds connected as read-only Tools and species knowledge packs for barramundi and grouper. Dynamic router Expert advisory panel · 5 Agents Evidence verification Safety gate · deterministic Telemetry cross-check Technical sign-off on escalation The controls - Evidence floor at 0.70, with a hard deterministic safety gate on anything touching dissolved oxygen, salinity, medication or harvest timing. - Autonomy graduated from full review to **extended sampling at 12%** over seven months, on measured accuracy, with automatic reversion armed. - Every abstention escalates to the specialist with the reading that would resolve it named explicitly. - Bilingual output in English and Simplified Chinese; mobile approval; durable runs to survive intermittent connectivity on the water. **Autonomy graduation** | Stage | Review rate | | - | - | | Months 1–3 · full review | 100% | | Months 4–5 · supervised sampling | 25% | | Months 6–7 · reverted on breach | 100% | | Months 8–12 · extended sampling | 12% | No accuracy figure is recorded for these stages. The month-six reversion was automatic and correct: a feed-supplier document set had been replaced without notice and accuracy fell below the bound within four days. Nobody had to notice it. **Commercials** - **Tier**\ Professional - **Subscription**\ S$2,400 / mo - **Aquaculture Operations Pack**\ S$1,000 / mo - **Credits drawn, managed keys**\ 218,400 / mo - **Top-ups purchased**\ None · 87% of allowance - **Specialist hours returned**\ 63 / month - **Ponds lost to chemistry events**\ 0, from 2 the prior year **What did not work** Adoption stalled at three of nine sites for four months. The crew was answering well; the technicians at those sites simply did not trust an answer that arrived on a phone. What changed it was not training — it was showing them the abstentions. Once technicians saw the system decline a question about a medication interaction and name the water test that would settle it, they started using it for everything. The lesson we did not expect: **the abstentions built more trust than the correct answers did.** We now put the abstention log in front of new sites during onboarding, before any of the success metrics. > "The morning it told my technician it could not answer and asked him to run an ammonia test first, I stopped worrying about it. Anything that will admit it does not know at five in the morning is something I can put in front of my crew." > > Lim Wei Sheng · Operations Director, Selat Aquafarms Sdn Bhd · Johor **How we publish these** Every figure comes from the tenant's own Cost & Outcomes records and evaluation history, not from a survey. Both customers reviewed and approved their numbers before publication, and both agreed to include the section that did not go well. We do not publish a case study without measured outcomes and at least one piece of friction. **Run the same arithmetic on your volume** Both of these started as an estimator output and a 45-minute session. Put your own volume and your own manual baseline in, and bring the result to the call. [Open the estimator](/estimator) [Book a demo](/book-demo) --- wGrow Technologies Pte Ltd · Singapore · Customer names and figures published with written consent. Individual results depend on document complexity, knowledge quality and review posture. Canonical page: https://dev.krewos.ai/cases