Sakana AI shipped a Japanese-language legal question-answering model aimed at corporate law teams, with deployment packages that keep prompts and retrieved statutes inside customer data centers. The release lands as listed companies review outsourcing rules following a wave of hosting-provider incidents that exposed partner credentials across travel and rail platforms.
Product shape
The model is not a chatbot for consumers. Sakana tuned it on publicly available statutes, ministry ordinances, and anonymized contract clauses contributed under license by two Tokyo law firms. Answers include paragraph citations—not bare summaries—and the system refuses to opine on live litigation dockets unless users upload redacted briefs under their own retention policy.
Hardware bundles range from a single-GPU tower for mid-cap issuers to a rack-scale appliance for banks that already run air-gapped analytics. Sakana said latency targets stay under three seconds for typical compliance queries, fast enough for meeting-room use but slow enough to discourage casual employee browsing.
Why in-house teams care now
Japan’s Financial Services Agency has pressed regional banks on third-party AI use, and general counsel offices face the same question: can we use generative tools without sending privileged drafts to U.S. clouds? Sakana’s pitch mirrors NEC’s on-premise retrieval appliances for lenders, but focuses on corporate law workflows—antitrust filings, labor rule changes, and disclosure checklists ahead of quarterly earnings.
Recent hosting breaches reminded boards that vendor risk is not only about credit card numbers. Legal departments worry about M&A teasers and personnel investigations leaking through poorly scoped SaaS trials. An on-prem bundle trades convenience for control.
Limits and hallucination controls
Sakana documents a two-step retrieval pass: first fetch candidate statutes, then generate an answer constrained to quoted spans. Human reviewers still sign outbound filings; the tool is positioned as a drafting assistant, not outside counsel. The company publishes benchmark scores on ministry-style multiple-choice exams, but warns that novel cross-border tax questions remain high-risk without specialist review.
Competing global models offer Japanese fluency, yet their default cloud terms conflict with internal policies at several insurers that tested them in sandbox environments this summer. Sakana’s differentiation is contractual and architectural—data stays inside the firewall.
Adoption outlook
Early customers named in the rollout deck include a pharmaceutical issuer and a logistics REIT, both experimenting on employee handbook updates and supplier contract renewals. Pricing is per-seat with hardware lease options, cheaper than retaining extra outside counsel hours for routine statute lookups but not trivial for subsidiaries that only need a handful of queries monthly.
For investors, Sakana remains a private company; the release is a product marker in Japan’s crowded enterprise-AI field. For general counsel, it is a test of whether localized models can earn a permanent desk license without repeating the cloud-security debates that slowed public-sector chatbot rollouts.