Hello, this is Hamamoto from TIMEWELL.
Until last spring I knew several companies whose internal rules said "Reviewing contracts with ChatGPT is forbidden because of information leakage risk." Coming into 2026, those same companies have switched to a workflow where "Claude Opus 4.7 produces the first draft, then a human checks it." The reason this changed in just one year is that legal AI moved from being a "promising experiment" to "a tool you lose without."
This series covers Vertical AI implementation patterns sharpened to specific business domains. The first installment is legal. Now that the world's major law firms have rolled out Harvey across their entire organizations and Japan's Big Four firms have finished distributing AI environments to every lawyer, I want to organize what corporate legal departments should think about next.
Four Problems Legal AI Is Solving
Boiled down, the motivations for putting AI into legal work converge on four areas: contract review time, accuracy of IP and patent search, comprehensiveness of compliance handling, and the obsolescence of knowledge management. They look different on the surface, but the root cause is the same. There are too many documents to read, and human eyes alone have hit a limit.
Let's start with contract review load. A mid-sized Japanese company's legal department processes anywhere from a few hundred to a few thousand contracts per year. The mix is broad: NDAs, service agreements, licenses, SaaS terms, and M&A-related contracts. A January 2026 report by LegalOn Technologies shows AI contract review adoption doubled year over year. The most common qualitative result reported by adopting companies is "first-pass time shrank from 30 minutes to 5 minutes." A 25-minute saving per contract sounds modest, but for 1,000 contracts per year that is over 400 hours of slack created.
The IP space is even more dramatic. A 2026 report from renue cites cases where generative AI and LLMs cut patent search time by up to 90%. Tools such as Patentfield, Patsnap Analytics, LexisNexis IP, and Questel now reach a level where they grasp the "intent" and "concept" of an invention through natural language processing rather than simple string matching. Running FTO (Freedom to Operate, the analysis of whether your product infringes others' patents) continuously starting from the requirements definition stage helps you avoid the disaster of hitting a patent wall after development is well underway.
Compliance keeps expanding year by year: the Personal Information Protection Act, the Specified Commercial Transactions Act, the Act against Unjustifiable Premiums and Misleading Representations, antitrust law, export controls, and ESG-related rules. Cross-checking internal documents and guidelines is an area where humans inevitably miss things. Finally, knowledge management. Notes from past matters that are buried in the legal department's SharePoint or Notion are notoriously hard to search, and "just ask the senior who handled that matter" remains the fastest path. AI is going after all four of these at once.
How AI Contract Review Works and How It Is Implemented
The internals of contract review AI all look similar on the surface. The system compares your contract against a playbook (a ruleset that defines your standard clauses and risk tolerance), extracts clauses, scores their risk, and proposes revisions. The differences emerge in how the playbook is built, how accurately clauses are extracted, and how the tool integrates into existing workflows.
Here is a comparison of the four leading tools.
| Tool | Environment | Strengths | Target Users |
|---|---|---|---|
| Harvey AI | Web app + Vault (document management) | M&A, litigation, regulatory work for large firms; Deep Research | Major law firms, mega-corp legal departments |
| LegalOn | Web app + Word integration | 50+ playbooks, Inline Citations, 5 AI agents | Legal departments domestic and overseas, mid-sized firms |
| Spellbook | Runs inside Microsoft Word | 2,300 contract types, Clause Library, Market Comparison | North American startup legal teams, contract attorneys |
| Hubble | Contract management SaaS + Contract Flow Agent | Matter management integration, legal request triage, 82% time reduction | Japanese in-house legal departments |
Harvey adopts a four-tier implementation pattern: Assistant, Vault, Workflows, and Deep Research. Assistant handles individual questions and drafts, Vault performs document review across thousands of files, Workflows automates multi-step sequences such as "contract review to risk report to compliance check," and Deep Research handles agentic multi-step reasoning. Allen & Overy (now A&O Shearman) ran 40,000 queries with 3,500 lawyers across 43 offices and operated multilingually on this stack.
LegalOn shipped a major update in spring 2026. The Inline Citations feature shows the source directly within the AI's answer. When the AI says "this clause is risky," you can trace its reasoning back to the relevant section of the playbook. In legal work where hallucinations can be fatal, verifiability is the most important feature of all. The ability to bring up to 20 files into the same conversation is also significant in practice; you can use it directly to compare multiple SPA drafts during M&A due diligence.
Japan's Hubble takes its own path. The Contract Flow Agent does more than review contracts; it acts as a triage layer that decides whether a request from the business unit is "true legal work" or "something a template can handle." This works because Japanese legal departments suffer from "legal requests being thrown over the wall and overwhelming the team." Just having AI sort traffic upstream lets legal staff focus on high-risk matters where they should really be spending their time. I think this design—where review AI plays the role of "triage agent in front of the lawyer" rather than "lawyer replacement"—is fundamentally sound.
Looking for AI training and consulting?
Learn about WARP training programs and consulting services in our materials.
IP and Patent AI Implementation
AI in the patent space is one step ahead of contract review. The reason is simple: patent databases have long existed as huge structured corpora, which is a great fit for machine learning.
There are roughly three implementation patterns: prior art search, citation network analysis, and specification drafting. In prior art search, you submit the claims of a planned invention in natural language, and the system pulls up similar inventions from global patent databases. Patentfield covers Japan, the US, China, Europe, and Korea, and Patsnap Analytics visualizes citation relationships between patents. Citation network analysis scores how often a patent is cited by others to surface "strong patents" or "foundational patents." When a business unit is entering a new domain and needs to read its patent landscape, this lets you see "the landmines you should not step on."
Specification drafting is still developing, but products like Questel's "Patent Mapping and Claim Analysis" have entered practical territory by automatically mapping product specifications to competitor patent claims. They can score which claim elements your product might fall under. There are case studies in the security industry where this alone shaved hundreds of thousands of yen off the initial estimate when commissioning outside patent counsel.
Japan's Patent Office under METI continues to publish surveys on AI-related patent applications. Since 2024, AI-related patent filings have surged worldwide, and the risk that your own AI implementation infringes another company's patent has grown year by year. Whether or not you can run FTO from the earliest stages of development is becoming a life-or-death issue for manufacturers and SaaS companies. Many Japanese IP departments are still just starting to use Patentfield or LexisNexis IP, but given the speed at which generative AI is being deployed, "using these tools" is likely to become the baseline within 2026.
Case Studies from Major Law Firms
The world's major law firms are past the phase of "how to introduce AI" and are now competing on "how to turn the AI we introduced into a differentiator."
A&O Shearman (former Allen & Overy) was Harvey's exclusive launch partner in 2023. In 2025 they announced agentic AI agents specialized for antitrust filings, cybersecurity, fund formation, and loan reviews. These agents are designed not only for internal use but also to be sold to clients and other law firms. Law firms are starting to pivot from buyers to sellers of AI products. That is a structural shift.
In August 2025, Latham & Watkins announced that Harvey would be deployed to all 3,600 of its lawyers. The fact that the second-largest firm in the US, with revenues over $7 billion, made this move at once is significant. At the same time they launched an "AI Academy," a structured training program. Beyond simply distributing tools, they are investing in the machinery to mass-produce "lawyers who can wield them." Norton Rose Fulbright is also strengthening AI regulation work as a practice area, led by Al Hounsell, Director of Strategic Innovation and Legal Design.
Japan's Big Four firms (Nishimura & Asahi, Mori Hamada & Matsumoto, Nagashima Ohno & Tsunematsu, and Anderson Mori & Tomotsune) all completed firmwide AI provisioning between 2024 and 2025. Nishimura & Asahi positions 2025 as its "AI Year One" and adopts a "best-of-breed" strategy that avoids dependence on any single vendor. They are accumulating results in information retrieval, document summarization, and drafting client-facing reports. At Anderson Mori & Tomotsune, CKO (Chief Knowledge Officer) Maki Kadonaga won the Legal Intrapreneur category of the Financial Times' Asia-Pacific Innovative Lawyers Award in 2025, and the firm is also stepping into outside legaltech businesses, advising Bengo4.com's "Legal Brain business." TMI Associates is asserting its presence in AI regulation and cross-border data by organizing the Global Data Protection Trends 2026 seminar.
What I find interesting about Japanese firm trends is that they explicitly pursue a multi-AI strategy rather than going "all in on Harvey" like US/UK peers. Generative AI capability rankings reshuffle every six months, so avoiding vendor lock-in makes sense. That said, juggling multiple tools sharply increases operational cost and governance complexity. Who controls that, and how, will be the next differentiator.
Claude Adoption in Domestic Legal Departments and Startup Legal Teams
The major firms make headlines, but the volume zone is mid-sized companies, small businesses, and startups. Looking at how they implement AI gives you a more realistic picture.
Anthropic's release of Claude Opus 4.7 on April 16, 2026 was a turning point that accelerated legal adoption. Its high instruction-following and tendency not to over-interpret are well-suited to legal work. Anthropic's own legal team uses Claude for contract review and verification of marketing materials, and reports that work that took several days now finishes in hours.
Here are three patterns I commonly see in Japanese startup legal teams.
First, generating first drafts of NDAs and service agreements. They train Claude Projects on their standard templates and history of past edits, and produce an NDA first draft for a new counterparty in 30 seconds. The legal team only checks clause consistency and business context. Second, redlining English-language contracts. As cross-border deals grow, English-language contract review becomes necessary, but many companies lack lawyers strong in English contracts. Handing Claude an English playbook to redline and asking it to summarize risky clauses in Japanese has become standard practice. Third, first-line response to internal compliance inquiries. A bot fields questions like "Is this promotion compliant with the Premiums and Representations Act?" or "Could this social media post violate the Pharmaceuticals Act?" by citing similar past matters and decision criteria. The final call is human, but research time drops to nearly zero.
Spellbook, which runs inside Microsoft Word, holds dominant share among North American startup legal teams. Completing contract review, Clause Library, and Market Comparison without leaving Word is decisive for small teams that do not want more tools. In Japan, both Hubble and LegalOn have strengthened their Word integration, and every vendor is moving in the direction of removing the friction of "switching to a separate browser app." My hypothesis—UI integration matters more than raw AI accuracy in determining adoption rates—has hardened into conviction in 2026.
Risks and Operational Design: Confidentiality, Verification Workflows, and Quality Assurance
I have written a lot of optimistic things, but there are three heavy pitfalls in legal AI: confidentiality, verifiability, and quality assurance.
The confidentiality risk is symbolized by the Samsung case. To boost productivity, engineers entered source code and meeting minutes into consumer ChatGPT, raising concerns that confidential information was used for model training. If a legal department does the same, the discussion expands all the way to attorney-client privilege. As of 2026 the minimum bar is to designate one enterprise service whose contract guarantees inputs are not used for training (Claude Enterprise, ChatGPT Enterprise, Gemini for Workspace, etc.) and to ban work-related inputs to other services at the level of company rules.
Verifiability is two sides of the same coin as hallucination defense. When AI says "this clause is high risk," you cannot trust the result without a way to trace back the reasoning. LegalOn's Inline Citations and Harvey's source-linking features are direct answers to this. If you build your own internal AI, you should incorporate not just vector search but also a knowledge graph (GraphRAG) so that you can trace relationships between clauses, precedents, and guidelines. That makes downstream verification work.
Quality assurance is a workflow design problem. AI reviews, then a human checks at the end. If the "final human check" becomes a formality, AI errors flow straight into production. Harvey's internal benchmarks show that combined Harvey + lawyer review beat lawyer-only review by 5% in accuracy, while AI-only review can be less accurate than human-only in some cases. Using AI as the starting point and having a human responsibly close the loop—a hybrid design—remains the best balance of accuracy and efficiency today.
GraphRAG has rapidly drawn attention in 2026 in the context of legal knowledge management. Contract clauses, case law, internal guidelines, and notes from past matters are structured data connected by relationships such as "citation," "override," and "exception." Vector search alone retrieves only "semantically similar documents," but GraphRAG can follow the relationships. It can answer compound questions like "If we include this clause, what problems have arisen in similar past matters?" with citations. This connects directly to the organizational deployment story I wrote about in installing-ai-agent-into-organization-5-phases and the strategic options in ai-agent-driven-management-3-strategic-options.
TIMEWELL's Legal AI Support
At TIMEWELL we support enterprise legal AI deployment along two axes.
WARP is our AI consulting service and walks alongside you from selecting a contract review AI through playbook design and operational governance. Whether Harvey, LegalOn, Hubble, or Spellbook fits depends on contract volume, matter complexity, existing workflows, and the share of English-language contracts. WARP NEXT bundles monthly research and implementation support to prevent AI selection mistakes and over-investment. Increasingly, hiring a monthly specialist consultant is more realistic than recruiting a legal DX leader full time.
ZEROCK is our enterprise AI foundation, offering GraphRAG running on AWS servers in Japan. You can ingest legal knowledge, internal guidelines, notes from past matters, and case-law notes in structured form, then build an internal AI that answers with citations. It is well-suited for companies that want to keep highly confidential legal data inside their own boundaries rather than handing it off to outside SaaS. The design also makes a future where you externalize your own AI assets—similar to Anderson Mori & Tomotsune's move into the Legal Brain business—realistic.
For a recent reference, I have summarized enterprise AI agent trends from Google Cloud Next 2025 in google-cloud-next-2025-ai-agents-enterprise. Rather than dropping in legal AI as a one-off, I recommend positioning it inside your overall organizational AI agent strategy.
Legal AI is no longer a "promising experiment." The gap between legal departments that wield Harvey, LegalOn, Hubble, and Claude well and those that do not produces 2x to 3x differences in both annual throughput and risk detection accuracy. The question for executives has shifted from "do we use it" to "which stack, controlled how, combined with which people." The first step is to inventory your contract volume and playbook readiness and decide where to start. If you need a partner to think this through together, please reach out.
References
[^1]: Harvey | AI platform for legal and professional services [^2]: What's New in LegalOn: Spring 2026 [^3]: A&O Shearman and Harvey to roll out agentic AI agents targeting complex legal workflows [^4]: Latham Announces Firmwide Deployment of Harvey [^5]: Hubble Contract Flow Agent [^6]: Lawyers Guide 2026 Nishimura & Asahi [^7]: Best Legal AI Contract Review Software 2026 | Spellbook [^8]: Legal Knowledge Management Trends for 2026 and Beyond [^9]: Anthropic Legal Team Boosts Productivity with Claude [^10]: What is Patent AI? Mechanisms of AI Patent Search and IP DX 2026 Edition | renue