Use Cases
AI Knowledge Management: Turning Precedents into a Firm Brain
AI knowledge management for law firms has four layers: a DMS connection that copies nothing, curation, measured retrieval and permissions. With a 90-day plan.
On this page
- The myth, and the versions that actually work
- Layer one: connect the DMS without copying it
- Layer two: curation is the product
- Layer three: retrieval, and why it fails
- Layer four: permissions, ethical walls and audit
- Measuring answer quality before anyone relies on it
- Buy, assemble or build
- The 90-day knowledge management plan
A "firm brain" is a retrieval system, and its answers are only as good as what was curated into it. Clifford Chance reclassified and summarised 400,000+ documents before its Knowledge Bank could be trusted with a question; firms that skip that step get what one practitioner study calls "plausible but partial answers". The working design has four layers: a DMS connection, curation, measured retrieval and permissions that survive the agent.
The myth, and the versions that actually work
The myth is that you point a frontier model at the document management system and ask it anything. The versions that work look different. Linklaters' in-house assistant Laila is used more than 200,000 times a month (Legal IT Insider). Addleshaw Goddard's AGPT reached 95% firm-wide adoption at roughly 2,000 prompts a day before the firm added Legora (Addleshaw Goddard, March 2025). Clifford Chance's Knowledge Bank, built with Microsoft and Epiq and announced on 20 July 2026, involved reclassifying and summarising 400,000+ documents (Artificial Lawyer). Harvey's Knowledge and Vault products gained Memory, matter-aware Spaces and context-inheriting agents in Harvey II on 18 August 2026 (Harvey). Lexis+ with Protégé added a Vault of up to 100,000 documents, Workrooms and customer-held encryption keys on 7 May 2026 (LawSites). DeepJudge, the Zurich search company that raised $41.2m in November 2025, is now a partner plugin in both Astra for Law and Perplexity's Computer for Counsel.
Every one of these has a curation layer and a permission layer. None is "the DMS plus a chatbot".
Two shortcuts do not work either. The first is long context. Claude Fable 5.1, Opus 5 and Sonnet 5 accept a million tokens, which Anthropic puts at about 555,000 words (Anthropic model overview); Fable 5.1 also cut cache-read prices by 75% on 1 September 2026. That is a matter, not a firm. Loading a whole data room into one context window is now cheap and sensible; loading twenty years of precedents is not possible. The second shortcut is fine-tuning a model on precedents. Spellbook's chief executive says fine-tuning "encourages hallucinations" (Non-Billable interview); Latham & Watkins does fine-tune open-weight Nemotron models, on its own Nvidia servers, with around 100 AI staff out of 900 in technology. That is a strategy for a firm with $8.3bn of revenue, and the build-or-buy guide explains why almost nobody else should copy it. Retrieval over a curated corpus, the pattern the RAG glossary entry describes, is the design that gives citations and respects walls.
Layer one: connect the DMS without copying it
The Model Context Protocol turned this from an integration project into a setting. iManage released its MCP Server on 14 May 2026: governed, permission-bound, fully logged access that "respects existing ethical walls", working with Claude, ChatGPT, Copilot, Harvey, Legora and custom agents; iManage cites 32% of firms naming integration complexity as an AI barrier (iManage). NetDocuments' ndMAX MCP runs every call under the user's own credentials, applies ethical walls and matter restrictions automatically, promises that "documents never leave the platform", and is enabled by IT with a single admin toggle (NetDocuments). The MCP specification of 28 July 2026 added mid-call confirmations and hardened OAuth, and the protocol itself has sat with the Linux Foundation's Agentic AI Foundation since December 2025.
Use the vendor's server. A report of 26 June 2026 found 82% of 2,614 public MCP servers vulnerable to path traversal, 43% to command injection and 53% relying on static API keys (MCP security statistics); a self-written connector to the DMS is the last place to learn those lessons. The setup guide walks through the iManage and NetDocuments steps.
What the connection gives you is search and reading under your own rights. What it does not give you is quality. The DMS holds every draft, every superseded version, every counterparty's markup and every clause a partner rejected. Retrieval over that returns all of it.
Layer two: curation is the product
A January 2026 practitioner analysis of why firm RAG projects disappoint concluded that the retriever was rarely the problem; legal knowledge is fragmented, written for different audiences, walled by ethical screens and unevenly updated, and the result is "plausible but partial answers" (Inside Practice). Legal retrieval also cannot tolerate fuzzy matches: a 2023 clause retrieved for a 2026 question is not roughly right, it is wrong.
Curation rules that survive contact with a practice group:
- A gold corpus. Approved precedents, signed off by a named owner per practice group, separate from the working DMS.
- Executed over drafts. The signed version, plus the one negotiated markup that shows what moved.
- Supersession. Every document carries a "replaced by" field; the retriever excludes superseded items by default.
- Audience tags. Client-facing, internal analysis, training material. An alert written for clients should not be retrieved as the firm's internal view.
- Retrieval exclusion. Matters behind walls, and matters whose engagement terms restrict reuse, carry a do-not-retrieve flag independent of DMS permissions.
| Metadata field | Example values | Why the retriever needs it |
|---|---|---|
| Governing law | Austria, Germany, England and Wales, Delaware | Prevents jurisdiction drift |
| Party side | Buyer, seller, lender, borrower, employer | Clauses are drafted for a side |
| Instrument | SPA, NDA, facility agreement, opinion, memo | Narrows the candidate set |
| Status | Executed, negotiated draft, template, superseded | Excludes stale and unapproved text |
| Date and regime | 2026; post-Omnibus; pre-DAC8 | Regulatory changes date precedents |
| Audience | Client, internal, training | Prevents audience mismatch |
| Owner | Named partner or knowledge lawyer | Someone answers for it |
Curation needs people. Baker McKenzie cut up to 10% of its business-services staff in February 2026, including knowledge, research and secretarial roles, citing AI (RollOnFriday), then rolled Legora out globally on 14 May 2026. Whatever the merits of the first decision, a platform is a retrieval engine, and the people who decided what was worth retrieving had just left. The knowledge-base playbook sets out the curation roles for a firm that does not have a KM department.
Layer three: retrieval, and why it fails
| Failure mode | What the lawyer sees | Fix |
|---|---|---|
| Chunking splits a clause from its definitions | A limitation clause without the defined "Loss" | Chunk by clause; attach the definitions section to every chunk |
| Near-duplicates crowd the results | Twenty versions of one SPA, none the executed one | Deduplicate on ingestion; prefer status "executed" |
| Stale precedent | A 2023 data-transfer clause after a 2026 regulatory change | Mandatory date and regime filters |
| Audience mismatch | A client alert retrieved as internal analysis | Audience tag filter |
| Jurisdiction drift | An English-law clause answering an Austrian question | Jurisdiction filter is mandatory, never optional; the Vals research study found every system lost 11 to 14 points on multi-jurisdiction questions |
| No negative answer | The system always finds "something" | A required "no precedent found" path, tested |
| Citation laundering | A real document cited for a proposition it does not contain | Require a verbatim quote and document ID in every answer |
The last row is the one to fear. Stanford's 2024 study found Lexis+ AI hallucinating 17% of the time and Westlaw's AI-Assisted Research 33% (Stanford HAI), on products that already used retrieval; the HAQQ benchmark of June 2026 found 24% of answers across current models cited or misapplied law that did not support the claim. Retrieval reduces fabrication; it does not remove the model's willingness to over-read a source. The rule that every answer quotes its source verbatim, with a document identifier a lawyer can open, is the only control that scales. The hallucinations guide explains the mechanism.
Layer four: permissions, ethical walls and audit
Three rules. Retrieval runs as the user, never as a service account; the iManage and NetDocuments servers enforce this, and Harvey added Intapp-based ethical-wall enforcement on 23 July 2026. The index must never become a side channel: an embedding store that contains walled-off documents and is queried by everyone is a breach with extra steps, which is the strongest argument for querying the DMS in place rather than copying it into a vector database. And every query is logged: Lexis Workrooms offer dual approvals, role-based permissions and audit trails; Anthropic's Compliance API retains Cowork and Claude Code session transcripts for six years by default, though it does not capture MCP server configurations, so keep those under change control yourself.
Test the wall before launch and after every change: a user outside a walled matter asks a question only that matter could answer, and must get nothing. For European firms the vendor sits under an Article 28 GDPR processor agreement, the data stays in the EU where the tier allows it (Gemini Enterprise added EU residency for NotebookLM Enterprise on 3 June 2026), and Germany's BRAK guidance treats any tool that forwards client data for model training as a breach of section 43a BRAO. The GDPR page covers the contract terms.
Measuring answer quality before anyone relies on it
Nobody should trust the system on the strength of a demo. Build a golden set per practice group: fifty real questions with the known best document, five trap questions where the correct answer is "no precedent", and one walled-matter probe.
| Measure | Target before launch | How |
|---|---|---|
| Right document in the top three | 90% | Compare against the known best document |
| Answer cites it with a verbatim quote | 95% of answers that found it | Check quote against source |
| Fabricated content | 0 | A reviewer reads every answer |
| Correct "no precedent" on trap questions | 5 of 5 | Any invented precedent fails the launch |
| Wall probe | 0 leaks | Run from a walled user's account |
| Time to a usable answer | Under 2 minutes | Stopwatch, including reading the source |
Closed-set question answering works when the set is right: the Vals Legal AI Report of February 2025 scored Harvey at 94.8% on document Q&A against a lawyer baseline of 70.1% (Vals). The scores collapse when the set is wrong. Re-run the golden set monthly, after every corpus change, and publish the numbers internally; a firm that cannot show the scores cannot answer a client or an insurer who asks how it knows.
Buy, assemble or build
| Route | What you get | Price signal (September 2026) | Trade-off |
|---|---|---|---|
| Harvey Knowledge and Vault | Curated firm corpus, Memory, Spaces, 500+ agents | Estimated $1,000 to $1,200 per seat per month, 20-seat minimum (secondary sources) | Best-integrated; pricing unpublished |
| Legora | Tabular Review to 100,000 documents, Monitors, Portal | Reportedly about $3,000 per user per year, 10-seat minimum | Strong at review; KM depends on your corpus |
| Lexis+ with Protégé Vault | 100,000-document Vault, Workrooms, BYOK | Bundled with Lexis subscriptions | Tied to Lexis content |
| CoCounsel Legal | Tabular Analysis over 10,000 documents, Workspaces | Published range $104 to $639 per user per month | Westlaw-centred |
| Claude Team or Enterprise plus iManage or NetDocuments MCP | Query the DMS in place under your permissions | Team $25 or $125 per seat; Enterprise custom | You supply curation; GC AI notes the Claude legal plugin resets context per session and lacks team playbooks |
| Gemini Notebook | Closed-universe notebooks, 50 sources free, 300 paid | Workspace add-on | Matter-level, not firm-level |
| Self-built (n8n, vector store, model API) | Full control | Model tokens plus hosting | You own every failure mode above |
The Harvey review and the pricing guide go into the numbers. For most firms below 50 lawyers the honest answer is the fifth row: connect the DMS, curate a small gold corpus, measure, and only then decide whether a platform's curation tools are worth the seat price.
The 90-day knowledge management plan
| Days | Work | Output |
|---|---|---|
| 1 to 30 | Pick one practice group. Appoint an owner. Enable the vendor MCP server on a business tier with training off. Write the golden set. Tag 200 gold documents with the metadata schema. | A connected, permissioned pilot with a test set |
| 31 to 60 | Run the golden set weekly; fix chunking, supersession and filters; add the "no precedent" path; run the wall probe. Train the group on quote-and-open verification. | Scores above the launch targets, or a documented reason why not |
| 61 to 90 | Open to the group. Log every question. Review the twenty worst answers with the owner; add or retire documents. Decide on platform versus assembled stack using the scores, not the demo. | A go or no-go for the second practice group, with numbers |
Module 4 of the AI Lab for Lawyers course covers connecting Claude to a document store and testing retrieval against a golden set, which is the part of this plan most firms have never done. Do it before the demo, not after the rollout.
Frequently asked questions
What is a firm brain in legal AI and does it exist?
The phrase describes a system that answers questions from a firm's own precedents, know-how and matter history. Working versions exist at Linklaters, Addleshaw Goddard and Clifford Chance, but each is a retrieval system over a curated corpus with permissions enforced, not a model that has absorbed the firm. The curation and permission layers are what make the answers usable.
Should a law firm fine-tune a model on its precedents?
Usually not. Fine-tuning changes how a model writes, not what it reliably knows, and Spellbook's chief executive has said it encourages hallucinations. Latham & Watkins fine-tunes open-weight models on its own servers with around 100 AI staff, which is not a path for most firms. Retrieval over a curated corpus gives citations and respects permissions; fine-tuning does neither.
How do iManage and NetDocuments MCP servers work with AI tools?
Both vendors released Model Context Protocol servers in May 2026. Every query runs under the user's own credentials, so ethical walls and matter restrictions apply automatically, calls are logged, and documents stay in the platform. They work with Claude, ChatGPT, Copilot, Harvey and Legora, and IT can switch them off with a single admin setting.
How do you measure whether a legal knowledge system gives good answers?
Build a golden set: fifty real questions per practice group with the known best document for each, plus five trap questions with no precedent and a walled-matter test. Score whether the right document appears in the top three, whether the answer cites it with a quote, whether anything was fabricated and whether the wall held. Re-run monthly and publish the scores internally.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Playbooks
Build a Firm Knowledge Base for AI: Curation, Retrieval and Permissions
A playbook for a law firm knowledge base AI can retrieve from: curating precedents, metadata, permission mapping, platform choice and a 30-question test.
11 min read
Playbooks
Connect Claude to iManage or NetDocuments with MCP: A Setup Guide
Connect Claude to iManage or NetDocuments with MCP in six admin steps, test on a read-only matter, enable audit logging and scope permissions before rollout.
11 min read
Glossary
Retrieval-Augmented Generation (RAG) — Definition for Lawyers
Retrieval-augmented generation (RAG) searches a document set first and drafts from what it finds. Every grounded legal tool uses it; here is why errors still get through.
4 min read
Glossary
Model Context Protocol (MCP) — Definition for Lawyers
Model Context Protocol (MCP) is the open standard that lets Claude, ChatGPT and Copilot read your DMS or docket feed through one connector. What a server exposes and what to check.
4 min read
Tools
Harvey AI in 2026: What It Does, What It Costs and Who Uses It
Harvey AI reviewed as of September 2026. The $15.5bn valuation, Harvey II, Tenet, 500-plus agents, who uses it, estimated seat prices and its own benchmark.
9 min read
Guides
Build or Buy Legal AI? Latham's GPUs, Harvey's Seats and the Firms in Between
Build vs buy legal AI in 2026: rent Harvey or Legora, assemble on Claude or Astra with MCP, or build like Latham on own GPUs. Costs, control, decision tree.
11 min read