Systems for matter files, contracts and precedent.

Start with the matter file

A law firm organizes work by matter, not by document and not by user. A matter has a client, a responsible partner, a team that changes as work proceeds, a conflicts position and a retention date, and nearly every question a fee earner asks is scoped to one. Outside law the same shape is an engagement or a job code.

The matter model comes first: how matter numbers map onto folders and document IDs, and who is on each team today. That record is split across a document management system such as iManage, NetDocuments or SharePoint, a practice management system, and email filed to the matter, which may hold the operative agreement. Each is a separate integration, and the metadata decides whether the rest is possible: document class, matter number, version, execution status and source security travel into the index with every passage.

What the implementation can include

A connector for the agreed document management and practice management systems, entitlement and barrier filtering applied at retrieval, pinpoint citations, a clause schema for the agreed document types, and version handling. Hybrid retrieval, reranking and clause comparison are options where the corpus and the evaluation set justify them. What a given build contains is settled in discovery; nothing here demonstrates a system already running in a firm.

Matter-based access and information barriers

Access here is not a role. Two partners of equal seniority can hold different entitlements because they are on different matters. How strictly that holds varies: some document estates are open by default, with security applied only to confidential and walled matters, while others are need-to-know throughout, and team membership may or may not be withdrawn when a fee earner comes off a matter. Which model the firm runs is established before the retrieval filter is designed.

Barriers arise where a firm holds confidential information that must not reach another team, after a lateral hire, a former-client conflict, or concurrent matters with adverse interests. For named individuals the barrier is absolute and has to hold on search, summarization, extraction, suggested precedent, autocomplete and anything cached. Entitlements are read live at query time, but a live filter does not unmake the copy indexing already made in an embedding store and any cache, outside the document management system's security boundary. Which matters are excluded from indexing, and how a new barrier clears documents from the index, is decided before ingestion.

Citations to the clause, and missing evidence

A link to the whole agreement lets a fee earner verify nothing, so a citation is pinpointed to the document, the version and the clause or page. That needs anchors captured at ingestion that survive back into the source document: page and region for a PDF, paragraph identifiers for a word processor file, and a document and version link into the document management system, alongside a parser that preserves clause numbering. Which anchors exist differs by document class.

Where retrieved material does not support an answer, the system states that the file does not contain it and offers what it did find rather than answering from general model knowledge. A citation that does not resolve, or that misdescribes what it points at, carries consequences the firm's own risk function defines, so the support threshold is explicit and refusal cases sit in the evaluation set.

Versions, amendments and what counts as current

A contract is rarely one document. There is a negotiated draft set, an execution version, amendments and side letters replacing particular clauses, perhaps a consolidated as-amended copy prepared by hand, and the same clause text appears across all of them. A system that treats them as equally valid passages can return a clause the parties have since replaced.

Version state therefore belongs in the index as data: execution status, effective date, what supersedes what, and whether a document is the operative instrument or a draft. Where that chain is not recorded anywhere, the system states which version it treated as operative rather than choosing silently.

Clause extraction, comparison and drafting

Clause extraction turns a document into typed fields: parties, governing law, termination rights, liability caps, change of control, indemnities. The output is a value, a confidence signal and a span pointing at the text it came from, so a reviewer corrects it against the source. A lease portfolio and a loan book need different schemas, agreed before the extractor is built.

Comparison is a different operation. A lawyer comparing two versions wants semantic difference rather than a character diff: which obligations changed, which were deleted, whether a cap moved, whether a carve-out appeared. That needs clause-level alignment between documents whose numbering differs, and it is the same machinery that checks a draft against a firm playbook. A draft is assembled from retrieved precedent rather than from the model's own knowledge, so each inserted clause carries the precedent and matter it came from. Ingestion, chunking, hybrid retrieval and reranking are the general engineering described under RAG chatbot development.

Assistance, not legal advice

The system assists with retrieval and drafting. It does not give legal advice. That is a design constraint rather than a disclaimer: scope is fixed in the retrieval sources and the tool contracts, so questions turning on the merits, on strategy, or on whether a clause is enforceable are routed to a fee earner. Where an interface reaches clients rather than firm staff, the line is tighter again.

A fee earner stays responsible for the work product, so drafts and extractions arrive reviewable, who accepted what is recorded, and the system files, sends and executes nothing. Conflicts checking is a separate problem, turning on party names, corporate groups and entity resolution across spellings and subsidiaries, and the clearance decision stays with the people whose professional obligation it is.

Privilege, retention and the client's own rules

Matter files hold privileged material, and a retrieval system moves it: into an embedding, into a prompt, into a provider's logs, into a cache, into a trace. Which provider sees what, in which region, under which retention and training terms, whether prompts are logged at all, and what the audit record shows, are settled in the architecture rather than inherited from a provider default.

Where outside counsel guidelines restrict which tools may touch a client's material, restrict where it may be processed, or reserve audit rights, those terms constrain the architecture, so the firm brings the relevant guidelines into the sprint. Matters also close under destruction schedules, so deletion propagates to embeddings and derived extractions. For firms in India this sits alongside applicable data protection and professional conduct rules, which we scope with the firm and its own advisers rather than advise on.

Evaluate against the firm's own matters

The evaluation set is drawn from the firm's own know-how requests and the questions partners put to juniors, and it has to include the cases only a matter corpus produces: one clause across several drafts of a negotiation where only one was executed, a question whose answer is in filed email and not in the document management system, a question answerable only from a matter the asking user is not on, and a barrier raised after indexing.

Each entry records the documents that should have been retrieved and the version that should have been treated as operative, so a citation failure is visible separately from a drafting failure. Access cases run as tests on every change to a connector, a filter or the index, and clause extraction is scored against a sample of the firm's own precedent reviewed by someone who signs off on that document type.

What a legal AI project costs

Every engagement starts with a fixed-price discovery sprint at USD 2,500 over two weeks, which produces the architecture, initial screen designs, acceptance criteria, a running-cost model and a fixed price for the build. The sprint fee is deducted from the build. For this work it also settles the matter and permission model, the connectors the firm will grant, the clause schema and the citation anchor per document class.

Indexing is charged per document version, and one agreement generates a version for each negotiated draft, the execution copy and every amendment, while a matter is re-indexed when a barrier or a team changes and not only when a document changes. Per-request inference rises with the passages a pinpoint citation forces into each request, and comparing two long agreements is a larger unit of cost than a question. Load may be uneven, since filing deadlines, renewals and closing periods can concentrate activity, so the firm's calendar sets the peaks the model is built on.

Part of the AI Feature Implementation Sprint → · Explore our work → · Discuss an AI project →