AI for legal teams: three deployments that work and three that don't

Legal AI gets sold harder than almost any other professional services category, and gets adopted more carefully than almost any other for the same reason: a hallucinated citation in a court bundle costs more than a misclassified email. The firms that succeed treat AI as deployment risk first and productivity second, which is exactly the inversion of how it’s pitched.

This is what we tell partners and operations leads on the first call. Three applications are worth the engagement. The fourth, fifth, and sixth on most pitch decks shouldn’t be near a production system yet.

1. Retrieval over your own files

This is the application with the clearest payback and the most predictable build. Public databases (Westlaw, LexisNexis) cover published case law, but your firm’s value is what you’ve done: drafted, negotiated, reasoned about. That work sits in a mix of PDFs, scanned bundles, Word documents, and partner inboxes, and is effectively unsearchable beyond keyword and folder structure.

A retrieval system over that corpus surfaces the right precedent in seconds, with the source paragraph attached. Not a generative answer, a retrieval result. The partner reads the cited passage, decides whether it’s actually on point, and proceeds.

The legal case study on our work page is a worked example. A UK commercial litigation firm with 2.3 million pages across fifteen years of case files. Eight weeks, on-premises deployment, 73% reduction in research time on partners’ own logs. The system cites every result back to the source document. No hallucinated citations because there’s no generative step doing the citing.

What makes this work. Three things, in order. First, paragraph-level (or clause-level) chunking. Partners don’t want documents, they want passages. Second, hybrid retrieval: semantic similarity plus metadata filters on practice area, year, document type. Third, a citation-first interface, where every result is a clickable link back to the source and “we couldn’t find it” is an acceptable answer.

Where it goes wrong. Document-level chunking, where the system retrieves a whole 80-page judgment instead of the relevant paragraph. Generative summaries layered in before partners trust the retrieval: the summaries hallucinate, partner trust collapses, the project fails. Skipping the OCR-and-cleanup phase on scanned material, because the system can only retrieve what it can read.

2. Contract drafting and review assistance

Standard agreements (NDAs, supplier terms, engagement letters, employment contracts, basic licensing) get drafted from templates that drift. Different partners use different versions. Junior associates copy from the most recent matter they remember, which may or may not reflect the firm’s current position. The non-bespoke work gets done badly while the bespoke work gets the senior time.

An assistive system improves both ends. The associate selects the template, answers a few structured questions, and gets a first draft with the firm’s preferred positions baked in. The system flags unusual clauses, missing sections, and diffs against your preferred language. The partner reviews drafts rather than writing from scratch.

Typical numbers from this kind of build: four hours to forty-five minutes for standard agreements. The partner doesn’t lose oversight. They review more drafts in less time. Senior capacity gets freed for the genuinely bespoke work, which is where the firm’s margin actually sits.

What makes this work. A clear taxonomy of what counts as “standard” and what doesn’t. Most firms underestimate how many agreements are genuinely bespoke. A library of preferred positions that’s actively maintained, not a one-time export. And a strict review step before anything goes to a client: the system drafts, the lawyer signs.

Where it goes wrong. Treating the AI draft as final, particularly under junior pressure. Templates that nobody owns: they drift, the system gets retrained on drifted templates, and the firm slowly loses control of its own positions. Trying to extend the system to negotiated terms, where AI can’t read the client’s room or hold a line under pressure.

3. Regulatory and compliance monitoring

FCA, SRA, ICO, and sector-specific regulatory updates land daily. The standard approach, a compliance officer reading a weekly digest and flagging anything relevant, burns most of a person’s week and still misses things, because the digest doesn’t know which of your clients or matters the update affects.

A monitoring system ingests the same updates, classifies them against your client base and matter types, and produces a daily two-minute review queue rather than a weekly half-day. It doesn’t replace the compliance officer; it makes them the reviewer of pre-sorted intelligence rather than the first reader of unsorted output.

This is the application where the audit-trail story matters most. The SRA or your insurer will ask, at some point, how you decided that update X applied (or didn’t apply) to matter Y. Every classification logged, every retrieved passage stored, every decision reviewable. Those are not features you bolt on at the end. They’re the architecture.

What makes this work. A defined client and matter taxonomy that the classifier can map to. Most firms don’t have this written down. Daily-cadence rather than batch processing, so the firm catches urgent changes the same day. And a clear escalation path for the cases the system isn’t sure about.

Where it goes wrong. Setting too high a confidence threshold and missing material updates. Setting too low a threshold and burying the reviewer in noise. Letting the system auto-action anything, when every output should be reviewed before it becomes a firm position.

What to ignore for now

Three categories of legal AI get pitched aggressively and shouldn’t be near a production deployment in 2026 unless you have unusually deep tolerance for risk.

Legal advice generation. Systems that draft client-facing advice end-to-end. The deployment risk is asymmetric: the upside is faster drafting, the downside is professional indemnity exposure on hallucinated advice. The economics don’t work for most firms.

Outcome prediction. “What are the odds we win this case?” systems. The underlying models are unreliable on novel matters, which is the only kind of matter where the prediction would actually be useful. Read the marketing carefully: the claimed accuracy almost always sits in a narrow band of routine cases where the partner already knows the answer.

Fully automated contract negotiation. AI back-and-forth on terms with a counterparty’s system. The technology is there in toy demos. The business case is not. The negotiated terms are where the firm’s actual judgment lives, and clients pay for that judgment.

The constraint hierarchy

For everything you do deploy, the constraints aren’t optional. We design every legal AI engagement around four of them:

Citation is non-negotiable. No retrieved fact without a clickable source. No generative answer without the underlying passages. Anything else is professional risk.

Privilege stays inside the firm. Self-hosted embeddings, vector store on your infrastructure or private cloud, nothing privileged leaves the network. UK data residency by default.

Hallucinations are deployment risk, not a UX problem. Retrieval-first architecture means the model surfaces what exists rather than inventing what doesn’t. Confidence thresholds and “we don’t know” responses are part of the design.

Audit trail for regulators. Every query, retrieved passage, and model response is logged and reviewable. If the SRA or insurer asks how the system reached a conclusion, the answer is in the logs.

Our legal sector page covers all four in more depth, and the legal case study shows what it looks like in delivery.

If you are earlier in the process and want to check whether your firm is ready, the AI readiness checklist is fourteen questions that surface the gaps most legal teams miss before committing to a build.

What to do this week

If you’re trying to work out where to start, pick one of the three applications above and measure the current cost. How many hours a week go into precedent search across the firm? How long does a standard NDA take from request to signed copy? How often does the compliance officer miss something material?

Those measurements are what the diagnostic starts from. Two weeks, fixed price, written report. If the numbers don’t justify a build, we tell you and refund the second week. The legal applications where AI works are narrow and clear, but where they work, the payback is fast and the technical risk is manageable.

Talk to us about your firm.