Legal services

AI systems for law firms and legal teams

In a law firm the product is judgement, and judgement isn't automatable. Everything wrapped around it — the intake, the conflict search, the chasing, the bundle, the time you reconstruct on a Friday afternoon — largely is. What makes legal work different from other document-heavy sectors is privilege: the material can't leave your control, which rules out most of what's currently being sold to firms and is the reason we build to run inside your own perimeter.

  • Privileged material never leaves your perimeter
  • Conflict searches run deterministically, never inferred
  • Every citable output verified by a person before it leaves
  • Fixed fee, agreed before we start

The work that isn't the work

Fee earners are expensive, and a large share of the week goes on tasks that generate no fee and require none of the qualification behind them.

  1. 01

    Matter intake that waits more than it works

    Engagement letter, identity checks, conflict search, source-of-funds evidence, professional clearance to the outgoing firm, authorisations, file opening. Each step is quick. The delay is entirely in the waiting between them — and nobody owns the waiting.

    We compressed exactly this pattern in a neighbouring practice: 3 days → 4 hours

  2. 02

    Chasing clients, opponents and experts, on repeat

    The other side's solicitor, the surveyor, the medical expert, the client who promised the bank statements a fortnight ago. Chasing is unbillable, requires no judgement, and is the single largest cause of a matter sitting still.

  3. 03

    Files that have to be reassembled before they can be used

    Disclosure review, bundle preparation, chronology building, pagination and indexing. The material already exists in the file — the cost is a junior fee earner or a paralegal putting it in order, then doing it again when something lands late.

  4. 04

    Precedent knowledge locked in closed matters and people's heads

    The clause you need has been drafted and negotiated four times before, in matters nobody can name. So it gets drafted a fifth time from a template, or from whoever happens to be in the office.

  5. 05

    Time recording reconstructed from memory

    Friday afternoon, working backwards through a calendar and a sent-items folder, trying to remember what Tuesday was. Under-recording is endemic, and it comes straight off the bottom line.

What we'd build for a law firm

All of these keep the fee earner's judgement exactly where it is. They remove the assembly, the waiting and the searching around it.

01

Client and matter onboarding worker

Orchestrates the whole sequence — engagement letter out and chased, identity documents collected, clearance drafted, authorisations tracked, file opened — with a live view of what's outstanding and on whom.

Our proven build. 3 days → 4 hours.

02

Conflict search and intake gate

Runs the conflict search across your existing matter and party records as an exact-match check, surfaces every candidate hit with the matter it came from, and holds the file closed until someone signs the search off.

Deterministic by design — never an inference.

03

AML and source-of-funds file assembly

Collates identity documents, screening output and source-of-funds evidence into a structured file with a gap checklist, and chases what's missing. Screening runs deterministically against official lists.

Assembles and chases. Your MLRO decides.

04

Disclosure and bundle preparation

De-duplicates, orders and indexes documents into a bundle to your court or house specification, builds the chronology from dates in the documents themselves, and reruns cleanly when late material arrives.

Assembles the bundle. Never decides what's disclosable.

05

Contract, lease and deed abstraction

Reads contracts, leases, deeds and titles, and pulls the terms, dates and obligations that matter into a structured summary against the matter — with anything ambiguous queued for review rather than guessed.

06

Precedent and know-how retrieval

Search and question-answering across your own precedents, past advice and closed matters, with every answer traceable to the document it came from — running on infrastructure you control, so nothing is sent to a third-party model.

The strongest case for on-premise deployment.

07

Chasing agent

Tracks what each client, opponent or third party owes you, escalates the chase in your firm's tone, files what arrives against the matter, and gives the partner a daily list of what's at risk.

08

Time recording assistant

Reads calendar, email and document activity and proposes time entries for the fee earner to confirm and narrate. It proposes; it never books.

09

Correspondence triage and file management

Routes inbound post and email to the right matter, files it, flags anything carrying a deadline or a limitation date, and drafts the routine acknowledgements for a person to send.

Verified client outcome

3 days → 4 hours

New client onboarding at a 22-person accountancy practice — not a law firm, and we'd rather say so than imply otherwise. We include it because the process is structurally identical to matter intake: a fixed sequence of quick steps where almost all the elapsed time is spent waiting between them. That's the pattern the build removes, and it transfers directly.

Read the case study →

Confidentiality, privilege, and who carries the responsibility

In legal work the constraint isn't a form to fill in — it's a professional duty attached personally to a named individual, and a privilege that doesn't survive careless handling. That shapes the architecture, not just the paperwork.

  • Nothing here prohibits AI, and nothing here transfers the responsibility. Regulators permit AI use and regulate by outcome: the solicitor remains personally accountable for everything that leaves the firm, exactly as they would be for a trainee's draft. We build systems that make review practical rather than nominal, because a review step nobody has time for is not a control.
  • Privileged and confidential material stays inside your boundary. Putting identifiable client material into an uncontrolled third-party model is the risk your regulator and your insurer are worried about — and it is an architectural problem, not a policy one. We deploy on your infrastructure or in your own tenancy so the question doesn't arise.
  • Conflict checks stay deterministic. A probabilistic near-miss on a conflict search is a negligence claim. We build conflict checking as an exact-match gate with human sign-off, and log every search so you can show it was run and what it returned.
  • Relevance and privilege calls stay with a fee earner. A system can de-duplicate, order, index and surface candidates at a scale no team can match. Deciding what is disclosable, and what is privileged, is a judgement with consequences that a person makes on the record.
  • Anti-money-laundering duties can't be delegated to software — and the thresholds moved on 30 June 2026. The Money Laundering and Terrorist Financing (Amendment) Regulations 2026 converted several euro thresholds to sterling, including the occasional-transaction threshold to £800. Systems that hard-code thresholds need those values built as configuration, not code. The risk decision and any sanctions hit stay with your nominated officer.
  • There is no UK AI Act, and most legal back-office work isn't high-risk under the EU one. Anyone telling you that you need to comply with UK AI legislation is selling you something. Document extraction, file assembly, chasing and bundle preparation don't fall into the EU AI Act's high-risk categories, and for a UK firm the Act only applies where the output is used in the EU.
  • Client disclosure, ready to use. Every system we build for this sector ships with a plain-English description of what the AI does and where a human decides — so you can put it in your engagement terms if you choose to.
Our full compliance & governance approach →

What we won't build for a law firm

Some of these have already caused real professional consequences for firms that got them wrong.

Legal research or citations that reach a client or a court unverified

Fabricated case authorities have been placed before the High Court by practitioners relying on AI output, and the court's response has been unambiguous. Anything citable is verified against the source by a qualified person before it leaves the building — and we design the workflow so that's enforced rather than optional.

Advice that goes out without a qualified person behind it

A system can draft from your precedents, structure the reasoning and flag what's missing. Whether the advice is right for this client on these facts is the thing you are paid for, and the thing your insurer priced.

Deciding what is disclosable or privileged

The consequence of getting this wrong is a waiver you cannot take back, or a disclosure failure in front of a judge. The system proposes and surfaces. A fee earner decides, on the record.

Suspicious activity reporting decisions

The nominated officer decides whether to report. A system prepares the file and highlights the pattern. It does not form the suspicion.

Autonomous client communication on contentious matters

Anything touching deadlines, limitation, money or a dispute goes out under a human name, after a human has read it. No exceptions, no matter how good the draft.

Can a law firm use AI under professional conduct rules?

Yes. Regulators permit AI use and regulate by outcome: the solicitor remains personally responsible for the work, exactly as they would be for a trainee's draft. The failures that have reached court weren't caused by using AI — they were caused by putting unverified AI output in front of a judge. A well-designed system makes verification part of the workflow rather than an extra step someone skips under deadline pressure.

Is it safe to put privileged material into an AI system?

Only if the system runs somewhere privilege is already protected. Sending privileged or identifiable client material to a general-purpose public model is the risk your regulator and your insurer are worried about, and it's avoidable rather than inherent. Systems built on your own infrastructure, or in a private tenancy you control, keep the material inside the same boundary it already sits in — which is why most of our legal work is deployed that way rather than as a cloud subscription.

Can AI draft contracts and legal documents?

It can produce a first draft from your own precedents, which is a genuine time saving and a poor substitute for review. The useful design is retrieval-based: the system assembles from clauses your firm has already approved rather than generating language from a general model, so what comes back is recognisably yours and traceable to a source. Whether those clauses are right for this transaction, and what needs negotiating, remains a fee earner's judgement.

Can AI run a conflict of interest check?

It can run the search and assemble the results; it should not make the call. We build conflict checking as a deterministic exact-match search across your matter and party records, surfacing every candidate hit with the matter it came from, and holding file opening until a person signs it off. A fuzzy conflict search that quietly misses a match is worse than no system at all, because it creates confidence you haven't earned.

Can AI do disclosure review?

It can do the mechanical majority of it. De-duplication, threading, ordering, indexing, chronology building and surfacing likely-relevant documents are all work a system does faster and more consistently than a team of paralegals. What it must not do is make the relevance or privilege call — that decision has consequences a person has to own, including a waiver you cannot take back. The right design gives a fee earner a prioritised, deduplicated set to review rather than a decision to rubber-stamp.

Do we have to tell clients we're using AI?

Increasingly, yes — and it's worth doing regardless. Some regulators expect AI use to be set out in terms of engagement with an opt-out; others suggest firms consider disclosure. We ship every legal build with a plain-English description of what the system does and where a human decides, so you can include it in your engagement terms without drafting it yourself.

Will this reduce our billable hours?

It reduces the hours you can't bill for. Nobody pays you to reconstruct time entries, chase an opponent for a document, or paginate a bundle. In the practices we've built for, headcount didn't fall — the capacity went into work that could actually be billed. If your model depends on billing clients for administrative time, that's a genuine strategic question and we'd rather discuss it honestly up front.

Start where the matter stalls.

The fastest wins in a firm are almost always the processes that sit still between steps rather than the ones that take a long time to do. Tell us where yours sit still.

Let's talk