Artificial intelligence (AI) for law firms: how to run intake, research, drafting and matters, process by process

Onboarding a client for regulated work is where the firm's anti-money-laundering duty begins, and it begins before you can act. For work such as conveyancing, company and trust work and managing client money, the Money Laundering Regulations 2017 require you to identify the client and any beneficial owner and to assess risk, whether simplified, standard or enhanced, before the retainer opens.

Most of that is structured, repetitive work: collecting and verifying names, dates of birth, addresses, source of funds and identity documents, in the same order every time. It is exactly the kind of capture an assisted flow can order, pre-fill and keep an audit trail of, so the file is complete and consistent before anyone makes a decision on it.

The supervision behind it is active and enforced, not nominal. The SRA supervises 5,569 firms for anti-money-laundering purposes, inspected or reviewed 833 in 2024-25, found about 32% non-compliant and issued fines totalling £1,498,983. A clean, ordered, retained due-diligence trail therefore has real value, because the record-keeping is checked.

An assistant built for your firm runs the identity and beneficial-owner capture, requests documents through a secure checklist, pre-fills what it can and logs every check as it happens. What it never does is make the call. The documented risk assessment, the ongoing monitoring and any suspicious-activity report to the National Crime Agency remain professional acts of the firm and its nominated officer.

One warning shapes the whole build. The SRA has cautioned that clients may use AI, including deepfakes and forged documents, to bypass identity checks, so verification has to be robust rather than a formality. The assistant standardises the checks and flags anomalies, but the firm must satisfy itself on identity, and the COLP is responsible for regulatory compliance when new technology is introduced into onboarding.

Research is where AI is most tempting and most dangerous at once. It can genuinely accelerate the work, summarising material, surfacing the points that matter and drafting a first-pass analysis in minutes. But a general language model is, in the Divisional Court's own words, not capable of conducting reliable legal research, and treating it as if it were is what gets solicitors into trouble.

The cautionary case is close to home and recent. In Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank, decided on 6 June 2025, the court found that generative AI had produced fictitious cases, fake citations and misstatements of law that were put before it. It referred the solicitors involved to the SRA and warned of wasted-costs and contempt exposure. The lesson is not that AI is useless for research; it is that every authority and quotation must be verified against the primary source before it is relied on.

The profession's caution is well founded and measurable: 77% of UK lawyers are concerned about AI producing inaccurate outputs, and roughly half of those who use it still reach for general-purpose models rather than tools grounded in real case law. That is precisely where the accuracy and confidentiality risk concentrates.

An assistant built for your firm can still do the useful part. It summarises, surfaces relevant points and drafts a first analysis grounded in verified legal sources, with citations you can open and confirm, and it runs on confidential tooling so no client fact goes into an open model. What it produces is a draft to be checked, never the last word.

The control is a hard rule, stated plainly: AI may draft and summarise, but no case, statute or citation reaches a client or a court until a solicitor has checked it against the real report. You remain responsible and accountable for every output, and that accountability cannot be delegated to a tool or a provider.

Document review is high-volume, repetitive reading. Leases, contracts, disclosure bundles and correspondence all have to be read, the clauses that matter extracted and the risks flagged, and it is precisely this first-read work where AI-assisted extraction and summarisation can save the most time.

An assistant can propose a first-pass summary, pull out the clauses that carry risk, break rights, indemnities, liability caps, assignment, and highlight anomalies, turning hours of first-read into a checked draft. On a large disclosure set it can go further, prioritising and clustering documents a person could not read in the time available. But the relevance call, and critically the privilege call, stay with the solicitor: the tool proposes, the solicitor disposes.

The confidentiality risk is what shapes the whole build. Documents under review are confidential and frequently privileged, and putting them into a public AI tool can be treated as disclosure to a third party and can risk waiving legal professional privilege. That risk is not theoretical when roughly half of the lawyers who use AI still reach for general-purpose models rather than confidential, contractually bound tooling.

An assistant built for your firm therefore runs on confidential tooling hosted in the UK or EU, under a written processor contract, so the review speeds up without putting privilege at risk. It reads the corpus, proposes a summary and flags what looks borderline, and a human checks anything the model marks as uncertain.

The line is clear. A missed clause or a wrong privilege call is the firm's responsibility, not the tool's. AI reads, extracts and proposes; the solicitor makes the relevance and privilege decisions, signs off the review and owns the outcome.

A great deal of drafting is repetitive and precedent-based. Client care and engagement letters, standard correspondence, first-draft pleadings and witness statements can all be produced as a first draft from the firm's own precedents, for a solicitor to finalise. An assistant can assemble a structured first draft in minutes, in the firm's house style, from the matter facts and the relevant template.

That leaves the fee-earner the higher-value work: tailoring, judgement and sign-off, rather than starting from a blank page at eight in the evening. And it matters at scale, because every one of the 158,295 solicitors holding a practising certificate in England and Wales remains personally responsible under the SRA Standards and Regulations for the work they sign, whatever tool produced the first version.

Some drafts are especially strong candidates because their required contents are stable. The client care letter is not optional: the SRA requires firms to give clients clear information about costs, the complaints process and how the firm is regulated. An assistant can populate that from the firm's standard wording and the matter details, and the solicitor checks and issues it.

But drafting is also where fabrication risk is highest. A first draft that looks polished can still contain a case that does not exist or a fact the client never gave, and the Ayinde ruling shows that AI-inserted fake authorities or invented facts in pleadings and statements can lead to referral to the SRA. A witness statement in particular must reflect the witness's own evidence, not the model's invention.

So the control is the same as for research, applied to every outgoing document. No citation, authority or asserted fact survives into a filed document until a solicitor has verified it. AI produces drafts, not advice; the solicitor verifies and owns every letter, pleading and statement, and remains responsible for it.

Most firms are sitting on years of their own best work, and can barely find it. The precedent used on the last share purchase, the note a partner wrote on a tricky point of construction, the engagement letter tailored for this kind of transaction: it all exists, scattered across matters, shared drives and personal inboxes. So a fee-earner facing a familiar problem too often starts from a blank page, or rebuilds from memory something the firm has already done well.

The value of a grounded assistant here is retrieval, not invention. Point it at the firm's own material and a fee-earner can ask for the clause we used last time, the template for this deal, the advice we gave on a similar point, and get an answer that carries a link to the source document. The starting point is the firm's accumulated work rather than a first draft written from nothing.

This is also the pattern the profession is best placed to adopt first, because it is closed. LexisNexis research puts generative AI use among UK lawyers at 61%, yet only 17% say it is embedded firm-wide, and the gap is trust: confidentiality, data protection and accuracy. A retrieval assistant over the firm's own data, hosted in the UK or EU under a processor contract, never needs client information to leave the firm's control, which makes it a controllable first build rather than a leap.

The internal confidentiality boundary still has to hold. The SRA duty of confidentiality is owed to each client, so a firm-wide assistant needs the same information barriers the firm already operates, and one client's confidential material must never surface in another client's matter. And the golden rule does not bend: an answer from the knowledge base is a starting point the fee-earner checks against the current law and the specific instruction. The solicitor owns the advice, always.

A surprising share of client contact is not legal work at all, it is chasing progress. Where are we, what happens next, has the other side replied yet. Clients grow anxious in the quiet stretches between milestones, and the natural response is to pick up the phone or fire off an email, usually to the one fee-earner who already has a full desk.

That pressure lands hardest at the small end of the profession. Around 364,000 people work in UK legal services, but at the thousands of small and high street firms a handful of fee-earners carry the fee-earning, the client care and the admin all at once. A fortnight of status questions from a dozen clients is enough to swallow a day that should have gone into the matters themselves.

An assistant can take the routine ones off the desk. From the matter's own record it drafts a clear, dated update, the last step taken, the next step due, what is currently awaited, for the fee-earner to approve and send. A recurring interruption becomes a quick review, and the client gets a faster, more consistent answer than a delayed callback would have given.

The boundary is where a status update starts to shade into advice. A view on the merits, on the likely outcome, on what the client should do next, is professional judgement, so the rule is simple: AI drafts, the solicitor decides what goes out. Keeping the client properly informed is an SRA obligation, not a courtesy, so anything bearing on the client's decision is confirmed by the fee-earner, and the client is always told plainly where they are interfacing with AI rather than their solicitor.

Few things frighten a firm like a missed limitation date, and with good reason: a deadline gone by is one of the leading causes of professional negligence claims against solicitors. The painful part is that these dates are not mysteries. The Limitation Act 1980 sets the periods, six years for most contract and tort claims, three years for personal injury, and the clock runs from facts that are already sitting in the matter.

On top of limitation sits the litigation calendar. Court and tribunal deadlines and directions under the Civil Procedure Rules, filing windows, service deadlines, hearing dates, give every contested matter a running set of dates. Across a busy caseload that is a lot to hold in one head, and every one of the 158,295 solicitors with a practising certificate in England and Wales carries personal responsibility for meeting them.

Because limitation periods and court directions follow rules and dates already in the matter, a diary is exactly computable rather than remembered. An assistant can hold that calendar matter by matter, compute who is due what and when, and start prompting the fee-earner early, so a full caseload does not resolve into a last-minute scramble or, far worse, a date that slips past unnoticed.

What it does not do is decide the date. A computed limitation date is decision support, not the decision: the golden rule is that the fee-earner validates it against the Limitation Act, the CPR and any court order, and owns it. The point of the diary is to raise the alarm earlier and more reliably than a person tracking dates by hand, not to remove the solicitor's obligation to get the date right.

Time recording is the quiet leak in a firm's finances. Work gets done, the day moves on, and the six minutes here and the ten minutes there never make it onto the matter. Unrecorded time is simply lost revenue, and it disappears fastest on exactly the days that were busiest, when there was no gap to stop and write the narrative.

That lost time is set against a market where the money is real. UK legal services revenue rose 11% to £52.3 billion in 2024, and margin at the individual firm turns on capturing the work it actually does. An assistant can draft a plausible, matter-linked time narrative from the fee-earner's own activity, for them to check and post, recovering time that a busy day would otherwise have swallowed while keeping the human in control of what is charged.

Billing has its own standards to meet. The SRA transparency and costs-information rules require firms to give clients clear price and service information, so a bill has to be accurate and understandable. The assistant can assemble a clear draft bill and a plain-English costs update from the recorded time and the agreed basis of charging, but the fee-earner confirms the figures and the firm is responsible for the bill that goes out.

The firm's own tax filing has gone digital too. All VAT-registered businesses, which most firms are, must keep digital records under Making Tax Digital for VAT, and from 6 April 2026 sole traders and landlords with qualifying income over £50,000 fall under MTD for Income Tax, a threshold falling to £30,000 in 2027 and £20,000 in 2028, which reaches sole-practitioner and partnership principals. Assisted data handling keeps the firm's records structured and current, so each VAT return and quarterly update starts from a trustworthy position, but the filing and the responsibility stay with the firm. Internal figures, charge-out rates, margins and work in progress, stay confidential and inside the firm.