Generative AI for Law Firms: Which Matters Suit It and Which Do Not

Generative AI suits routine, document-heavy matters where a lawyer can check output against a source in minutes: standard letters, summaries of papers you already hold, first-draft templates. It suits advisory work only as a drafting aid, and contentious work only for internal organising. Anything citing authorities, filed or sent as advice needs full lawyer verification.

Think of it as supervision by matter type, not by tool. The question for each file is: what is the worst thing an unchecked error could do here, and how quickly can a qualified person catch it? Where the cost is low and the check is fast, use AI freely. Where the cost is high or the check is slow, narrow its role or leave it out.

Follow me on Instagram@sagnikteaches

A grid you can adopt: stakes against checkability

Two questions place any task. First, what happens if the output is wrong and nobody notices: an embarrassing email, a missed deadline, a misleading statement to a court or client, or a breach of confidence? Second, how checkable is the output: can the reviewer compare it to a document in front of them in minutes, or does verification mean researching from scratch? Low stakes with fast checks is the green zone. High stakes with slow checks is the red one.

Connect on LinkedInSagnik Bhattacharya

The grid is a starting point. Your firm's insurer, regulator and clients may push particular rows into a different colour, and you should check each before adopting it. What matters is writing it down so supervision is a rule, not a hunch.

Subscribe on YouTube@sagnikteaches

Routine matters: where the gain is quickest

Take a small firm with a busy residential practice. Each completion generates a dozen near-identical letters. A fee earner supplies the template, the matter details (client name, property, dates, amounts) and asks for a draft in the firm's tone. The reviewer ticks every variable against the file. The model adds little beyond layout and phrasing, which is exactly the point: the risk is confined to transcription, and transcription is checkable.

Draft a client update letter from the template below.
Use ONLY the details I give you. Do not add legal advice, dates or amounts.
Put [CHECK] after any sentence that depends on a date or an amount.
Template: [paste firm template]
Details: Client [first name]; exchange expected 14 days after the buyer's searches are returned;
searches ordered; no enquiries outstanding.

The [CHECK] markers are a practical device. They push the reviewer's eye to the sentences most likely to harm if wrong. How small law firms use AI to draft letters and routine documents goes deeper on templates, tone and client update messages.

Document summaries sit here too, with a condition. A summary of a 40-page lease that you have on file is checkable because the reviewer can open the page. Require page references in the output, and reject any summary that cannot supply them. An illustrative good result lists "Break clause: tenant may end the term on six months' notice (p. 12, clause 7.3)". The reviewer opens page 12 and reads the clause, which takes under a minute.

Advisory matters: drafting aid, never the adviser

For advice work, the model can speed the first draft and the reading, but it cannot hold the professional responsibility. A lawyer reviewing a supplier contract for a small client might ask AI to list unusual clauses, then read the full contract anyway. The list is a second opinion on where to look, not a substitute for looking. AI contract review for small firms shows what these tools catch and what they miss, and it is worth reading before you promise clients a faster turnaround.

Three rules help in advisory work:

1. Facts in, not out. Give the model the facts as the client stated them and ask it to flag gaps. Do not let it fill gaps from "typical" situations.
2. Law from sources. Where a point of law matters, the lawyer finds the authority in an official or trusted legal database and reads it. If AI suggested the authority, treat it as a hypothesis.
3. Your voice, your name. The letter goes out under a lawyer's name after that lawyer has read every sentence. If they would not say it aloud to the client, it is not ready.

A realistic failure: a first draft of an advice note states that a notice period is "typically 28 days" in a way that reads as a rule, when the client's agreement said 56 days. The model had blended general patterns with the document. The reviewer caught it because the note was checked against the contract clause, not against common sense. For rules on authorities specifically, see how to stop AI inventing case law.

Contentious matters: organise internally, verify externally

Litigation and disputes punish unverified material hardest, because a false statement to a court is a professional problem, not only an embarrassment. Fabricated citations have been in the news repeatedly, so the rule here is blunt: no authority, quotation or procedural rule reaches a document unless a person has read it in an original source.

Useful, lower-risk internal uses exist. Say a four-person firm acts in a commercial dispute with 600 emails. A fee earner can ask AI to build a draft chronology from the emails, each entry with the sender, date and a quote reference. A paralegal then checks entries against the emails. The chronology saves hours, and since every entry carries a reference, wrong entries are findable. Another use is turning the lawyer's own rough notes into a clean skeleton argument outline. The words and law remain the lawyer's.

Privilege and confidentiality need special care. Sending privileged documents into a tool is a decision with consequences: confirm the plan's terms on training and retention, and keep the document set minimal. Summarising bundles and transcripts with AI safely covers the practical steps.

A chronology prompt and what its output looks like

Here is the contentious-side use in detail, because it shows both the benefit and the fix. The fee earner pastes a batch of emails (client-approved, minimal, no privileged advice) and asks:

Build a chronology from the emails below.
Table columns: date, event, who, source (email subject and sender), confidence (clear / inferred).
Only include events stated in the emails. Do not infer motives.
If two emails disagree about a date, list both and mark the row DISPUTED.

An illustrative output row reads: "3 March | Supplier says the delivery was rejected | Supplier's manager | 'Re: pallet 14', from the supplier | clear". Another reads: "10 March | Client says payment was made | Client | 'Payment sent', from the client | inferred", because the email said "sent" but gave no bank reference. What you fix: the "inferred" rows are the ones to chase for evidence, and anything the model marked DISPUTED goes to the partner. What you must not do is let the chronology become the exhibit. It is an index to the emails, and the emails are the evidence.

Before and after: an advice paragraph, first draft to sendable

Take a short advice paragraph on an employment query. The first AI draft says: "You are entitled to a minimum of four weeks' notice and the company must follow a fair process." The lawyer reads it against the client's contract and the facts. The contract says eight weeks. The facts show a short length of service, and what process the law that applies requires for that length of service is a question the lawyer checks, not one the draft can assume. The lawyer rewrites: "Your contract provides for eight weeks' notice (clause 11). Whether a formal process was required depends on your length of service, the law that applies to your employment and the contract's disciplinary terms, which we need to confirm." The changes are small in words but large in substance, and every one came from reading the source. That is the value of review. A firm that sees only the first draft's fluency, and not the gap between it and the contract, would have sent a wrong letter.

The four mistakes small firms make first

• Treating a polished paragraph as a checked one. Fluent writing hides errors better than clumsy writing does. Review by reading against the file, not by skimming for sense.
• Using a free consumer account for client facts. The plan matters more than the model. Use a business plan with no training on your content, or a legal-specific tool with contractual confidentiality terms, and check the current terms on the vendor's page.
• Setting supervision by who is cheapest. The least experienced person ends up checking the highest-risk output. Match reviewer to matter type, and put the name on the grid.
• No record. When a client or an insurer asks how a document was produced, "someone used a chatbot" is not an answer. A file-note line fixes this at almost no cost.

A quick sum: what a green task saves

Say the firm sends 30 completion-stage client update letters a month. Each takes a fee earner about 10 minutes to write from scratch. With a template-driven AI draft and a check against the file, it takes 4 minutes. That saves 6 minutes a letter, or 3 hours a month, roughly half a working day. Over a year, about 36 hours. It is modest, and that is the honest scale of most green-zone gains. The bigger prize is usually in summaries and chronologies on contentious files, where an hour saved is a billable hour redeployed. Fix your expectations before you pilot, so a modest saving is not read as failure.

Probate and wills: a worked walk-through of one matter

To see the grid on a single file, take an estate administration handled by a small firm. At intake, the client gives a list of assets and family details. A fee earner asks AI to turn the intake notes into a checklist of documents to request (bank statements, policy details, property papers), phrased as a client-friendly letter. That is green: the checklist is checkable against the firm's own standard list. Next, the firm gathers statements and wants a schedule of assets. AI may extract balances from pasted statement text into a table, amber, because every figure must be ticked to the statement and the total reconciled by hand. When the estate account is drawn up, the figures live in the accounts system and AI has no part in them, red. The distribution letter to beneficiaries is drafted from the approved account figures, amber again, with the reviewer ticking each name and amount. One matter, four rows of the grid, and the reviewer's job is clear at each step.

The same approach applies to a landlord and tenant file or a small business sale: break the matter into tasks, place each task, name the reviewer. It is the task, not the file, that carries the risk.

What to tell clients and how to word it

Clients do not need a technical lecture. A short paragraph in the engagement letter usually covers it: that the firm may use AI tools to assist with drafting and summarising, that confidential information is handled under the firm's data-protection terms, that a qualified lawyer reviews all work before it is sent, and that the client can ask questions or object. Have a partner approve the wording and check it against your regulator's guidance. If a client objects, you need a way to mark the file as no-AI, so build that flag into your case-management notes from the start.

Matters that should stay off the tool

Some work is better kept out. Client money handling, conflict decisions, limitation and deadline calculations and anything where the reasoning must be the professional's own judgement belong in your accounts, case-management and diary systems, not in a chat window. Which legal tasks a small firm should never hand to AI lists them. Client money is the clearest example: the ledger and the reconciliation must come from the accounts system, with a person signing them off.

Matters for vulnerable clients, or those involving children, safeguarding or serious criminal charges, deserve a conservative default. That does not mean a ban on every use. It means the grid is set stricter, and a partner decides case by case.

Filling in the grid for a four-lawyer firm

Say a four-lawyer firm does residential conveyancing, wills and probate, employment advice and a small litigation practice. In an hour at a whiteboard they list tasks per team and place each in the grid. For conveyancing: client updates and standard letters are green; reading a lease summary is amber with page references; anything on title issues is red for AI drafting. For wills and probate: estate account layouts are amber because figures are involved; the will itself is drafted by a lawyer from instructions. For employment: first drafts of policies are amber; advice on a dismissal is the lawyer's own work, with AI limited to organising facts. For litigation: chronologies are amber; submissions are red for AI content.

The output is a single page per team. Each row has a task, a colour, a named reviewer and a one-line verification step. Once that exists, the firm can answer the most common questions from clients, insurers and staff without improvising.

Supervision, records and insurance

Three practical records keep you comfortable. A register of tools and plans in use, showing the vendor terms checked and the date. A matter-level note where AI assisted, recording the task and the reviewer. And a short training record so staff know the grid. None of this has to be heavy. A line in the file note, "first draft produced with AI; reviewed in full by [name]", is usually enough.

Ask your professional indemnity insurer what it expects. A proposal form or renewal questionnaire may ask about AI use, and it is far better to have written answers prepared than to improvise on the form. If you use a legal-specific tool, compare it with a general assistant on data handling and cost; ChatGPT or a legal AI tool: which should a small firm use sets out the trade-offs, including cost.

A pilot you can run in a month

Choose one green and one amber task. For four weeks, record for each use: time taken with AI, time the reviewer spent, number of corrections and the severity of the worst one. Compare against a baseline from similar earlier files. If the green task saves time and produces nothing worse than wording changes, keep it. If the amber task produces substantive errors that took longer to fix than the drafting saved, narrow its role or drop it.

Review the grid every quarter and whenever a vendor changes terms, a regulator publishes guidance, or a complaint or near-miss occurs. A near-miss caught by a reviewer is the system working. Write it down, share it with the team, and adjust the grid. Over time, the sensible pattern is a growing green column and an amber column with sharper verification steps, not a creeping habit of trusting output because it reads well.

If you would rather build the grid with someone who has seen how small firms apply it, that is what my AI implementation consultation covers, at the contact page.

Questions partners ask before using generative AI on matters

Which matters are safest to start with?

Routine, repeatable work with low stakes and a clear template: standard letters, document tidying, client update drafts, internal summaries of documents you already hold. The common factor is that a qualified person can check the output quickly against a source they have in front of them.

Should contentious matters use AI at all?

Yes for internal tasks such as organising a chronology or drafting a first outline, as long as nothing AI produces is filed or relied on without verification. Authorities, quotations and facts must be checked against primary sources by a person, every time.

Do I need client consent?

It depends on your professional rules, your engagement terms and what data goes into the tool. Check your regulator's guidance and your insurer's requirements, and consider telling clients in the engagement letter how AI-assisted drafting is supervised.

Can a junior use it unsupervised?

Set the supervision level by matter type, not by seniority. A trainee drafting a standard letter with a template and a reviewer is different from a trainee asking a chatbot for case law. Write the rule down and put a name against review.

What should be logged?

Record the matter reference, the task, the tool and plan used, who reviewed the output and when. You do not need to keep every prompt, but you need enough to show how a particular document was produced if asked.

Further reads

• AI Readiness Checklist for Accountants, Solicitors, Consultants — A readiness check across a whole professional firm.
• AI Client Intake for Law Firms: Qualify Enquiries Out of Hours — Handling new enquiries out of hours before a lawyer reviews them.
• AI for Small Law Firms: What to Automate First — A shortlist of first automations for a small firm.
• AI Software for Solicitors: Data, Supervision and Supplier Checks — Data, supervision and supplier checks before you buy.
• AI Time Capture for Solicitors: Recover Hours You Never Billed — Recovering unbilled time, a low-risk first use.
• How to Mention AI Use in Client Contracts and Proposals — Wording to tell clients how AI is used on their matters.
• Can Solicitors Use ChatGPT Without Breaching Confidentiality? — Confidentiality questions before pasting client facts.
• AI Implementation Plan for a Small Law Firm: The First 90 Days — A 90-day sequence for rolling the grid out.
• E-Discovery Software for Small Law Firms: Costs, Limits and AI Review — Six e-discovery pricing models, one matter costed from your own data, and a sampling test for AI review, written for a firm with a handful of lawyers.
• Family Law Document Automation: Gavel Workflows or Generative AI? — Rules-based assembly against prompting for family law documents: costs for a four-fee-earner firm, how each fails, and a hybrid that uses both.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: vendor business-plan privacy pages and professional-regulator guidance on AI use, checked in 2026. The matter grid and examples are illustrative and do not describe any real matter.