A small law firm can use AI to draft guides, posts, newsletters and review replies, provided a qualified lawyer approves every statement of law before it's published. Request reviews from all clients the same way, answer them without confirming anyone was a client, and keep out anything that could mislead: success claims, fees, specialisms and invented testimonials. Check your regulator's publicity rules first.
The rules bite harder for law firms than for most businesses. A practice guide that explains the law is read as advice, so a wrong time limit or a case that doesn't exist becomes a professional problem rather than an embarrassing typo. And confidentiality usually covers the fact that someone is your client, which changes how you handle every review, case study and testimonial.
Which marketing jobs AI can take, and who signs them off
Not every marketing task carries the same risk. Sort them before anyone opens a chat window:
| Task | What AI does | Who approves | Risk |
|---|---|---|---|
| Practice-area guides and FAQs | Drafts from a lawyer's notes, suggests structure and plain wording | Fee earner in that area | High: states the law |
| Newsletter | Summarises your own recent guides and firm news | Partner or compliance lead | Medium |
| Social posts | Cuts approved guides into short posts | Marketing, spot-checked by a lawyer | Medium: shortening can distort |
| Review requests | Personalises a standard request | Set once by compliance lead | Medium: platform rules |
| Review replies | Drafts a reply within fixed limits | Named person before posting | High: confidentiality |
| Case studies | Drafts from an anonymised summary with written consent | Supervising lawyer and client | High |
| Ad copy and landing pages | Variants of approved wording | Compliance lead | High: claims and fees |
The pattern: AI works from material a lawyer has already produced or approved, never from its own general knowledge of the law.
Practice-area content: a drafting workflow that keeps the law right
General chat assistants know a lot of law in a blurred, out-of-date, multi-jurisdiction way. That's the worst possible source for a guide with your firm's name on it. The fix is to make the lawyer the source and the AI the typist.
- Get 15 minutes of the fee earner's knowledge. A voice note or bullet list: the question clients actually ask, the answer, the time limits, the three things people get wrong, when to call a lawyer. Transcribe it.
- Draft from those notes only. Tell the AI which jurisdiction the guide covers and forbid it from adding legal content that isn't in the notes.
- Ask for the gaps, not the fill. Have the model list questions a reader would still have, so the lawyer can answer them rather than the AI guessing.
- Lawyer review against a checklist (a filled-in example follows).
- Publish with a date and a review date. Law changes; a guide that was right in March may not be in October.
A drafting prompt that works:
You are drafting a client guide for a small law firm.
Jurisdiction: [state it]. Audience: members of the public, no legal
background. Length: about 700 words.
Use ONLY the lawyer's notes below for any statement about the law,
time limits, costs or procedure. Do not add legal content, cases or
statistics that are not in the notes.
Structure: the question in plain words; the short answer; what happens
step by step; time limits; when to speak to a lawyer.
After the draft, list any questions a reader would still have that the
notes do not answer.
LAWYER'S NOTES:
[paste transcript]Run on a lawyer's notes about contesting a will, the draft came back like this (an illustrative sample):
If you believe a will does not reflect what the person really wanted,
you may be able to challenge it. Most claims must be brought within a
strict time limit, usually six months from the grant of probate, so it
is important to act quickly. In a landmark case, the court confirmed
that adult children can always claim a share of the estate.
Questions the notes do not answer: what it typically costs; whether a
claim can be settled without going to court.Two sentences need the lawyer's pen. The "six months" was in the notes for one type of claim, but the draft applies it to "most claims", which may be wrong for other routes. The "landmark case" isn't in the notes at all: the model invented a legal proposition with a vague citation, the classic failure described in stopping AI from inventing case law. The questions list at the end, on the other hand, is useful: both are worth a paragraph from the lawyer.
A pre-publication checklist, filled in for one guide
A blank checklist gets ticked without thought. Here's the same list completed for the will-contest guide after review:
| Check | Result | Action |
|---|---|---|
| Jurisdiction stated in the first paragraph | Missing | Added "This guide covers [jurisdiction] only" |
| Every time limit matches the lawyer's notes | One overgeneralised | Rewritten to name the claim type |
| No cases, statutes or statistics that weren't in the notes | One invented case | Deleted |
| No guarantee or outcome language ("you will", "we always win") | Clear | None |
| Fees described accurately or not at all | Not mentioned | Linked to the published fees page |
| General-information note and "speak to a lawyer" route | Present | None |
| Author, approval date and next review date recorded | Done | Review in six months |
That review took the lawyer 20 minutes. Without the checklist it would have been a skim, and the invented case would have gone live.
Social posts need the same care, because shortening is where hedges disappear. The approved guide says "you may be able to challenge a will if you were financially dependent on the person who died". The AI's 40-word post version said "Financially dependent on someone who died? You can challenge their will." One word turned a possibility into a promise. Tell the model to keep every "may", "usually" and "in some cases" from the source, and have a lawyer glance at a week's posts in one batch rather than each one separately.
Newsletters and adverts have a different weak point: words that need proof. In an illustrative newsletter draft, "Welcome to [name], our new family law specialist" came from a line in the lawyer's own notes that said "[name] joins the family team, ten years' experience in children matters". The AI upgraded experience into a title. Where your regulator restricts "specialist", that's a problem; the notes' own wording says the same thing without it. Similarly, asked for five variants of an approved advert, one illustrative variant opened "No win, no fee on all family claims", a fee arrangement the firm doesn't offer. Variants should only rearrange approved phrases, so tell the model to use no fee, outcome or status wording that isn't in the approved text, and check each variant against that text line by line. Keep a short list of words that always get a second look (specialist, expert, leading, guaranteed, no win no fee, free) and run every draft past it before approval.
Reviews: ask every client the same way
Review platforms have their own rules on top of the law. Google's policy for Maps content prohibits merchants from discouraging negative reviews or selectively soliciting positive ones, offering incentives for reviews, and asking staff to collect a set number of reviews or reviews with specific content. Consumer-protection regulators in several major markets have also made fake reviews expressly unlawful, including AI-written ones, alongside buying reviews and suppressing negative ones.
For a law firm, that means one standard request sent to every client when a matter closes, not just to the ones the partner thinks were happy. The practice of filtering who gets asked, often called review gating, is explained in what review gating is and why it's risky.
You will still need some exclusions, and they must be based on safety and client wishes, never on expected rating. Reasonable ones: the client asked not to be contacted, the matter involved domestic abuse or another situation where an email could put someone at risk, or the matter ended in a complaint that's still open. Write them down so the rule is visible. Then check once a quarter that the rule is what actually happens: list the matters closed and the requests sent, and account for every gap. An illustrative month: 40 matters closed, 34 requests sent, six exclusions (three at the client's request, two for safety, one open complaint). Every exclusion has a written reason from the list. If "partner felt it wasn't appropriate" starts appearing as a reason, gating has crept back in.
AI can personalise the request lightly. A filled-in example for a closed conveyancing matter:
Subject: Thank you, and a quick request
Dear [client name],
Now that your purchase has completed, thank you for choosing us. If you
have a moment, we'd be grateful for an honest review of your experience,
whatever it was. It helps other people decide whether we're right for
them. [link]
With thanks, [name]Keep the matter type vague in the email itself, and never mention the review in the same message as a bill.
Replying to reviews without confirming anyone was your client
This is where AI drafts go wrong most often, because a helpful model wants to address the specifics. Under most legal professional rules, the fact that someone instructed you is itself confidential, so a reply must not confirm the reviewer was a client, mention their matter, or correct their version of events with facts from the file.
Suppose a one-star review says the firm "took months to reply and got my divorce paperwork wrong":
AI's first draft (don't post): "We're sorry you feel this way. Our records show we responded to your emails within five working days and that the delay was caused by your former partner's solicitor."
After editing: "Thank you for taking the time to leave feedback. We take every comment seriously, but our duty of confidentiality means we can't discuss any individual matter here. If you'd like to talk about your experience, please contact our practice manager on [number] and we'll look into it properly."
The first draft confirms a client relationship, reveals the matter type and discloses file details. Build the limits into the prompt: "Do not confirm or deny the reviewer is or was a client. Do not refer to any legal matter, date or fact. Invite them to contact [role] directly. Under 60 words." Then a named person reads every reply before it's posted. For the harder cases, replying to negative reviews with AI has more examples.
Positive reviews carry the same trap, and it's easier to miss because the reviewer has already said who they are. A five-star review reads: "Thank you for sorting out Mum's estate so kindly, the whole team was brilliant." The AI's warm first draft: "It was a pleasure helping you through probate, and we're so glad the estate is now settled." That reply confirms the matter, its type and its outcome. The reviewer chose to share some of it; the firm hasn't been given permission to add the rest. A safe version: "Thank you for your kind words. We'll pass them on to the team." Short replies to good reviews look gracious, not cold, and the prompt limits above apply to every reply whatever the star rating.
Case studies: anonymising is harder than removing the name
A case study is the most persuasive thing a firm can publish and the easiest way to identify a client by accident. Written consent comes first, from the client, for the exact text. Then the anonymisation, which AI is surprisingly bad at, because it treats "anonymise" as "remove the name". An illustrative draft from a lawyer's summary of a family matter:
We recently helped a retired head teacher whose husband ran the only
dairy farm in their village. After his death, his brother claimed the
farm had been promised to him...No name, and still identifiable to anyone who lives nearby: the profession, the only dairy farm and the family relationship pin it down. The edited version keeps the lesson and loses the fingerprints: "We recently helped a client whose late husband's family business was claimed by a relative who said it had been promised to him." Ask the model to list every detail that could identify the client (occupation, place, unusual facts, dates, family relationships) before it drafts, then decide with the client which, if any, stay. Send the client the final text, not a summary of it, and keep their written approval with the matter file.
The rules to check before anything goes live
Four layers of rules apply, and your regulator's publicity rules are the one only you can look up. Check each of these against the actual wording:
- Platform rules. The review policies above, plus each social network's rules on advertising legal services, which some restrict.
- Consumer-protection law. No fake or AI-invented reviews or testimonials, no misleading claims, clear pricing where you show prices at all.
- Your professional regulator. Look for rules on: accuracy and not misleading the public; publishing price information for certain services, which some regulators require; words like "specialist" or "expert", which some restrict; statements about past results or success rates; testimonials and endorsements; comparisons with other firms; and unsolicited approaches to people who may need legal help, which many regulators limit.
- AI-specific duties. If a chatbot on your site talks to visitors in the EU, they must be told it's an AI. Client data used in marketing (for a case study, say) needs a lawful basis and usually written consent; whether solicitors can use ChatGPT without breaching confidentiality covers the tool-side settings.
If a rule is ambiguous, ask your regulator's ethics helpline or your professional indemnity insurer's risk team before publishing, not after a complaint.
A month of marketing for a four-lawyer firm
Picture, for illustration, a four-lawyer practice handling wills, family and residential property, with a marketing assistant two days a week. Their month:
- Two practice guides. Before: about 6 hours each for the assistant plus an hour of lawyer time. After: 15 minutes of dictation, 90 minutes of drafting and editing, 20 minutes of lawyer review against the checklist, 15 minutes of formatting. About 2.5 hours each.
- Eight social posts cut from the approved guides: 30 minutes each by hand, about 12 minutes each with AI, with one lawyer spot-check a week.
- One newsletter: down from 3 hours to about 1.5.
- Review requests to around 40 closed matters, sent automatically from a list the practice manager updates weekly. Replies to the reviews that arrive: about 5 minutes each including approval.
That's roughly 10 hours of the assistant's month back. The cost that matters is not the software (two seats of Claude Team or ChatGPT Business, the minimum on both, is about $50 a month on monthly billing) but the lawyer review time, around 90 minutes a month. Budget it explicitly. If it gets squeezed, the checklist gets skipped, and a skipped checklist is how invented cases end up on your website.
Measure more than hours. Track enquiries that mention a guide, review volume and average rating over a quarter, and any complaint or regulator query linked to marketing. The last number should stay at zero.
Law firm marketing and AI: follow-up questions
Can a chatbot on our website answer legal questions?
Keep it to intake and logistics: opening hours, which matters you handle, how fees work in general, and booking a first call. Anything that reads like a request for advice should go to a person, because an answer from a bot can look like advice from the firm. Tell visitors they are talking to an AI, which has been a legal duty for EU visitors since August 2026, and keep the conversation logs.
Can we use AI-generated images in our marketing?
Yes, with care. Avoid images that look like real clients, real staff or real events, because that can mislead. Check the image tool's terms allow commercial use, keep a record of which images are generated, and never produce anything that imitates a real person. Stock-style illustrations of offices or documents carry far less risk than realistic photographs of people.
Who should sign off AI-drafted marketing?
A qualified lawyer for anything that states the law, describes a service or mentions outcomes; your compliance lead for anything touching fees, specialisms or client stories; and whoever owns the brand for tone. Keep a simple log of the version approved, who approved it and when. If a complaint arrives two years later, that log is what shows the content was checked.
Further reads
- How to Fact-Check AI Marketing Copy Before It Goes Live — A general fact-check routine for any copy before it goes live.
- How to Write an AI Content Policy for Your Marketing — Put your sign-off rules in writing so staff follow them.
- Can You Legally Use AI-Generated Images in Your Marketing? — The wider rules on using AI-generated images commercially.
- How to Get More Google Reviews With Automated Review Requests — Set up automated review requests step by step.
- AI for Small Law Firms: What to Automate First — Where marketing sits among a small firm's automation priorities.
- Which Legal Tasks Should a Small Firm Never Hand to AI? — The work that should stay entirely with lawyers.
- Can You Use AI to Write Testimonials? Where the Legal Line Is — Where the legal line sits for AI and testimonials: an allowed-grey-never table, what the rules prohibit, an approved-edit workflow and a consent email.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Google Maps user-contributed content policy (prohibited and restricted content, fake engagement); consumer-protection regulators' published rules on fake reviews; EU AI Act Article 50.