How to Use AI in a Law Practice: A Four-Week Starter Plan

Spend week one writing a short AI policy and choosing one approved tool, week two piloting one low-risk task with no client confidences, week three building a citation-verification routine, and week four settling client disclosure and reviewing results. Add client-matter work only after the fourth week passes the review.

The sequence is deliberately slow at the start. The failures that reach the news, such as invented case citations in filings, come from skipping week three, and the failures nobody reports, such as client details pasted into a free chatbot, come from skipping week one. A practice that does the weeks in order spends perhaps six to ten hours across the month and ends with something it can defend to a client, an insurer or a regulator.

Follow me on Instagram@sagnikteaches

Before week one: who owns the plan

Name one person as the AI lead. In a four-person practice it is usually a partner with a technical streak, or the practice manager with a partner as sponsor. Their job is small and specific: keep the policy, keep the approved-tools list, run the verification log and call the monthly review. Give them an hour a week for the first month.

Connect on LinkedInSagnik Bhattacharya

Illustrative practice for this plan: a four-person firm with one partner, two associates and one paralegal, doing a mix of property, wills and small-business advice. The numbers below come from that firm. Your own mix will differ; the plan doesn't depend on the specialism. For specialism-specific tool choices, which tasks a small practice should start with and the seven-tool comparison go further.

Subscribe on YouTube@sagnikteaches

Week 1: Policy, tool and boundaries

The aim this week is that everyone knows what is allowed, in one page.

Day 1 to 2: Write the one-page policy

Keep it short enough to be read. Cover six points:

1. Approved tools. A named list. Everything else is not approved for work. Start with one.
2. Confidentiality. What may never be entered: client names, matter details, anything covered by privilege, opposing party details and personal data. Until the tool is vetted, nothing from a live matter goes in.
3. Verification. Every legal citation, quotation and statement of law produced by AI is verified against a primary or trusted source by a named person before it is used.
4. Human responsibility. The lawyer who signs the document is responsible for all of it, whoever or whatever drafted it.
5. Records. Log the task, the tool and who verified, in the matter file where a matter is involved.
6. Review. The policy is reviewed after four weeks and every quarter after that.

Your professional body may have published guidance on AI and technology competence; read it, and align your policy. Rules about confidentiality and supervision differ from place to place, and this tutorial is practical guidance, not legal advice, so ask a qualified adviser to look at the policy once.

Day 3: Choose and vet one tool

Two routes suit a small practice. A general assistant on a business plan (ChatGPT Business or Claude Team, both of which do not train on business content by default, and both need at least two seats) is cheapest to start. A legal-specific platform, or your practice-management system's built-in AI, may be better when it sits inside the matter system you already use. Clio, for instance, listed four plans starting at $49 per user a month on its pricing page in October 2026, with AI drafting added from the Core plan up; check the current plan names, prices and what the AI actually does before you decide. Lexis+ with Protégé is another route; its product page shows no price and offers a trial and a sales contact instead, so ask for a quote. The comparison of Clio, Smokeball and Lawmatics sets out what each platform offers.

Whatever you choose, ask the vendor in writing four questions and file the replies: Is our content used to train models? How long is it retained and can we delete it? Where is it processed? Who can access it? A vendor who cannot answer in plain words does not pass.

Day 4 to 5: Brief the team and sign

Hold a 30-minute session. Show the policy, show one live demonstration on a non-client task and show one example of a wrong answer so nobody leaves thinking the tool is reliable. Everyone signs the page. The sign-off is the first item in your AI file.

Week 2: One low-risk task, tested properly

Pick one task with three properties: no client confidences, a clear way to check the result, and a person who does it often. Good candidates for a small practice:

• Rewriting the firm's own client information leaflets in plainer English.
• Drafting an internal checklist for a standard process, such as opening a new file.
• Summarising a public document, such as a published guidance note, into a one-page briefing for the team.
• Drafting the firm's own marketing or newsletter content from the partner's bullet points.

Say the practice picks the leaflets. The paralegal gives the tool an existing leaflet on "what to expect when you buy a home" and asks for a plain-English rewrite aimed at a first-time buyer, at a reading level around 12 years old, with all legal terms kept and explained in a clause.

Rewrite this client leaflet for first-time home buyers in plain English. Keep every legal term but explain each in a short clause the first time it appears. Do not add any legal information that is not in the original. Keep it under 500 words. After the rewrite, list any sentence where you changed the legal meaning, however slightly, and any term you were unsure how to explain.

A sample result, illustrative: the rewrite is clearer and 420 words, and the "changed meaning" list contains one item: it turned "exchange of contracts is binding" into "once you exchange, you must complete", a simplification that drops the exceptions. The associate restores the original wording and adds the exceptions in plain language. That is exactly the review step working: the model's list points to the sentence a lawyer should check, and the lawyer decides.

Time the task. In this illustration the old way took the paralegal 2.5 hours, the tool-assisted way took 45 minutes plus 30 minutes of associate review, so 1 hour 15 minutes against 2.5 hours. A saving of about half is plausible, but record your own numbers in the log.

What to record from the pilot

• Time taken, with and without the tool, for the same task type.
• Errors found in review, by type (meaning changed, term missed, tone).
• Whether the reviewer would have preferred writing from scratch.
• Anything the policy did not anticipate.

This is the evidence for week four. A practice with no numbers can only argue from opinion.

A filled-in policy page you can adapt

A blank template gives little to work from, so here is a short filled-in version for the illustrative four-person practice. Change every line to fit your own rules, and have a qualified adviser review it.

AI use policy, version 1, dated at the start of week one. Approved tool: one business-plan assistant, named by the AI lead. No other AI tool is approved for work. We do not enter client names, matter details, privileged material, or personal data into any AI tool until the AI lead has recorded that the tool has been assessed for that use. Every authority, quotation and statement of law produced with AI help is checked against a trusted source by the lawyer responsible before use, and the check is recorded. The signing lawyer is responsible for the whole document. We log AI-assisted tasks on the matter file. The policy is reviewed after four weeks and then every quarter.

It runs to under 130 words, which is the point: a policy of eight pages that nobody reads protects nobody. If a rule cannot be said in a sentence, it will not be followed on a Friday afternoon.

Making a prompt safe: before and after

Week one's confidentiality rule works only if people can see what a compliant prompt looks like. This is the same request written carelessly and then written to the policy. Both are illustrative.

Careless: "Summarise the attached email chain between our client and the other side about the boundary dispute at the client's address and say what we should reply."

Compliant: "I am a lawyer drafting a reply. In a boundary dispute between neighbours, one side has offered to move a fence at their own cost if the other agrees not to claim for earlier damage. List the questions I should ask my client before answering, and the points a reply should cover. Use general principles only and flag where the answer depends on the local rules."

The second contains no party, address or private fact, and the answer it produces is a checklist of questions, which is where a general assistant helps most. The lawyer then applies the checklist to the real file in their own system. This shift, from "analyse my matter" to "give me the questions to ask about this kind of matter", is the most useful habit a small practice can adopt in month one, because it gets value without moving a single confidence.

A second pilot candidate: the intake summary

Once the leaflet pilot has passed review, the next step up is turning a client's intake form into a structured summary. The risk rises, because intake forms hold confidences, so use it only after the tool has been vetted in writing for that use, and after the client has been told in the engagement letter. Illustrative example: a new wills client completes a form listing assets, family members and wishes. The paralegal works in the practice-management system's own AI feature, which is covered by the vendor terms the practice checked in week one, and asks for a one-page summary with headings: family, assets, wishes, open questions.

The associate reads the summary against the form. A realistic fault: the summary lists a "second property" the client mentioned as a holiday let, but the form says it is held in a company. The error would change the advice, so it is caught only because the associate reads the original, not the summary. Record this in the log as an example of why reading the source is not optional. Time saved in the illustration is about 20 minutes per intake, against 5 minutes of added checking, and the associate prefers to see the open-questions list first.

Week 3: Verification, especially of citations

This is the week that matters most. Courts in several places have sanctioned lawyers who filed documents containing case citations that an AI tool invented. The citations looked right: plausible names, plausible reporters, plausible years. The mistake was treating a good-looking citation as an existing one.

The verification routine

1. Extract. List every authority, statutory reference and quotation in the AI-assisted draft in a table.
2. Find. Locate each in a trusted source: the court's website, an official statute database, or your licensed legal research platform. Do not use the AI tool to confirm itself.
3. Open and read. Open the authority. Confirm name, citation, court, date, and that it says what the draft says it does. For a quotation, compare word for word.
4. Check the status. Confirm the authority is still good law: not overruled, not superseded, not amended. Your research platform's citator does this. Lexis+ with Protégé, for example, includes Shepard's citations for the status and treatment of an authority, and its product page lists Shepard's Verify, which checks citations in AI answers and drafts and flags unsupported statements.
5. Record. Initial the table, with the date. Keep it on the matter file.

A trap to teach the team: a real case cited for a proposition it does not support. A tool may give a genuine case name with a summary that sounds right and is wrong. Only opening the authority finds this.

A practice exercise for the team

Give each person a short AI-generated memo on a simple point of law with five citations, two of which you have made up and one of which is real but misdescribed. Time them as they verify. In this illustration, the associates find the two invented authorities within ten minutes, and nobody catches the misdescribed one until they open the case. That result is the case for the routine. Repeat the exercise at the quarterly review with a different memo.

The same discipline extends beyond citations. Verify dates, thresholds, deadlines, monetary limits and any statement of what the law "requires". Where a deadline matters, check it against the rule itself and then against your own calendar system, and never put a limitation or filing date into the diary from an AI answer alone.

Week 4: Client disclosure and the review

Disclosure and engagement letters

Decide what clients are told. Rules and expectations vary by regulator and by client, so start by reading your professional body's guidance. A reasonable minimum is a short clause in the engagement letter, such as: "We may use AI-assisted tools to support research and drafting. We do not enter your confidential information into tools that are not approved by the firm and assessed for confidentiality. A lawyer reviews all work before it reaches you." Where a client's instructions or a regulator restrict AI, follow them. Ask a qualified adviser to review the wording. For some matters, especially where AI use affects the billing basis, say so explicitly and agree how time saved is passed on.

If a client asks directly whether AI was used on their matter, the answer should be available from your log and should be honest. That is the reason for the record-keeping in week one.

The four-week review meeting

Hold a 45-minute meeting with the whole practice. Use a fixed agenda and a simple scorecard:

Illustrative scorecard entries, but notice the near-miss. A near-miss in the first month is information, not a disaster: it shows where the habit needs a prompt, such as a pinned reminder at the top of every saved prompt that reads "No client names or matter details".

After week four: what to add and what to leave

With the policy, the pilot and the verification routine working, add one task at a time, in this order of rising risk. First, internal and marketing work (done). Second, administrative drafting on client matters: letters, attendance notes summarised from your own dictation, with confidences handled under the approved tool's terms. Third, first-pass document review or summarising for a matter, with a lawyer reading the underlying documents for anything that matters. Last and slowest, research and drafting of advice, where every statement of law is verified as in week three.

Leave alone, for now, anything where an error is costly and hard to see: limitation calculations, conflict checks done by AI alone, advice to clients without lawyer review, and any filing without citation verification. You can still use AI to prepare for those jobs; it must never be the final check. How a firm in a specific area builds on this is in the first-month setup for a two-lawyer estate planning firm, which follows a similar sequence.

Four ways a small practice's rollout stalls

The policy nobody reads. It sits in a shared folder. Fix: put the one-page version on the wall by the printer, and put the confidentiality line at the top of every saved prompt.

The enthusiastic associate. One person starts using a personal account for speed. Fix: make the approved tool easy to reach, and address the reason: if the approved tool is slower, find out why.

The skipped verification. Under deadline pressure, the table is not filled in. Fix: the partner signing the document asks to see the table, every time, for the first three months.

The vanishing lead. The AI lead gets busy and the review is cancelled. Fix: put the monthly review in the diary now, and pair the lead with a deputy.

If you want help turning this into a plan for your own practice, with the first task chosen and the policy drafted around how you actually work, that is what my AI implementation consultation is for.

Questions a small practice asks before starting

Which task should a law practice try first?

Choose work with no client-confidential content and low consequence if wrong: rewriting your own firm's website copy, drafting an internal checklist, or summarising a public document. Move to client matters only after the policy exists and the team has practised the verification step.

Can I use a free chatbot for client matters?

Avoid it. Free and personal accounts may use your inputs differently and give you less control over retention. Use a business plan, or a legal-specific product, whose terms you have read for confidentiality, training and retention. Your regulator's guidance on confidentiality still applies.

How do I check an AI citation?

Find the authority in a trusted legal database or the court's own site, open it, and confirm the name, citation, court, date and that it says what the draft claims. Never accept a citation because the format looks right. Record who checked and when.

Do I have to tell clients I use AI?

Rules differ by regulator and by engagement. At minimum, check your professional body's guidance and your client agreements, and consider a short statement in your engagement letter. Where AI use affects fees or confidentiality, disclose it clearly.

What if a partner wants to skip the plan?

Offer a smaller pilot, not an exemption. A one-task pilot with the same verification rule respects their time and still protects the practice. The risk of one unreviewed filing outweighs any speed gained.

Further reads

• AI Readiness Checklist for Accountants, Solicitors, Consultants — A readiness checklist for professional firms.
• AI in Planning Law: Drafting Representations and Appeal Statements — Drafting representations and appeal statements with AI.
• Best Legal AI for Family Law: What to Pay For and What to Skip — What a family law practice should pay for or skip.
• Best AI for Immigration Lawyers: Intake to Filing, Stage by Stage — Immigration work from intake to filing, stage by stage.
• Best Legal AI for Criminal Law: A Small Defence Firm's Shortlist — A small defence firm's shortlist.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI ChatGPT Business pricing page, Claude pricing page, Clio pricing page (archived 20 October 2026), LexisNexis Lexis+ with Protégé product page, the AI Hallucination Cases database of court decisions; your professional regulator's guidance on technology and confidentiality. Rules differ by jurisdiction; ask a qualified adviser.