Using AI in Law: Which Tasks a Small Practice Should Start With

A small law practice can safely give AI summaries of documents it already holds, first drafts of routine letters and notes, and internal admin first. Legal advice, anything filed or sent to a court or counterparty, and any citation should never leave the office unreviewed by a lawyer.

The ranking below runs by what happens if the output is wrong and nobody notices. A muddled meeting summary costs ten minutes. A fabricated case citation in a filed document can cost a reputation. So the rule is not "which tasks are easy" but "which tasks fail cheaply and visibly". Start at the cheap end, build a habit of checking, and only then move up.

Follow me on Instagram@sagnikteaches

The risk ladder in five rungs

Every task a small practice might hand to a tool sits somewhere on one ladder. The rungs are defined by consequence, not by difficulty, because a task can be technically easy for the model and still dangerous to ship unchecked.

Connect on LinkedInSagnik Bhattacharya

Rungs 1 and 2 are where I would put the first month. Rung 3 follows once the team has checked a few dozen outputs and knows how this tool tends to go wrong. Rung 4 is a different kind of work and deserves its own trial, ideally with a research tool built for law rather than a general assistant (see general models versus CoCounsel for that choice). Rung 5 stays human.

Subscribe on YouTube@sagnikteaches

Three rules that sit under every rung

These stay the same whichever task you pick, so settle them before the first document goes in.

Confidentiality comes first. Client information is protected by your duty of confidence and, in many matters, by privilege. Before a single document goes into a tool, find out three things from the vendor's current business terms: whether your inputs are used to train the model, how long they are retained, and where they are processed. Business plans from the main assistants generally do not train on business content by default, while free and personal plans can differ, so check the current settings rather than assuming. If you cannot answer all three, the tool is not approved yet. Strip names, addresses and account numbers from anything you test with in week one. Some regulators now say this in terms: a joint statement from three legal regulators in December 2024 warns that lawyers cannot safely put confidential, sensitive or privileged client information into public chatbots, and that anyone using a commercial tool with client information must review its contract terms first.

A named lawyer owns every output. Supervision does not change because a machine produced the draft. Whoever sends it is responsible for it, and the file should show who reviewed it. A simple line in the file note, "AI-assisted first draft, reviewed and edited by [initials]", is enough and costs ten seconds.

Nothing is cited that has not been opened. Language models can produce plausible case names, section numbers and quotations that do not exist. Any authority in any document is opened in the official or subscription database and read before it is relied on. That rule applies whether the text came from a chatbot, a junior or your own memory.

Rung 1: summarising documents you already hold

The best first job is the one where the source sits next to the output, so checking takes a minute. Say a three-lawyer commercial practice receives a 40-page expert report in a dispute. The associate uploads the report (anonymised, or into an approved business-plan tool) and asks for a structured summary.

You are assisting a solicitor. Summarise the attached report for an internal file note.
Return: (1) the author and date; (2) the five conclusions the report reaches, each with the page number it comes from; (3) any assumptions the author states; (4) anything the report says it could not verify.
Use only the text provided. If a point is not in the text, say "not stated".

An illustrative output might list five conclusions with page references, flag that the author assumed the contract was performed on the original timetable, and note that the author could not inspect the site. What you would fix: the tool will sometimes cite page 14 for a point that sits on page 17, or merge two conclusions into one. The associate opens each cited page, which takes about eight minutes on a 40-page report against perhaps forty minutes to read cold. Keep the page references in the instruction, because they are what makes checking quick.

A second example in the same rung is a long email chain with a client. Ask for a chronology of decisions and open questions, with the date of each email. Before, the partner scrolled thirty messages to find when the client agreed to the revised fee. After, there is a dated list, and she checks the one date she needs against the original. The risk is a missing message, so ask the tool to state how many emails it read and compare that with the count in the inbox.

Rung 2: internal admin that never leaves the building

File notes, time narratives and checklists are low-risk because a person reads them before they matter, and a clumsy sentence does no harm. Voice-to-text file notes are the quickest win. A fee earner dictates a two-minute note after a call, a tool turns it into a tidy paragraph with the next steps in a list, and the earner reads it and saves it. If you use an AI note-taker on live calls the rules are stricter, because it records the client and the other party; see AI note takers in client meetings for consent and retention.

Time narratives are another good fit. Many practices lose billable time because narratives are written days later from memory. A filled-in example: the fee earner types "call w client re: lease break, 0.4, agreed to write to landlord Fri" and the tool expands it to "Telephone attendance on client regarding exercise of the lease break option; advising on notice requirements and agreeing to write to the landlord by Friday." The earner checks that nothing in the expanded version overstates what was done. The sum is simple: if six fee earners each save six minutes a day on narratives, that is 36 minutes a day across the team, roughly 130 hours over a year of 220 working days, before any recovered time. Treat that as an illustration, not a promise.

Rung 3: first drafts of routine outward documents

This is where most practices want to go first, and where I would ask them to wait a month. The reason is not that drafts are risky in themselves; it is that an outward document carries your name. Once the team has a feel for the tool, drafts of client update emails, standard engagement letters and simple covering letters are a sensible next step.

Give the tool your own precedent, not a blank page. A filled-in prompt for a property practice might read:

Draft a client update email in the practice's tone: plain, warm, no legal Latin.
Facts (use only these): searches have been returned, one query raised on drainage, we have asked the seller's side to respond, expected reply within 7 days, exchange cannot happen until it is answered.
End by asking the client whether they want a call. Do not predict the completion date.
Do not give any advice on whether to proceed.

The line "Do not predict the completion date" earns its place. Without it, assistants like to fill the gap with an optimistic date that the client then remembers. A realistic mistake from a trial: the draft said "the seller has agreed to resolve the query" when the facts only said it had been raised. Rung 3 work fails by smoothing over uncertainty, so the reviewer reads for sentences that sound more certain than the facts allow.

Keep a one-page prompt library per task so the instruction is the same each time. Over a few weeks you will see the same corrections repeat; those corrections go into the prompt.

Rung 4: research starting points, with the sources opened

Research is where hallucination does real damage. A general assistant can be useful for mapping an unfamiliar area ("what are the headings I should consider in a retention-of-title dispute?") but it must never be the last word on what the law is. Use it to build a list of questions, then answer those questions in a proper legal database. Tools built for legal research that link answers to sources are a different category; compare them in best AI tools for attorneys and test them in a short scored trial (the two-week scorecard).

A check you can run in five minutes: ask the tool a question you already know the answer to in your own field, with a twist, and see whether it follows the twist or answers a more common version. Repeat with three questions. If it fails one, you know the pattern of errors to look for.

Rung 5: advice and filings, which stay with a lawyer

Advice depends on judgement about this client, this counterparty and this risk appetite. A filing carries a duty to the court. Neither belongs on a tool's output. Permitted assistance here is narrow: proofreading a document you wrote, checking a bundle index for consistency, or reformatting a table. Several courts have issued practice notes or directions about generative AI in proceedings, and some require a declaration. One court's practice note in force since February 2025, for example, bars generative AI from generating the content of affidavits and witness statements, requires those documents to say it was not used, and makes the author of any AI-assisted submission confirm that every authority exists, is accurate and is relevant, a check that cannot be done by an AI tool alone. Look up the current note for every court you appear in rather than assuming.

A realistic mistake: a trainee asked a chatbot for authorities on a limitation point and pasted two into a draft skeleton argument. One case name was real but the proposition was wrong; the other did not exist. A supervising partner caught it because her rule was "open every case". Nothing was filed, and the practice added a line to its policy requiring a citation check initialled by the reviewer.

A one-page policy you can adopt this week

Write the policy before the trial, not after the first incident. A filled-in version for a four-person practice:

1. Approved tools: one business-plan assistant, one note tool, nothing else. New tools are requested from the managing partner.
2. Banned inputs: identity documents, medical records, children's details, and anything under a confidentiality order, unless the tool is specifically approved for it.
3. Review: every output is read by the lawyer who owns the matter; rung 3 and above is initialled on the file.
4. Citations: no authority is cited unless opened in a primary source.
5. Disclosure: the engagement letter says the practice may use AI tools under lawyer supervision, with confidentiality protections.
6. Incidents: anything sent to the wrong place or wrongly relied on is reported the same day to the managing partner.

If your practice management system already includes AI features, ask the vendor the same three data questions. Clio, for instance, says some AI features come with every plan while others, including AI drafting and billing, need select plans or add-ons, and its help centre notes that AI queries may be processed outside the firm's home jurisdiction. That last point matters if your regulator or client terms restrict where data goes. Read the vendor's current pages for your plan rather than relying on a summary. A comparison of the platforms is in Clio, Smokeball and Lawmatics compared.

Measuring whether it worked after a month

Pick numbers you can collect without a project. For rungs 1 and 2: minutes saved per document, counted by timing five tasks with and without the tool; the number of errors the reviewer found per ten outputs; the number of times an output was thrown away entirely. For rung 3: the proportion of drafts sent after light edits versus heavy rewrites. If heavy rewrites exceed half, the prompt or the task is wrong, not the lawyers.

A worked check with illustrative numbers: over four weeks, a team of four logs 60 summaries and 40 file notes. Reviewers found a mistake in 9 of the 60 summaries, mostly wrong page references, and none that would have changed advice. They timed ten summaries both ways and averaged 11 minutes with the tool and 34 without. That supports keeping rungs 1 and 2 and trying three rung-3 drafts a week. If the error rate were 30 percent, or errors changed meaning, you would stop and rewrite the prompts before going on.

Plan a decision date. Four weeks after the start, the partners meet for thirty minutes and answer three questions: what did we save, what did we catch, what do we change in the policy. Without a date, trials drift into permanent habits no one chose.

What to do when a check fails

Checks will fail, and the practice should decide in advance what happens. Take a tenancy practice that tries rung 3 for client updates. In week two the reviewer finds that a draft said a deposit "has been protected" when the file only showed that the paperwork was requested. The fix has three parts. First, the draft is corrected and sent. Second, the prompt gains a line: "Never state that an action is complete unless the facts say so; use 'we have asked' or 'we are waiting for'." Third, the incident goes in a one-line log with the date, the task and what the check caught.

That log is the most useful artefact of the first quarter. After twenty entries you will see that most failures cluster: wrong dates, overstated certainty, invented detail where the facts were thin. Each cluster becomes either a prompt rule or a reason to keep the task a rung lower. It also gives you something concrete to show an insurer or a client who asks how you supervise AI use, without inventing a process after the fact.

A second case from the same practice: the reviewer found nothing wrong in six consecutive file notes and stopped reading them closely. In week five a note recorded a call with the wrong tenant name, copied from an earlier prompt. The lesson is that review fatigue is a failure mode of its own. Rotate who reviews, and keep a rule that a note is read in full, not skimmed, before it is saved.

Three jobs to leave off the list for now

Three jobs tempt small practices and should wait. First, automated client-facing chat that answers legal questions: the answer is advice, and the client cannot tell. Second, drafting pleadings or witness statements from instructions: the voice and the facts must be the witness's own. Third, bulk document review in disclosure without a validated process; that needs a method, sampling and a defensible protocol, which is a project of its own.

If you want a weekly rhythm rather than a task list, a four-week starter plan for a law practice sets it out, and if you would rather map it with someone, that is what my AI implementation consultation covers. Keep this ladder pinned where the team can see it, and move a task up only when the checks on the rung below have gone quiet.

One last practical point: the ladder is per task, not per person. A partner using AI on a rung 3 draft needs the same checks as a trainee, because the risk sits in the output. Review the ladder every quarter, and move a task down a rung the moment something goes wrong with it.

Questions partners ask before the first trial

Do I have to tell clients I use AI?

Rules differ by regulator and by what the tool touches. Many regulators expect you to be open about AI use where it affects the work or the bill, and some engagement letters now carry a short clause. Check your own regulator's guidance, and put a plain sentence in your terms of engagement.

Can a paralegal or trainee run the AI tasks?

Yes, as long as a named lawyer reviews the output and owns it. Treat the tool like a very fast, very junior assistant: the supervising lawyer remains responsible for everything that leaves the office, exactly as with any delegated work.

Is a free chatbot acceptable for the early rungs?

Only with no client-identifying content, and even then a business plan is safer. Free and personal plans often have different data terms from business plans. Read the current terms and the training setting before any client material goes in.

What should the AI policy contain?

One page: approved tools, banned inputs, who reviews what, how citations are checked, how AI use is recorded on the file, and who to tell if something goes wrong. Review it every quarter, because tools and settings change quickly.

Further reads

• Best AI for Solo Lawyers: A Stack That Fits One Person — A tool stack sized for one fee earner.
• AI Billing Software for Lawyers: Which Features Earn Their Cost — Which billing features earn their cost.
• Free ChatGPT for Lawyers: Where the Free Plan Stops Being Safe — Where the free tier stops being safe for client work.
• AI Voice Agent for Lawyers: Cost and Safe Handling of Intake Calls — Intake calls are the next rung once drafting is stable.
• AI Setup for a Two-Lawyer Estate Planning Firm: The First Month — One practice area walked through in its first month.
• AI Time Capture for Solicitors: Recover Hours You Never Billed — Tools and traps for recovering unbilled time.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI and Anthropic business pricing pages (training defaults), Clio's AI features page and help centre, a joint legal regulators' statement on AI (December 2024) and one court's generative AI practice note. Check your own regulator's and court's current guidance.