Law Firm AI Programs: Training Staff on a New Tool in Four Weeks

Roll out a new AI program in four weeks: week one baseline and rules, week two a pilot with two or three people on set tasks, week three review and correction with a supervising partner, week four measure and decide whether to widen. One named partner owns the plan, and every output is checked before it leaves the firm.

The tool matters less than the sequence. Firms that hand everyone a licence and a policy PDF get a few enthusiasts, a few quiet refusers and a risk nobody is watching. A short, supervised pilot gives you evidence, trained champions and rules written from real mistakes instead of fear.

Follow me on Instagram@sagnikteaches

Why a short, supervised plan beats a firm-wide launch

Regulators agree that AI does not change who is responsible. The AI warning notice that one solicitors' regulator published on 17 August 2026 works as a checklist even if your own regulator has not written one yet. It expects firms to have governance, systems and controls to manage AI risks; solicitors to keep the knowledge of the people they manage up to date and to supervise work for clients effectively; every case put to a court to be genuine, relevant and backed by a verifiable citation (it quotes a court saying that reliance on AI output is not a defence); and client information to go into AI tools only with safeguards, such as not being used to train models unless explicitly authorised. Its summary line is that the use of AI "does not diminish or transfer your professional responsibilities". Read the notice, or your own regulator's equivalent, before you finalise your rules.

Connect on LinkedInSagnik Bhattacharya

That list is practically a rollout plan. Governance means written rules and a named owner. Training means a session for every role, not a link to a video. Supervision means a named person watching. Verification means a review step with a checklist. Confidentiality means deciding in advance what may be pasted where. Accountability means no output goes out unchecked. A four-week structure makes each of those a visible task with an owner and a date.

Subscribe on YouTube@sagnikteaches

Before week one: pick the owner, the tasks and the numbers

Three decisions come first, and they take an afternoon.

The supervising partner. One partner owns the rollout. Their job is to approve the rules, sit in on the pilot review, decide what is in and out of scope and report to the other partners at the end. Give them two hours a week in their diary. In a firm with four fee earners it is usually the partner who already deals with risk and compliance.

Two or three tasks. Choose tasks that are frequent, low risk and easy to check. For a law firm that might be summarising long email chains for a file note, first drafts of routine client letters, and turning a meeting recording into an attendance note. Avoid anything where a wrong answer reaches a client or court without a person reading the whole document. Our guide to what small law firms should automate first helps rank candidates.

A baseline. Time five real examples of each task by hand, now. Without that number you will have opinions in week four and not evidence. Write down minutes per task and the number of corrections a supervising fee earner usually makes.

Week one: write the rules on one page and set the baseline

The rules should fit on one page and be written by the supervising partner, with the pilot group's comments. A longer policy comes later, after the pilot has shown what it needs to cover; for a template see our acceptable-use policy for a professional firm. For week one, six rules are enough. An illustrative version:

1. Use only the firm-approved AI program, signed in with your firm account.
2. Client names, matter numbers and anything identifying a client go in only where the
   program's settings have been approved for it. When unsure, anonymise first.
3. Every output is a draft. A fee earner reads all of it before it is used or sent.
4. Every case, statute, quotation or date is checked against the original source.
5. If you used AI on a document, record it on the matter file (one line is enough).
6. If something looks wrong, tell the supervising partner the same day. Nobody is blamed
   for reporting a bad output; hiding one is the problem.

Rule 2 needs real thought. The program may have different data settings by plan, and your supplier's terms control whether prompts are retained or used for training. Check them before the pilot, not after: whether prompts and uploads are kept, for how long, whether they can be used to train models, and whether an administrator can switch that off for every account. Write the answer to each in one line on the rules page, so the pilot group knows what the setting is rather than guessing.

Also in week one, set up the program accounts for the pilot only, and run a 30-minute walkthrough of the six rules. The pilot group is the two or three people who volunteer, mixed across roles. In a four fee-earner firm, that might be one partner, one associate and the practice manager or a paralegal. People who are sceptical but fair are more useful than evangelists.

Week two: the pilot, with prompts the group writes together

The pilot group uses the program only on the chosen tasks. The first session, about an hour, should have the group working on real but low-risk material, with the supervising partner present. Spend the time on three skills in order: giving the program clear context, asking for a specific output and checking the result.

Here is a starter prompt for the file-note task, illustrative and with client details removed:

You are helping a solicitor write an internal file note. Below is an email chain
with identifying details removed. Write a file note of no more than 200 words
covering: what was agreed, what is outstanding, and who is to do what by when.
If a date or amount is unclear, write "check source" and do not guess.

[paste anonymised email chain]

An illustrative result might list three agreed points and two outstanding actions, with one line reading "completion date: check source" because the chain gave two conflicting dates. The trainer's point is the last line: a good prompt gives the program permission to say it does not know. Then the group checks every line against the emails, which is the habit you are training. Our one-hour team session plan gives a fuller agenda for this meeting.

For the rest of the week, the pilot group works through ten real items per task, logging for each: time taken, number of corrections they made and whether the output was usable. A simple shared sheet with those columns is enough.

Row 4 is the one to prize. The program invented a deadline; the checker caught it; the group now has a real example for the rule on verification. Keep a "gallery" of such failures and show it in week three. It teaches more than any slide.

Week three: the review routine and what the partner looks at

Week three is about making checking systematic, not ad hoc. Each output goes through a short checklist before use. An illustrative one-card version for fee earners:

• Facts: names, dates, amounts, addresses and references all match the source documents.
• Law: any case, statute or quotation has been found in the original source, not trusted because the program cited it.
• Gaps: nothing in the output claims something the file does not support.
• Tone: it sounds like the firm and is right for the client.
• Confidentiality: nothing is in the output that should not go to this reader.
• Record: a one-line note of AI use has been added to the matter file.

The supervising partner then does two things. First, a 45-minute review session with the pilot group where they look at the week-two log, discuss each failure and update the six rules. Second, a spot check: pick five outputs at random from the pilot and review them as if they were going out under the partner's name. Note where checks were missed. If reviewers did not catch an error that the partner finds, the checklist needs to change or the task is not yet safe to widen. For the case-law check in particular, the routine in how to stop AI inventing case law is worth teaching verbatim.

Week three is also the time to deal with how people feel. Ask each pilot member three questions in confidence: what saved you time, what annoyed you, what did you almost miss? The third question surfaces the near-misses that never reach a log.

What each role needs to learn, and what to show them

One training session for everyone wastes time, because the risks differ by role. Keep the shared hour for rules and checking, and add ten minutes per role.

Notice that the partner exercise is about spotting errors, not about typing prompts. The people who sign work off need the checking skill most, and in many firms they are the ones who have touched the program least.

A before and after to show in the training hour

Showing a weak prompt next to a strong one persuades faster than explaining prompting. Use something like the following, which is illustrative and anonymised.

Weak prompt: "Write a letter to the client about the delay."

Result: a generic apology that blames "unforeseen circumstances", promises "a swift resolution" and invents a new date. Nobody can send that.

Better prompt:

Draft a letter of under 150 words from a solicitor to a client. Facts, and use only these:
the other side's replies are three weeks overdue; we sent a chaser on [date];
we will chase again on [date]; completion cannot be fixed until they reply.
Do not give any new date or promise an outcome. Tone: calm and direct.
Leave [client name] and [your name] as placeholders.

Result: a short letter that states the delay, the chaser dates and that no date can be fixed yet. The fee earner changes one phrase and sends it. Ask the group what made the difference, and write their answers on the board: facts supplied, word limit, a ban on inventing dates, placeholders for anything unknown.

Handling the objections you will hear in week two

Resistance is information. The four objections below come up in most firms, with a response that respects the point:

• "It will take longer to check than to write." Sometimes true. Use the pilot log to find the tasks where it is not: summaries of documents the reviewer already knows, or first drafts of standard letters. Drop the others.
• "I do not trust it with client data." A reasonable instinct. Show the data settings you checked, the rule on what may be pasted and the anonymising exercise. Where a client has asked that no AI be used on their matter, record it and honour it.
• "My clients pay for my judgement, not a machine's." They do, and the program drafts while the fee earner judges. Show the file-note example, where the machine's output is an input to judgement and not a substitute.
• "What if I get blamed for a mistake it made?" Say plainly that responsibility sits with the person who signs the work off, which is why checking is built into the job, and that reporting a bad output early is rewarded.

What the four weeks cost in time

A supervising partner needs a rough budget so the plan does not get squeezed. Illustratively, for the four fee-earner firm above:

• Week one: partner 3 hours (rules, baseline planning); pilot group 1 hour each for the walkthrough.
• Week two: pilot group about 3 hours each across the week, including the training hour and logging.
• Week three: partner 3 hours (45-minute review session plus the spot check of five outputs); group 1 hour each.
• Week four: partner 2 hours measuring and preparing the decision; 1 hour with all partners.

That is about 11 partner hours (nine for the supervising partner, one in the week-two training hour and one for the other partner at the decision meeting) and 15 hours across the pilot group, or roughly 26 hours in total. At any sensible internal hourly cost that is the real price of the rollout, not the licence, and it is a modest price for evidence that the program saves time without raising risk.

Week four: measure, decide and widen in stages

Pull the numbers. For each task compare minutes with AI against baseline minutes, count corrections per item, and count how many items were unusable. Add one usage figure: how many pilot members used the program in each of the last two weeks.

The 25% figure is a starting suggestion, not a standard; pick a threshold that covers your licence cost plus the time spent checking, and agree it with the other partners before you look at the results, so nobody moves the bar after seeing the numbers.

If the decision is to widen, do it in stages: the pilot group trains the next group, using the failure gallery and the one-page rules, over the following fortnight. Each person trained does their first five items with a pilot member beside them. Do not announce firm-wide availability and walk away.

A worked example: a law firm with four fee earners

Illustratively, take a firm with four fee earners (two partners, two associates), a practice manager and a paralegal, with a new AI program on a firm-wide licence at a cost of, say, $30 per user per month for the whole team of seven, which is $210 a month. The two tasks chosen are file notes from email chains and first-draft routine letters.

Baseline. Five file notes average 24 minutes each; five letters average 21 minutes with an average of four corrections from the reviewing partner.

Pilot. One partner, one associate and the practice manager volunteer. Over a week they run 12 file notes and 8 letters. File notes average 11 minutes (a saving of 13), with 1.5 corrections each. Letters average 17 minutes (a saving of 4) with 5 corrections each, and two of eight were unusable, one with an invented deadline.

Decision. File notes show a 54% time saving with stable corrections, so they go firm-wide. Letters show a 19% saving and rising corrections, so they stay in the pilot with a rule that a partner reviews any AI-assisted letter that states a deadline. At 40 file notes a month firm-wide, the saving is about 40 × 13 = 520 minutes, or nearly nine hours, against a licence cost of $210. If the fee-earner time is worth far more than that, the cost is not the issue; the controls are.

All numbers are illustrative. The shape to copy: measure before, pilot narrow, decide task by task and keep the supervising partner's spot check going after launch.

Realistic ways a rollout stalls

• Nobody owns it after week one. The partner is busy and the pilot drifts. Fix: put the review sessions in the diary in advance.
• The rules are never read. A long policy PDF sits in the shared drive. Fix: the one-page version, discussed aloud in week one.
• Quiet use of other tools. Staff keep using free tools because the approved one is slower to start. Fix: ask in week three what else people use, without blame; the usage audit tutorial shows how to find out properly.
• Time saved is eaten by checking. The checking step takes longer than writing. Fix: narrow to tasks where checking is quick, such as summaries of a document the reviewer already knows.
• A good result gets overtrusted. After ten fine outputs, the eleventh goes out unread. Fix: keep the monthly spot check going after the rollout ends.

If you would like help building the plan around your own matters and team, an AI implementation consultation covers choosing the first tasks and setting up the checks; you can get in touch here. For a wider plan beyond the first tool, the first 90 days for a small law firm picks up where this four-week plan ends.

Questions partners ask before rolling out an AI program

Should everyone be trained at once?

No. Start with a pilot of two or three people from different roles, run it for two weeks, then widen. A small group finds the problems cheaply, writes the examples the rest will learn from and gives you real numbers instead of enthusiasm.

Who should supervise the rollout?

One partner, named in writing, with an hour or two a week set aside. Someone with authority has to decide what is allowed, sign off the rules and say stop when a result is wrong. A committee slows every decision and leaves nobody accountable.

What if some staff refuse to use it?

Find out why before pushing. The usual reasons are fear of looking incompetent, worry about confidentiality and a first bad result. Address those with examples and a clear rule that checking AI output is part of the job. Use is encouraged, not forced, until the pilot data says it helps.

How do we know the rollout worked?

Measure three things against a baseline taken in week one: time on a chosen task, errors found at review and how many staff use it weekly. If time falls, errors do not rise and use holds after week four, keep going. If errors rise, narrow the tasks.

Further reads

• AI Software for Solicitors: Data, Supervision and Supplier Checks — Data, supervision and supplier checks before you choose.
• AI Paralegal Software: What a Small Firm Should Test Before Paying — What to test in a paralegal-style tool first.
• ChatGPT or a Legal AI Tool: Which Should a Small Firm Use? — General chatbot or legal tool for a small firm.
• Shadow AI: Is Your Team Using AI Without Telling You? — Spot staff using AI without telling you.
• How to Roll Out an AI Policy So Staff Actually Follow It — Make a policy stick once the pilot ends.
• AI Training Programme for a Company: A Four-Week Staff Rollout Plan — The same four-week idea for a non-law team.
• What to Check in an AI Tool's Privacy Policy and Terms — What to read in the supplier's data terms before the pilot.
• How Much Time and Money Does AI Staff Training Take? — How much time and money staff training takes.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: SRA warning notice Misuse of AI (published 17 August 2026), read November 2026; no vendor pricing is quoted, and all pilot figures are illustrative.