AI Training for Solicitors' Firms: A One-Hour Team Session Plan

A small firm can run a useful one-hour internal AI session by spending 10 minutes on rules, 15 on a redacted live demo, 15 on group checking, 10 on a role-based exercise and 10 on a quiz and commitments. Use only an approved tool and synthetic or properly redacted material. A solicitor supervises every legal output.

The session should teach a repeatable decision: is this task allowed, is this data safe for this tool, and who verifies the result? Professional guidance on AI in legal work emphasises confidentiality, human verification and effective supervision. A polished output does not show whether a citation exists or whether a private fact was sent to an unsuitable service. The manager should leave with a rule sheet and a 30-day check, not merely a lively demonstration.

Prepare the room, tool and sample file

Choose a manager, a solicitor who can explain supervision, and someone to record questions. Eight to 12 participants is manageable for one shared exercise; a larger team can work in small groups with a named reviewer in each. Book a screen everyone can read and a way to display a source document beside the AI output. Tell staff in advance that the training uses a synthetic file and that real matter material must not be pasted into the demo tool.

Pick one approved tool that the firm has already checked against its contract, privacy terms, access settings and internal policy. If the firm has not completed that review, use a synthetic example in a trial account and make the approval decision a follow-up action. OpenAI says it does not train its models on ChatGPT Business data by default, but that fact alone does not decide whether a firm may upload client information. The firm's own confidentiality and supplier review still govern use. The solicitors' software buying checklist covers that separate decision.

Create a one-page fictional matter pack. For example: “Matter 014, service contract dispute. Supplier says notice was received on 4 April; customer says it was sent on 2 April. Contract extract says notice must be received by 5 April. The file contains two emails with dates and a short call note. One email is a duplicate.” Do not borrow real names, addresses, client documents or distinctive facts and call them redacted. A truly synthetic pack is safer for a live room and lets the trainer plant deliberate errors.

Prepare the answer key before the session. In the fictional pack, one email shows a sent date but no proof of receipt, and the duplicate creates an apparent third item if counted carelessly. The correct training answer is a chronology of what the pack states, with the missing receipt evidence flagged. The AI should not opine on the legal merits. Keep the answer key hidden until the group review.

Have a fallback if the live tool is unavailable. Save a screenshot or plain-text copy of an output generated earlier from the same synthetic pack, clearly marked “illustrative training output”. Then run the source-checking exercise on that output. Do not improvise by moving the pack into a different unapproved service because the preferred tool is slow. The lesson is the verification process, which works just as well with a prepared output. Record the tool and date used for the example so nobody mistakes it for a current product test.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Use a 60-minute agenda with visible outcomes

Display a countdown. The manager can spend the whole hour explaining AI in general and still leave staff unsure what to do tomorrow. Keep general theory to the minimum needed for the decision rule. The first ten minutes should cover: never use unapproved tools for client-sensitive input; restrict data to the minimum needed; ask for source references; verify every legal proposition and citation in authoritative material; and apply the firm's existing supervision route. Staff should know where to raise an uncertain case.

Say explicitly that roles differ. A support colleague may use an approved tool to format a synthetic chronology or draft an internal checklist under instruction. A solicitor determines legal relevance and signs off client-facing advice. An administrator who manages access should understand retention and sharing settings. The session is not an invitation for every person to choose their own AI account. If the firm needs a deeper confidentiality discussion, link staff to the solicitor confidentiality tutorial after the session.

Run the live demo as a source-checking exercise

At minute ten, paste only the synthetic pack. Ask the tool for a chronology with columns for event, stated date, source, uncertainty and question for the solicitor. Instruct it not to supply law or invent missing evidence. A prompt that demands provenance is more useful for training than “summarise this case”. It gives participants a way to inspect the result rather than judging how polished it sounds.

Using only the fictional file below, make a chronology.
For every event, give the exact source document and date stated.
Mark duplicates, missing evidence and ambiguous receipt dates.
Do not add legal conclusions, citations or facts from outside
the file. Write "not shown" where the file is silent.

An illustrative AI output might say: “2 April: notice received by supplier, source customer email.” The pack actually shows that the customer sent an email on 2 April; it does not prove receipt. Pause the demo and ask the room to identify the unsupported verb. The corrected chronology says “2 April: customer email sent; receipt not shown.” This is a concrete lesson in how fluent wording can turn an evidential gap into an apparent fact.

A second illustrative error is duplicate counting. If the same email appears once in a mailbox export and once as a PDF attachment, the model may list two notices. Show participants the matching message ID or identical text and ask them to mark one copy as a duplicate. Explain that an AI tool can help find likely duplicates, but a person checks whether they are truly the same communication. The exercise transfers to bundles and transcripts; the safe bundle-summary workflow explores that task in detail.

Do not let the demo end at “the AI got two things wrong”. Show a correction loop: point to the original source, amend the chronology, record the uncertainty and decide who owns the next enquiry. If a participant asks the tool to supply a legal authority, demonstrate the safer response: locate the authority in the firm's approved research source, read it in context and have a solicitor assess its application. Never treat a citation in generated text as proof that the authority exists or supports the proposition.

Let the room challenge a good-looking sentence as well as an obvious mistake. An illustrative draft says, “The supplier acknowledged the notice on 4 April.” The call note actually says a staff member “discussed the message” without confirming which message or when it arrived. Ask what the source supports, what it does not, and what further evidence would resolve the ambiguity. This avoids a common training failure: staff learn to hunt for absurd hallucinations but miss modest overstatements that look normal in professional prose.

Give staff a one-page rules sheet

Keep the sheet on the firm's intranet or in the matter workflow, not buried in training slides. Use four boxes. Allowed tasks lists approved examples with the tool and role. Blocked inputs lists client-sensitive material and any other data the firm has not approved for the tool. Review gate says who checks what before the output enters a matter file or reaches a client. Escalation names the person to ask and how to report a mistaken upload or suspicious output.

Written out in full, an illustrative rule could say: “Task: create a chronology from a synthetic training pack. Tool: approved business workspace. Input: synthetic text only. Output: draft working document. Reviewer: supervising solicitor checks against each document. Prohibited: real client pack until the tool and matter use are authorised.” A second rule could permit drafting a neutral appointment reminder from fields already approved for email, with a person checking recipient, date and attachment before sending. These are examples of how to write a rule, not permission for every firm.

Use a decision tree in spoken form: “Would this input identify a client or reveal their matter? If yes, is this specific tool approved for that material and role? If not, stop. If yes, is this task on the permitted list? If uncertain, ask the supervisor. Before using the output, can you check each claim against a reliable source?” Staff can remember that sequence. A long policy with no worked example is less likely to change behaviour.

Include an incident route. If someone pastes a document into the wrong account, they should preserve enough detail to investigate, stop further use and tell the named internal lead promptly. They should not ask the model to “forget” the file and assume that solves retention. The firm's adviser should decide the necessary legal and contractual response. Training should make reporting an error straightforward, because concealment makes a data problem harder to contain.

Print a decision card for staff desks if the firm works with paper files. One side can show the three questions “Approved task? Approved data and tool? Named reviewer?” The other can show the escalation contact and the current policy location. Do not print confidential examples on the card. When a support colleague receives a late request to “put this into AI quickly”, the card gives them a concise way to pause and ask for approval without having to recall a slide from the session.

Make the group exercise specific to each role

At minute 40, hand each group one short scenario. A solicitor receives an AI-drafted letter with two case citations; a paralegal has a 40-page bundle to organise; a receptionist wants to enter a client's phone message into a public chatbot; an accounts colleague wants to classify matter expenses. Ask each group to identify the input tier, approved tool, useful output, reviewer and stop condition. The point is to practise the rule sheet under realistic pressure.

For the letter scenario, the correct response is to open and verify both cited authorities and the relevant proposition before using the draft. If one citation cannot be found, the letter is not ready. For the bundle, a checked index and chronology may help, but the supervisor chooses what is legally material. For the phone message, a public chatbot is not an approved place for a client's account of a dispute. The receptionist records it in the firm's authorised system and asks the supervisor how to process it.

The accounts scenario (illustrative) has a receipt for $240 that the AI labels “court fee” because the word “filing” appears in a note. The receipt is actually for document storage. The accounts colleague checks the supplier and matter coding before posting it. This example shows that verification is not reserved for legal research; routine records can also create client billing errors. Give the group two minutes to state the wrong route and the corrected route, then collect the answer.

Use one boundary that staff are likely to test: “Can AI draft a first version of a client letter?” The answer depends on the approved tool, input, matter risk and a solicitor's review. It is not a blanket yes. The legal tasks that should stay with people tutorial can help managers define the blocked side of the sheet without turning the hour into a product comparison.

Close with a quiz that tests decisions

Use five questions, answered individually in the last ten minutes. One: “A tool gives a citation; what do you do before using it?” Answer: find the source and check the proposition in context. Two: “Can a personal account receive a real client bundle because it has a privacy switch?” Answer: only if the firm has specifically approved that tool and use; otherwise stop. Three: “Who signs off a client-facing legal conclusion?” Answer: the responsible solicitor under the firm's supervision process.

Four: “A synthetic chronology says a notice was received when the file proves only sending. What changes?” Answer: correct the event and mark receipt unproved. Five: “You uploaded material into the wrong workspace. What now?” Answer: stop, report through the internal incident route and follow the firm's instructions. Collect answers without turning the quiz into a blame exercise. Review wrong answers in the room; they show which rule needs clearer wording.

For an illustrative eight-person firm, suppose six staff answer all five correctly and two miss the data-approval question. The manager should not declare the session complete because the average score is high. Clarify that specific tool approval is required, give the two staff a scenario to practise, and check again. Record the training date, attendance, version of the rules sheet and the questions that required follow-up. The goal is reliable behaviour, not a ceremonial score.

Check changed behaviour after thirty days

Assign a manager to inspect a small sample after thirty days. Ask how many approved tasks were attempted, how many drafts needed material correction, whether any unapproved tools were used for client data, and whether the named reviewer can trace outputs to sources. Compare the results with the session's promises. If staff did not try the permitted task, ask whether the route was too awkward or the rule unclear before concluding they resisted AI.

Run the numbers on an illustrative pilot with eight staff, each allowed to test one synthetic or approved administrative task per week for four weeks: at most 32 task attempts. If 20 attempts occur, 16 produce usable drafts after review and four are rejected for unsupported facts, the team has a useful correction pattern. Do not call that an 80% accuracy rate for legal work; the tasks were selected and the reviewer caught the errors. Classify the four failures and improve the prompt, input or blocked-task rule as appropriate.

Count the true staff cost of the hour. Eight people in a 60-minute session use eight staff hours, plus manager preparation and review. If preparation takes two hours and the thirty-day audit takes one, the illustrative investment is 11 hours before any tool fee. That is reasonable to compare with correction time and avoided rework, but do not promise a return from one class. The question at day thirty is whether the firm has a safe, repeatable route for a small number of tasks.

Refresh the rules sheet when the tool, supplier terms, professional guidance or firm's work changes. An internal one-hour session cannot make every future AI use safe. It can give solicitors and support staff a shared language for deciding what to try, what to check, and when to stop and ask the person responsible for the matter.

At the thirty-day review, choose one rejected output and walk it back through the workflow. Was the input incomplete, the prompt too broad, the wrong task attempted, or the human review too late? In one rejected letter (illustrative), the model invented a deadline that was absent from the file. The team adds “do not infer deadlines” to the permitted draft instruction, but also requires the reviewer to compare every stated date with the matter record. A prompt change alone cannot remove the duty to check. Feed this example into the next induction session so new staff learn from a real process finding without seeing a client's information.

Further reads

• AI Readiness Checklist for Accountants, Solicitors, Consultants — Check whether tools and records are ready before staff practise.
• How Much Time and Money Does AI Staff Training Take? — Budget staff time and follow-up for a larger programme.
• AI Literacy Requirements: What Your Staff Need to Know — Map baseline knowledge by role after the first session.
• Law Firm AI Programs: Training Staff on a New Tool in Four Weeks — Extend one hour into a supervised four-week rollout.
• ChatGPT or a Legal AI Tool: Which Should a Small Firm Use? — Choose the tool category after defining safe tasks.
• Best AI for Solicitors in the UK: Matched to Each Task — Match AI tools to research, drafting, client emails and contracts in a small UK firm, with list prices, data-location answers and a supervision record.
• ChatGPT for Solicitors: Five Safe Tasks and Three to Avoid — Five tasks a solicitor can safely give ChatGPT, three to avoid, and the settings and client-data rules that go with them.
• Law Firms AI UK: What Solicitors Should Check Before Using AI — A pre-use checklist for solicitors' firms: eight checks with how to verify each, a go or no-go sheet and the evidence to keep on file.
• AI Training Programme for a Company: A Four-Week Staff Rollout Plan — A week-by-week plan for training your whole team on AI: sessions, homework tasks, a staff rules sheet and a month-end check.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: The Law Society, Generative AI – the essentials; OpenAI, Business data privacy, security and compliance (both checked 4 November 2026).