In an AI training session, I want people to do the task themselves and find a mistake in the answer. Watching a demonstration cannot tell you whether they know what to check.
Some colleagues will have experimented with AI already. Others may be opening a tool for the first time. Start with an explanation of what the firm is asking them to do, which service they can use and where their responsibility for the work begins.
Use an exercise that fits the role
A litigation solicitor might review a draft chronology. A paralegal might compare a summary with its sources. An administrator could prepare an internal handover from fictional enquiry notes. Each exercise should use material the participant can understand and question.
Managers should see the review involved too. If they expect a finished answer as soon as a draft appears, the team will be under pressure to skip checks. Include preparation, corrections and review when discussing the time a task takes.
Teach the request and the review together
Ask participants to describe the task, identify the permitted source material and specify the output. For a summary, they might ask for key events, disputed points and missing information, with references to the documents.
Then put the answer beside those documents. Can they locate the important statements? Has the answer changed a person’s account into an established fact? Is there a qualification on the next page? Asking the tool to check itself may help explore a question, but it does not replace independent verification.
Give them a mistake to find
For a fictional exercise, use a letter dated 8 March that reports receiving a document on 6 March. Give participants a prepared response that records 8 March as the receipt date. Include a disputed account that the response presents as agreed.
Ask them to identify the passages, correct the wording and explain what the answer leaves out. The discussion shows why a response that reads well may still be unreliable. It also gives people a way to raise doubts about an output rather than accepting it because the software produced it.
Practise choosing what to enter
Use fictional or suitably prepared non-confidential material in training. Explain the firm’s approved tools and uses before anyone enters real information. Removing a client name alone may leave enough detail to identify the matter.
Make information handling part of the exercise. Participants should check whether they have the right material, whether access is appropriate and whether the request includes information it does not need. They should know whom to ask if they cannot decide.
Let everyone finish the work
Give participants time to prepare the input, write instructions, check the result and produce a corrected version. Discuss the decisions they made. Two reasonable reviewers may organise a chronology differently while agreeing on the underlying facts.
Feedback should go beyond whether the prompt was well written. Did someone notice a missing attachment? Did they preserve an uncertainty? Did they decide an output needed more work? Leave short instructions that capture those steps and the common mistakes from the session.
Keep help available afterwards
Nominate a colleague to gather questions and arrange support. Users need a route for reporting unreliable answers and a way to distinguish a technical problem from a question for their supervisor or the person responsible for the legal work.
Revisit recurring problems in follow-up sessions. If the software or task changes, update the instructions. Assess what people can now do: complete the task, identify an error and apply the required checks. Attendance alone tells you little about that.
For further practice, visit LawAI Skills Lab, the free desktop legal AI training tool. Its fictional scenarios are for training and awareness, including verification and human judgement.
Talk to me about training your team, whether they are starting from scratch or already have tools in use.