A team attends an AI demonstration on Monday. By Friday, someone has pasted a customer email into a chatbot to draft a reply. The demo showed what the tool could do, but did anyone explain which account to use, what information could be shared, or how to check the draft?
A useful training session answers those questions in the context of people’s work. It should give employees a chance to practice, make a mistake in a safe exercise, and leave knowing where to get help.
Choose a task people recognize
Start with one task the group does regularly, such as turning meeting notes into a draft action list or preparing a first response to a routine inquiry. Ask the people doing the work where the task gets difficult and what a good result looks like. A support team and a finance team may need different exercises, even when they use the same assistant.
Keep the first exercise narrow. For example, give participants a fictional customer question and an approved sample policy, then ask them to draft a response. Include a detail that changes the answer, such as a return deadline, and a question the policy does not answer. That gives the group something concrete to check.
Use fictional or approved training material. Removing names alone may not make a record anonymous; other details can still identify someone, as the UK Information Commissioner’s Office explains. Ask whoever manages privacy and security which material and accounts are suitable before the session.
Teach the boundaries before the prompt
Tell staff which tools and accounts are approved for the exercise and what types of information may be entered. Explain that approval can depend on the account, its settings, the intended use, and the sensitivity of the material. A personal account and an organization-managed account for the same product may have different terms or controls.
If the answer is unclear, give people a route to ask before they paste. For example: “Use the fictional cases in this workshop. For a customer record, check our data-handling guidance or ask the privacy lead first.” Avoid promising that a vendor never retains data or uses it for model improvement unless your organization has verified the relevant agreement and settings.
The NIST generative AI profile discusses privacy and information-security risks involving third-party integrations and recommends guidance for people using generative AI. For organizations subject to UK data protection law, the UK Information Commissioner’s Office also explains considerations for using an AI tool to draft email responses. Requirements depend on the circumstances and jurisdiction, so involve your privacy or legal advisers when needed.
Practice giving context and checking the answer
Have participants write an instruction that identifies the task, the intended reader, and the approved source. For the support exercise, it might say:
Draft a reply to the customer using only the sample returns policy provided. Mention any deadline that applies. If the policy does not answer a question, say what needs a human decision. Do not send the reply.
Then compare the draft with the policy, line by line. If the policy says returns are accepted within 30 days, check that the draft does not promise 60. Look for invented exceptions and answers the policy cannot support. A fluent answer can still be wrong. Give participants time to find and correct problems before walking through the review together.
Try the same exercise again with a missing detail or a conflicting instruction. Discuss when the right next step is to ask a person rather than accept a plausible answer. This is a good place to use the team’s own review checklist, if it has one. Our guide to testing AI answers covers how to build a broader set of examples before relying on a workflow.
Make human review a job someone can do
“A human checks it” is too vague to run a process. Decide who reviews the output, what source they compare it against, and whether they can fix an error or need to escalate it. The reviewer also needs enough time and authority to reject the draft.
For a customer reply, the person sending it might check the policy, verify case-specific facts in the approved system, and remove unsupported promises. If the policy is ambiguous, they should follow the existing escalation process. For a consequential decision about a person, an organization may need a different process and specialist advice; a quick read of an AI summary is not a substitute for that assessment.
Check what changed after the workshop
End with a short exercise participants complete on their own. Ask them to identify an unsafe input, spot an unsupported claim, and explain where they would send an uncertain case. Review the answers together. A quiz can reveal gaps in the session, but it does not prove that the workflow is safe in production.
Follow up after people have tried the approved workflow. Ask what they used, where review took time, and what was confusing. Track recurring errors or questions without collecting sensitive prompts unnecessarily. Update the examples and guidance when the tool, its settings, or the underlying process changes.
A one-hour workshop can be a useful start. On the next workday, employees should be able to find the approved tool, recognize information they should not enter, and know who can help when an answer is uncertain.
If your organization wants help designing an exercise around its own work, explore our AI training and AI strategy services.