1. Identify the approved environment.
Begin by showing people which application and account they may use. A consumer chat account and an organisational deployment can have different contracts, administrative controls and data settings even when the interface looks similar. Staff need an explicit route to the approved tool rather than a general instruction to “use AI responsibly”.
Microsoft 365 Copilot and Google Workspace with Gemini connect AI features to business productivity environments. Their suitability depends on the organisation's configuration, permissions and contractual arrangements. Use the relevant Microsoft documentation and Google Workspace administrator guidance to check the deployment in question. Do not teach a product name as a guarantee about every account bearing that name.
2. Explain what may be shared.
Translate the organisation's information rules into recognisable examples. Distinguish public material, internal working information, personal data and restricted records. Explain whether each category can enter the approved tool and who can answer a question about an unfamiliar document. Permission to read information does not necessarily include permission to upload it to another service.
Use training material that does not expose customer or employee records. Removing a person's name may leave identifying detail elsewhere in the text. Show people how to describe a task without attaching the underlying confidential material, and give them a clear escalation route when the necessary information falls outside the permitted boundary.
3. Give clear instructions and context.
A prompt is the instruction and context supplied to a model. Ask people to state the task, the intended audience, the source material and the required output. For example, a summarisation task should specify which text to summarise and which exceptions must remain visible. Separate the source content from the instructions so it is easier to review both.
Complex wording is not inherently better. Start with a plain request, inspect the result and revise the instruction where the task is misunderstood. If information is missing, ask the tool to identify the gap rather than invent a plausible answer. Keep prompts reusable where appropriate, but explain that a saved prompt still requires review when the input or purpose changes.
4. Practise checking the result.
Ask learners to compare an answer with its source. Check names, dates, quantities, exclusions and conclusions rather than focusing only on grammar. Generated references can be incorrect, and a genuine link can still fail to support the attached claim. Open cited material and inspect the relevant passage before relying on it.
For calculations, use an appropriate calculation tool and review the inputs. For legal, medical or financial matters, involve a qualified professional where individual advice is needed. Training should make the limits of the task clear: an assistant can help draft or organise information without becoming the person responsible for a consequential decision.
5. Match practice to each role.
Choose exercises based on the work people actually perform. A communications team may need to preserve approved wording and check external claims. An operations team may need to validate extracted fields. A manager may need to distinguish an incomplete summary from an accurate account of a policy. These tasks require different checks even when they use the same model.
Include examples that cannot be completed from the supplied information. Practising a refusal, clarification or escalation is as important as producing a polished answer. Provide accessible written instructions alongside demonstrations so people can repeat the task without depending on a fast-moving presentation or a remembered sequence of clicks.
6. Leave usable guidance behind.
Record approved tasks, prohibited information, review requirements and the route for reporting a problem. Assign an owner to the guidance and explain how changes will reach staff. If the organisation changes tools or connects new data sources, revisit the examples rather than assuming previous instructions still apply.
A training engagement should produce task-specific practice and working guidance, not a claim that attendance establishes competence for every AI use. Check learning through an actual task and its review. Connect any recurring errors to the evaluation process, and revisit the project scope if people are being asked to rely on outputs beyond the approved purpose.