Choose a small set of approved working tasks
Begin with recurring tasks that have a clear owner and an observable output. Examples include preparing an internal summary from approved material, drafting a response for review or organising information for a project handoff. Distinguish these from consequential decisions that require specialist judgement or authority.
Ask staff where the work slows down and what they currently do to check it. Select tasks where the team can compare an assisted output with its normal standard. Avoid launching a broad instruction to “use AI more” without giving people a practical purpose and a way to recognise an unsuitable use.
Assess the starting environment before the workshop
Inventory the tools already approved, the accounts people actually use and the information they may put into them. Review access, contractual arrangements and internal policies with the relevant owners. A useful training session should teach the approved route, not encourage people to improvise one around an unresolved restriction.
Check the operational prerequisites too: document quality, naming conventions, access to source material and ownership of templates. If staff cannot locate an approved answer, training them to write better prompts will not solve that underlying problem. Separate the information-management work from the skills programme and coordinate both.
Teach verification as part of doing the task
Use representative examples from the team’s work, with appropriate permission and redaction. Demonstrate how to provide a clear purpose, relevant source context and an expected format. Then spend equal attention on checking the output against its source and the organisation’s own requirements.
Show an incorrect, incomplete and misleading result as well as a strong one. Staff should know when to correct an answer, seek an authorised source, request help or abandon the tool. The ability to recognise a bad output is a working skill, rather than a brief warning at the end of a presentation.
Give each role the authority it actually needs
Separate drafting from sending, recommending from approving and retrieving from disclosing. A person authorised to prepare an internal summary may not be authorised to share the source externally. The training materials should make those boundaries clear using realistic situations relevant to the department.
Where an AI workflow can act in another system, teach the approval and exception route as well as the user interface. Explain what information an operator should inspect before accepting a proposed action. Do not describe automation as a reason to stop owning the business outcome.
Pilot with managers and measure correction work
Run a limited pilot with a named business owner, representative users and a defined review period. Establish a baseline for the task before measuring improvement. Include the time spent collecting inputs, checking outputs and resolving mistakes, rather than recording generation time alone.
Look at useful completion, correction effort, missed information and user confidence in handling exceptions. Give staff a simple way to report a failure without hiding it to protect an adoption target. Feedback should improve the task design, examples and tool configuration, not merely be counted as resistance to change.
Build an operating programme beyond the first session
Provide short task playbooks, approved examples, a question route and an owner for updating guidance. Appoint internal champions with a defined responsibility and time allocation. Managers need to know what good use looks like and which questions should be escalated to IT, privacy, legal or a process owner.
Review the programme when tasks, tools or permissions change. Retire examples that no longer match the approved environment. Expand to additional teams when there is a reusable working pattern and the necessary support, rather than assuming one general workshop gives every department the same capability.