Practical guide
Start with a useful business problem
Identify a task that is repetitive, time-consuming or difficult to scale, then define what a good outcome would look like. Consider whether AI is appropriate or whether a simpler process change would solve the problem more reliably.
Choose a limited use case and a small group to evaluate it. Agree what the tool may do, what remains a human responsibility, how you will measure quality and when the trial will stop or be reviewed.
- Name the business owner and people who will use or be affected by the tool.
- Set a measurable purpose, boundaries and a review date.
- Test outputs against real requirements before relying on them.
- Consider accessibility, bias, accuracy and the effect on customers or staff.
Discover existing tools and shadow AI
Before choosing a new product, find out what staff already use. Ask about public AI websites, personal accounts, browser add-ons, meeting transcription tools, coding assistants and AI features embedded in software the business already pays for.
Treat this as discovery, not a blame exercise. People may be trying to solve genuine problems or save time. Understand the task, tool, account type and information involved, then provide an easy route to review the use case and offer a suitable approved option where one is available.
- Talk to teams and review software inventories, subscriptions and enabled features.
- Record the purpose, owner, user group and account type for each use.
- Check whether tools connect to mailboxes, cloud files, customer systems or other applications.
- Prioritise review based on the sensitivity of the data and the impact of the task.
Map where information flows
For each proposed service, trace what people enter or upload, where the service processes it, what may be retained and whether it is shared with another provider. Check the exact product and account: practices can differ between free and paid plans, business subscriptions, features and settings.
Look beyond the prompt itself. A connected assistant may retrieve files, emails or records; an integration may pass data to another service; and an automated agent may take actions. Review permissions, logs, storage locations, deletion options and supplier responsibilities. If a provider cannot answer an important question, record it as an unresolved risk before using sensitive information.
Information sent
Prompts, files, recordings, images, code, identifiers and context supplied by users or connected systems.
Recipients and processing
The AI provider, hosting services, integration partners and other parties that may receive or process the information.
Retention and reuse
How long inputs and outputs are kept, who can access them and whether they may be used to improve services or models.
Outputs and actions
Where generated content is stored, who validates it and whether the tool can update records or trigger actions.
Set clear rules for information and decisions
Write simple guidance that staff can apply. Explain which tools are approved, what information may be used, what must never be entered, how outputs should be checked and how to report a concern. Tailor the rules to your data and obligations rather than assuming one policy fits every tool.
Consider personal information, confidential client material, commercially sensitive plans, credentials, intellectual property and information subject to contractual or legal requirements. Involve appropriate privacy, legal, compliance or data-protection advisers when a use case needs that review. This guide is practical IT guidance, not legal advice.
- Require human review for important customer, financial, employment, legal or operational decisions.
- Do not enter passwords, access tokens or restricted data into an unapproved service.
- Explain how to check for inaccurate, out-of-date or misleading output.
- Define who approves new tools, integrations and higher-risk uses.
Prepare access, security and support
Use managed business accounts where suitable and confirm that the organisation can administer access, apply settings and remove accounts when people leave. Apply strong authentication, least-privilege permissions and appropriate controls to connected data sources.
Decide who monitors the service, handles incidents, reviews supplier changes and supports users. Include AI services in your supplier and software records, access reviews, security discussions and continuity planning. For existing systems and responsibilities, see the IT health check guide.
Train, test and improve
Give people practical examples that match their roles. Show them how to use approved tools, protect information, verify outputs and ask for help. Make clear that AI can make confident mistakes and that human accountability does not transfer to the tool.
During a trial, check quality, time saved, user experience, data handling and unexpected effects. Record decisions and limitations, then expand, change or stop the use based on evidence. Review the arrangement when the service, integrations, data or business purpose changes.
- Record the approved purpose, owner, tool, users and data categories.
- Test the normal case and likely failure cases before wider rollout.
- Provide a route to report inaccurate outputs, unintended disclosure or unexpected actions.
- Schedule reviews for access, supplier terms, outcomes and staff guidance.
