Managing AI tools in the workplace — use policy, training and human review

Managing AI tools in the workplace means publishing a clear policy, assigning roles in tools such as Microsoft 365 Copilot and enforcing at least one human review step for outbound material; combine role-based access, staff training and a single approval gate to reduce reputational and compliance risk.

Policy and access: set rules before you deploy

Start by writing a short, focused policy that states what is acceptable and what is forbidden when staff use AI tools. Make the document a living one-page policy with three clear sections: permitted use cases (drafting internal notes, summarising documents), prohibited actions (sharing customer PII, generating formal legal advice) and required controls (authentication, logging, approvals). Keep the language practical: name the tools you permit (for example, Microsoft 365 Copilot or an approved enterprise chatbot), identify who can approve new tools, and list the data classes that must never be pasted into an AI prompt.

Practical controls to include:

Role-based access: only senior or trained staff should access tools that can read or write client data.
Data classification rules: mark personal data and commercial secrets and block them from prompts.
Logging and rotation: log usage for 90 days and review logs monthly to spot unusual behaviour.

Use short, named checklists rather than long prose so staff can follow the rules quickly. Link the policy to your procurement process so any new AI subscription gets a security and privacy sign-off before purchase — that sign-off can be part of your wider managed IT arrangements by outsourcing oversight to a specialist if you prefer. For example, include a link to your managed IT and AIOps service so decision makers know where to go for help: managed IT and AIOps service. If you want independent guidance on secure defaults and sensible technical controls, consult NCSC’s guidance on secure use.

Operational controls: training, review gates and monitoring

Policy alone won’t stop mistakes. Train staff on the policy, run tabletop exercises using the actual tools your teams will use, and set clear escalation paths for when AI output looks wrong. Use short, scenario-based training: give a three-line brief and ask staff to spot what data must not be included in prompts, who should approve the resulting text, and where the final copy is stored. Combine this learning with simple daily routines — a five-minute checklist before sending anything externally that was drafted or edited by AI.

AI-drafted client-facing communications carry a live business risk without a review step — we have seen an AI-drafted invoice email confidently name-drop the wrong client. Insert a human review gate anywhere AI is writing outbound content on behalf of the business. In practice this means at least one named approver who checks tone, facts and client identifiers before anything is sent. Make that approver accountable: record who approved the message and store the final version in your document management system.

Operational controls to operationalise:

Human review gate: mandatory sign-off by a named person for all client-facing AI output.
Templates and prompts: standardise prompts and response templates so outputs are predictable.
Monitoring and KPIs: track monthly exceptions and near-misses — treat them as learning, not blame.

An initial audit of your use can usually be done in a week and will reveal where role changes, training or simple template fixes will remove most of the risk. Keep the controls lightweight: a one-click approval for routine messages and a longer checklist for invoices, contracts or regulatory submissions is often enough. Finish by setting a review cadence — revisit your policy and logs every quarter, or sooner if you change the tools you use.

FAQ

How should a small UK firm control staff use of ChatGPT or similar tools?

Publish a short acceptable-use policy, limit access to named users for anything involving client data, and require a human approver for any output that will be shared externally; this keeps risk low without blocking useful internal summaries.

Do I need a written AI policy to use generative tools with clients?

Yes — even a single-page written policy prevents common mistakes and gives staff a default action; aim to write it in one afternoon and review quarterly.

How many people should approve AI-written client correspondence?

Require at least one human reviewer per item of client-facing content and record who approved it; for high-risk documents (invoices, contracts) add a second reviewer or manager sign-off.

How often should I check AI tool logs and access rights?

Review access rights at least monthly and scan usage logs every month; escalate anomalies immediately and run a full policy review every three months or after any serious near-miss.