Rules for the Robot in the Room: An AI Use Policy That Survives Scrutiny
Photo: Foreign, Commonwealth & Development Office, CC BY 2.0, via Wikimedia Commons
Generative AI is no longer a novelty confined to innovation teams. It sits in email clients, word processors, customer-service tools and the browser tabs of employees looking for a faster draft. For UK directors, the risk is not that staff will use AI; it is that they will use it without a shared understanding of what may be entered, what must be checked and who remains accountable for the output.
A policy is not a ban. A well-drafted acceptable use policy gives employees a clear route to productivity while setting boundaries that protect client confidentiality, personal data and the company's intellectual property. The objective is governed adoption, not performative caution.
Start with the data, not the tool
AI policy discussions often begin with a list of approved products. That is the wrong starting point. The first question is what data the company holds, where it sits and which categories are sensitive. Client contracts, employee records, pricing models, unreleased product plans and personal data all carry different obligations.
Under the UK GDPR and the Data Protection Act 2018, personal data entered into a third-party AI tool may constitute a processing operation. The company remains the data controller for that processing unless the facts clearly indicate otherwise. Directors should establish whether a data protection impact assessment is required, what the vendor's retention and training practices are, and where processing takes place. A tool that stores prompts or uses inputs to train models may be unsuitable for confidential material, regardless of how convenient its interface appears.
The four pillars of a workable policy
A credible policy rests on four commitments.
First, confidentiality. Employees should know which information may never be entered into an external AI service. This includes client data, unpublished financial information, personal data and material covered by non-disclosure agreements. Where an approved enterprise tool has been configured with appropriate controls, the policy can permit defined use cases.
Second, data protection. The policy should link to the company's existing data protection framework rather than sit apart from it. It should identify the lawful basis for processing where relevant, set retention expectations, and require employees to report suspected data incidents promptly. A prompt that contains personal data can be a personal data breach if it is disclosed to an unauthorised recipient.
Third, intellectual property. UK copyright and database rights do not map neatly onto AI-generated material. Employees should understand that AI output may not be original, may reproduce third-party content and may not attract the same protection as human-authored work. The policy should require review before output is used in client deliverables, marketing copy, product specifications or anything that may be registered.
Fourth, accountability. AI may draft, summarise or suggest, but a named human being remains responsible for the result. The policy should make clear that employees must verify factual claims, check calculations and apply professional judgement. A disclaimer that 'the AI wrote it' is not a defence.
A policy outline employees can follow
A practical policy need not be long. It should be specific enough to guide daily decisions and short enough to be read. A workable structure includes:
- Purpose and scope, including which staff, contractors and devices are covered.
- Approved tools and the process for requesting a new one.
- Prohibited inputs, with clear examples rather than abstract warnings.
- Verification and review requirements before output is used.
- Disclosure rules where AI has contributed to client-facing work.
- Ownership and licensing of inputs and outputs.
- Incident reporting and escalation routes.
- Training, monitoring and review arrangements.
The language should be direct. Employees need to know what to do when a client sends a document and asks for a summary, or when a manager suggests using a free AI tool to speed up a tender response. Examples are more effective than principles alone.
A director-level checklist
Board oversight should focus on governance rather than technical detail. Directors should be able to confirm the following:
- A named owner is responsible for AI policy and its review.
- A register of AI tools in use has been compiled, including tools embedded in existing software.
- Material vendors have been subject to due diligence on security, retention and sub-processors.
- Contract terms address data use, confidentiality, liability and termination.
- A data protection impact assessment has been completed where required.
- Employees have received training that reflects their actual use cases.
- There is a simple route for reporting mistakes and suspected incidents.
- The policy is reviewed at least annually and after significant changes in law, guidance or tooling.
Making it stick
The greatest risk is shadow AI: employees using unapproved tools because the approved route is slow or unclear. Directors should therefore pair the policy with a workable approval process and a small set of supported tools. A policy that says no to everything will be ignored; a policy that says yes without boundaries will not protect the company.
AI adoption is now a governance question, not a technology experiment. The companies that manage it well will be those that treat the policy as a living control, test it against real workflows and keep the accountability for output firmly in human hands.