Employee rules for generative AI: what company data can you use?

1. Publish an approved-tool register before writing prompt rules

List the exact service, account type, owner and allowed purpose. “ChatGPT is allowed” is not precise enough: a personal account, a company-managed business workspace and an API integration can have different terms, retention, access and administration. Record who may add users, whether sharing links are allowed and where employees report a missing feature or a new tool.

Check current first-party documentation and the signed agreement. For example, OpenAI says that its named business products and API do not use organisation data for model training by default and describes encryption and retention controls. That is a product-specific claim, not permission to assume that every plan, account or other provider behaves the same way. Recheck the register when a plan, setting, integration or supplier term changes.

• Exact product, plan and company workspace.
• Approved tasks and prohibited tasks.
• Workspace owner, user access and sharing rules.
• Date and person responsible for the next review.

2. Classify inputs into three lanes employees can recognise

Green means public, intentionally fictional or safely redacted material used for an approved low-risk task. Examples include rewriting a published product description, creating a meeting agenda from generic bullet points or testing a prompt with invented customer names and figures. Redaction must remove details that can be combined to identify a person, client, negotiation or internal decision.

Amber means non-public material that may be useful but requires the approved business workspace and a named process owner: an internal procedure, an unpublished draft or a limited dataset prepared for a specific workflow. Red means stop and ask: passwords and API keys, special-category or sensitive personal data, legal advice, employment cases, confidential customer documents, source code or commercial terms that the company is not authorised to disclose. Adapt the examples to your contracts and information-classification rules.

• Green: public, fictional or genuinely de-identified input.
• Amber: approved internal input in an approved workflow only.
• Red: secrets, sensitive personal data or restricted third-party material.
• Unclear: stop and ask the information owner before uploading.

3. Define the task, not only the data

The same text can be acceptable for one purpose and inappropriate for another. Drafting three subject-line options from a public announcement is different from deciding which customer receives credit or which candidate advances. For each approved use case, state the allowed input, expected output, required reviewer and action the tool must never take.

Start with reversible work: outlines, alternative wording, categorising a small redacted sample or extracting fields for a person to check. Do not let an experiment silently become an automated decision or external message. If the output can affect money, employment, legal rights, safety, customer commitments or access to a service, involve the appropriate business, legal, security or data-protection owner before deployment.

• Purpose and the employee role using the tool.
• Allowed input and source of that information.
• Person who checks and accepts the output.
• Actions and decisions that remain outside the tool.

4. Make human review observable, not ceremonial

“A human checks it” is not a control unless the reviewer knows what to verify and has enough time and source material. Give each workflow a review list: factual accuracy, missing context, unsupported claims, confidential content, biased wording, calculations, links and consistency with company policy. The person who sends or saves the result remains responsible for the final work.

The NCSC notes that large language models can produce convincing but incorrect answers. Review intensity should therefore follow consequence, not writing quality. A polished internal brainstorm may need a quick sense check; a customer quote, financial instruction, HR document or technical change needs independent evidence and, where relevant, specialist approval. Keep a few failed examples in training so staff learn what plausible errors look like.

5. Give employees a safe response when something is pasted by mistake

A useful policy explains what to do after an error without encouraging concealment. Tell the employee to stop the conversation, avoid pasting more context, preserve the minimum details needed to investigate, and report the event through the existing security or privacy route. The responsible owner can then check account controls, shared links, retention options, contractual duties and whether affected people or clients require further action.

Do not promise that deleting a chat reverses every form of processing, and do not improvise a legal conclusion. Record what was entered, which account and workspace were used, when it happened, who could access it and what immediate action was taken. Use the incident to improve examples, access settings or training rather than merely adding another warning sentence.

• Stop and report through the named internal route.
• Record the tool, workspace, time and type of information.
• Check sharing, retention and access with the responsible owner.
• Update the control or training after the review.

6. Train by role and review the rules with real work

A twenty-minute briefing with examples from sales, finance, operations and HR is more useful than one generic prompt lesson. Ask employees to classify sample inputs, identify the reviewer and practise a stop-and-ask case. The European Commission says AI-literacy measures should reflect staff knowledge, the systems used, their risks and the context; it does not prescribe one identical course for every organisation.

Review the policy after the first month, then whenever tools, workflows or supplier terms change. Look at questions, rejected outputs, near misses and useful approved cases. The Commission's repository shows varied literacy practices, but also warns that copying an example does not automatically demonstrate compliance. Treat this guide as an operational starting point, not legal, security or data-protection advice for your specific organisation.