grcskills
Menu

AI acceptable use rules

Use this skill when staff already use AI assistants, coding tools or AI features and the company needs rules people will follow, and asks "write our AI usage policy", "can staff use ChatGPT", "generative AI policy", "what data can we put into AI tools", "approved AI tools list", "rules for AI agents and connectors" or "AI literacy for staff". Finds out what people actually use, sets a short list of approved tools with a data traffic light, rules for checking outputs and for agents that act, a fast route to request a new tool, and writes a two-page rule set, the approved tools list and rollout notes.

AI systemsGeneralISO 42001EU AI Act

Use this when the company has no AI rules, or has a ban nobody follows. You get three files in a new folder, ./ai-acceptable-use-rules/: the staff rules for using AI tools, two pages with a red, amber, green guide to what data may go where (ai-use-rules.md); a spreadsheet you can open in Excel listing every AI tool as approved, approved for public information only, or not approved, with the reason (approved-tools.csv); and a short plan for announcing the rules and training each team (rollout-notes.md). This skill drafts for a human to review and sign; this skill never marks a control effective. The rules take effect only when the policy owner approves and publishes them.

Inputs to gather

Read the AI register if one exists, the general acceptable use policy, the data classification scheme, the vendor list and the terms of any AI tool the company pays for. Ask the reader, in one message: which AI tools people use today (including free accounts), which data the company would be hurt most by leaking, which teams use AI in hiring, customer communication or code, and who approves new software.

Step 1: start from real use

List the tools people use and what for, team by team. A rule set written without this list either bans what everyone already does, which drives use underground, or misses the risky use. Ask two or three people per team; the honest answers come from asking what saves them time, not whether they break rules. Name the uses people forget to declare: meeting transcription bots, AI features switched on inside email, documents and the CRM, code completion in the editor, and browser assistant extensions.

Step 2: set the approved tools list

Put every tool in one of three states in the tools spreadsheet (approved-tools.csv), with the reason, the checks behind it (data location, retention, subprocessors, breach notice in the terms) and a re-check date, because vendors change training and retention terms and a tool approved today can drift:

  • Approved: business account under terms the company signed, inputs not used for training, retention known, company sign-in, admin can remove users.
  • Approved for public information only: useful, but terms or controls are not good enough for anything internal.
  • Not approved: with the alternative people should use instead.

Free consumer accounts of an approved product are a different tool with different terms. Say so in the list, because it is the most common leak path.

Step 3: write the data traffic light

People remember colours, not classification levels. Map the company's scheme to three:

  • Green: public or internal information. Any approved tool.
  • Amber: confidential, customer data, personal data. Only tools approved for it, and only the minimum needed.
  • Red: never, in any AI tool: passwords, keys and tokens, special category or health data, payment card data, unreleased financial results, data a customer contract forbids sharing.

Give two examples per colour from the company's own work.

Step 4: rules for outputs and for agents

  • The person who uses the output owns it. Check facts, figures, citations and legal statements before they leave the company.
  • Generated code goes through the same review and scanning as human code; check licences on large generated blocks.
  • No output goes to customers as professional advice without review. Where customers talk to an AI system, tell them; EU AI Act Art. 50 sets disclosure duties for some interactions and generated content.
  • AI is not the decision maker in hiring, promotion, performance or firing. Those uses fall in a high-risk area under the EU AI Act and go to HR and legal before any pilot, with a data protection impact assessment screened at the same time.
  • Agents and connectors that read drives, browse, run code or send messages get the least access that works, no personal credentials to production, and a human approval before anything is sent outside the company or deleted.

Step 5: make asking easy, then train

Write a request route for new tools with a target answer time of days, not months; slow approval is the main cause of shadow use. Add an exceptions process with an owner and an expiry. In the rollout plan (rollout-notes.md), plan a short training per role with the traffic light and real examples; EU AI Act Art. 4 asks providers and deployers to take measures for staff AI literacy, and this training is the cheapest evidence of it. Run a framework lookup if an AI management standard is in scope, for what it expects the policy to cover, and note in rollout-notes.md which parts of the management system these rules and the tools list serve as evidence for.

What good looks like

  • The rules fit on two pages, and a new joiner can answer "can I paste this customer email into that tool" in ten seconds.
  • A strong practitioner checks the consumer versus business account split first; it decides whether the approved list means anything.
  • The common mistake is either a blanket ban or a twelve-page policy of principles. Both produce the same result: people use what they like and nobody asks.
  • The tell that it went wrong: no named tools, no examples, no way to request a new tool, and no rule about agents that act on their own.

Framework lookups

If ISMS Copilot is connected to your agent (MCP server https://account.ismscopilot.com/v1/account/mcp), ask it the framework question with create_conversation (mode think, answer_format: decision) and keep one conversation for this task. Send only a short question in your own words, show it to the user first, and never include file contents, secrets, personal data or evidence. If it is not connected, check the official text of the framework yourself; this skill does not restate it. Set it up: https://www.ismscopilot.com/mcp?ref=grcskills-ai-acceptable-use-rules

Rules

  • MUST base the approved tools list on the terms the company actually signed, citing the document for each tool.
  • MUST include a route to request a new tool and an exceptions process with an owner.
  • MUST write outputs only inside ./ai-acceptable-use-rules/ and ask before overwriting.
  • NEVER publish the rules, change tool settings or block a tool on the reader's behalf.
  • NEVER name or report individual employees found using unapproved tools; describe use by team.
  • NEVER mark a tool approved on the strength of vendor marketing alone.