Playbook · for Leadership, Security
Shadow AI: a usage policy people follow
Give people approved AI tools and clear data rules, then back the policy with access controls and monitoring.
By Mr.DLast reviewed 2 min read
Shadow AI is already happening
If your company has no approved AI tools, people use their own. They paste customer emails into a chatbot to draft a reply, upload a contract to summarise it, or connect a browser extension to their inbox. Most do it to work faster, not to break rules. The risk is that you no longer know where your data is, and under GDPR or HIPAA that can be a reportable incident in itself.
Banning AI does not work; it moves use out of sight. The fix is a safe default plus clear rules.
1. Offer approved tools first
Pick an enterprise option for the common needs (writing, coding, summarising, meeting notes) with:
- single sign-on and the ability to remove access when someone leaves;
- no training on your data, and retention you can set;
- admin controls and audit logs;
- a data processing agreement and a known processing location.
Publish the list where people look for it, and say how to request a new tool.
2. Write a one-page policy
Keep it short enough that people read it. A template:
textAI usage policy (v1) 1. Use only tools on the approved list for company work. 2. Never put these into any AI tool: passwords and keys, health data, payment card data, customer data outside an approved tool, unreleased financials, source code outside approved coding tools. 3. Check AI output before you use it. You are responsible for it. 4. Say when content sent to customers was mostly AI-generated. 5. Report mistakes quickly (wrong data in a tool, a strange answer, a tool asking for access). Early reports are never punished. 6. New AI tools, agents or integrations need approval from <owner>. Owner: <name>. Review: every six months.
3. Back it with technical controls
- Use SSO for approved tools and block personal accounts where your identity provider allows it.
- Add the common AI domains to your web filter or secure web gateway: allow approved tools, warn or block others for teams that handle regulated data.
- Use data-loss prevention on endpoints or the browser to stop the most sensitive data classes.
- Review OAuth grants regularly; AI add-ons often ask for full mailbox or drive access.
4. Train with real examples
Ten minutes in onboarding and once a year: what may go into which tool, two real incidents, and how to report a mistake. People follow rules they understand.
5. Keep an inventory and an owner
Track approved tools, who uses them and what data they may hold. Name one owner for AI risk. This inventory is also the starting point for ISO 42001 and for EU AI Act questions about which AI systems you deploy.
How it maps to audits
The policy, approved list, access reviews and training records are evidence for ISO 27001 controls on acceptable use, supplier relationships and awareness, for SOC 2, and for the AI literacy duty in the EU AI Act. WardBee tracks them as policy and manual controls next to your automated checks.
Related standards: ISO/IEC 27001 · ISO/IEC 42001 · NIS2 / GDPR / HIPAA


