The short answer
Shadow AI is staff using AI tools the organization has not approved: a personal chatbot account used to draft a client letter, a browser extension that summarizes web pages, a free transcription app in a meeting, or a new AI feature in software you already pay for. It happens because the tools are useful and the approved route is slow or missing.
The risks are real: confidential and personal information entered into services with unclear terms, inaccurate output used in decisions, and privacy obligations nobody checked. The fix is not a ban. Find out what is in use, approve tools that meet the need, set a short policy, train people and back it with data protection controls.
Why it happens
- The tools genuinely help. Drafting, summarizing and rewriting are faster with AI, and busy people use what works.
- There is no approved option, or it is harder to reach than a free website.
- Nobody said what is allowed. Without a policy, people make their own judgment calls, and they vary.
- AI arrives inside other software. Vendors keep adding AI features to tools you already use, so new AI can appear without anyone installing anything.
Most shadow AI is not malicious. Treat it as a signal of demand, not as misconduct.
The real risks
| Risk | What it looks like | Why it matters |
|---|---|---|
| Confidential data in consumer tools | A contract, client list or staff issue pasted into a personal chatbot account | The information has left your control, under terms you never agreed |
| Unclear retention and training terms | Prompts kept for an unknown period, possibly used to improve the vendor's models | You cannot tell clients or regulators where their information went |
| Inaccurate output used in decisions | A summary that drops a key condition, or a confident but wrong figure | Errors reach clients and boards with your name on them |
| Privacy obligations missed | Personal information processed by a service with no contract or assessment | The FTC Act, state privacy laws and, for health data, HIPAA still apply |
| Unvetted access | Extensions and apps granted broad access to mailboxes, files or browsers | A new route for data loss or attack |
NIST's Generative AI Profile (NIST AI 600-1), a companion to its AI Risk Management Framework, lists data privacy (leakage or unauthorized disclosure of personal or sensitive information), information security and confabulation (confidently stated but false content) among the risks of generative AI, and suggests actions organizations can take to manage them (NIST). Retention and model-training terms vary by service and plan, and they change, so check them rather than assume.
On privacy, the accountability stays with you. The FTC can treat a failure to protect personal information, or to keep your own privacy promises, as an unfair or deceptive practice, and its business guidance expects you to make sure the service providers you share data with protect it too (FTC). Where they apply, state privacy laws such as the CCPA (as amended by the CPRA) and the Texas Data Privacy and Security Act also call for a contract with each service provider that processes personal data for you. A personal chatbot account has no such contract with your organization. Health information adds HIPAA: an AI vendor that handles protected health information for a covered entity is a business associate and needs a business associate agreement (HHS), which a personal account never provides; our guide to AI and HIPAA explains more. This is general information, not legal advice.
How to discover shadow AI
Start with evidence, not accusations.
- Your security tools. Firewall and secure web gateway logs already show which AI sites people reach. Microsoft Defender for Cloud Apps uses traffic logs from firewalls and proxies, or device data through Microsoft Defender for Endpoint, to give visibility of unsanctioned cloud apps (Microsoft Learn). Its app list includes a Generative AI category, so you can see which AI apps are used, review a risk assessment for each and sanction or block them (Microsoft Learn). Where your licenses include it, Microsoft Purview can add insight into sensitive data in AI interactions.
- App and consent inventories. Device management shows installed apps and browser extensions, and your identity platform shows which third-party apps users have connected to their work accounts.
- Vendor feature reviews. Check the admin settings of your main business platforms for AI features, what they do with your data and whether they are switched on.
- A staff survey without blame. Ask which tools people use, for which tasks and why, and say plainly that the goal is better tools, not discipline. You will learn what logs cannot show, such as AI on personal phones, and which tasks people most want help with.
Network tools only see managed devices and networks, which is why the survey matters.
Rather talk it through? If you suspect staff are pasting work into free AI tools but cannot see how much, we can help you find out and set up approved options quickly. Talk to a Promatics specialist
How to get it under control without banning AI
Bans tend to push AI onto personal devices, where you cannot see it at all. A better sequence:
- Approve tools that meet the need. Choose business versions of the tools people want, with data use and retention terms you have actually read. For Microsoft 365 organizations, Microsoft 365 Copilot is often the natural candidate; our comparison of Copilot, ChatGPT and Gemini covers the alternatives.
- Write a short policy. Approved tools, what must never be entered, when output needs human review, when to disclose AI use and how to request a new tool. Our guide to an AI acceptable use policy walks through it.
- Train people on real tasks. Show how to use the approved tools well, what not to paste and how to check output.
- Add data protection controls. Sensitivity labels and data loss prevention can stop sensitive content reaching unapproved sites where your licenses support them. Block high-risk apps; warn and educate for the rest.
- Keep a register. List approved tools and use cases with owners, and keep the request process fast, so the approved route stays easier than the workaround.
- We know which AI tools and AI features are in use, from logs and a staff survey.
- At least one approved AI tool is available to staff who need it.
- Vendor terms on retention and model training are reviewed for each approved tool.
- A short AI use policy is published and acknowledged.
- Staff are trained on what may and may not be entered.
- High-risk AI apps are blocked, and data protection controls cover sensitive content.
- Requests for new tools have an owner and a quick turnaround.
When to bring in help
A small organization on Microsoft 365 can make a good start alone: run the survey, publish a one-page policy and switch on an approved tool.
Bring in help when you handle sensitive personal information (health, financial or children's data), when a board, funder, auditor or public agency client wants evidence, when discovery and data loss prevention need configuring in your security tools, or when AI features in vendor software need a risk assessment. Our AI governance and security service covers AI use policies, shadow AI discovery, risk assessments and security testing of AI applications, with alignment to frameworks such as the NIST AI RMF and ISO/IEC 42001. Because we also provide cybersecurity and Microsoft 365 services, we configure the discovery and protection controls ourselves, not just recommend them.
Sources and further reading
Product capabilities and guidance change. These are the primary sources this article relies on, checked on the review date above.
- AI Risk Management Framework and Generative AI Profile (NIST AI 600-1), National Institute of Standards and Technology
- How do I discover AI apps and the sensitive data these use in my organization?, Microsoft Learn
- Cloud app discovery overview (Microsoft Defender for Cloud Apps), Microsoft Learn
- Privacy and security guidance for business, Federal Trade Commission
- Health information privacy (HIPAA), U.S. Department of Health and Human Services
This article is general information, not legal, accounting or security advice for your specific situation. Examples are hypothetical unless stated otherwise.