Shadow AI is the use of AI tools at work without the knowledge or approval of IT or leadership: staff pasting client data into a free chatbot, installing an AI browser extension, connecting an AI note-taker to every meeting, or turning on AI features inside an existing SaaS app. It isn't a sign of bad employees. It's a sign that people find AI useful and the business hasn't given them a safe way to use it. The answer isn't a blanket ban, which drives usage further out of sight. It's to discover what's in use, approve good tools, restrict risky ones and give people a sanctioned alternative.
Why shadow AI is a real risk
- Data leaves your control. Free and consumer tiers of AI tools may retain prompts and, depending on their terms and settings, use them to improve models. Client files, source code, financials and personal data pasted into them are outside your contracts and retention policies.
- Over-broad permissions. AI note-takers, email assistants and "connect your Google Drive" plugins often request read access to entire mailboxes, calendars or file stores, through OAuth consents a user can grant in seconds.
- Compliance exposure. Sharing protected health information, customer financial data or controlled data with an unvetted vendor can breach HIPAA, the FTC Safeguards Rule, client contracts or CMMC requirements.
- Unreliable output. AI-generated content used in client deliverables, contracts or code without review can introduce errors and liability.
- A bigger attack surface. Every extension and integration is another vendor that can be compromised. See SaaS sprawl.
For a wider view, read the AI security risks SMBs are already living with and from PII to BII.
Where shadow AI hides
| Type | Examples | How to find it |
|---|---|---|
| Standalone chat assistants | Personal accounts on public AI chat services | Web proxy, DNS or secure web gateway logs; firewall application reports |
| Browser extensions | Writing assistants, summarizers, "AI for Gmail" add-ons | Browser management in Intune, Google admin or your endpoint tool |
| OAuth-connected apps | Meeting note-takers, email and calendar assistants, file analysis tools | Enterprise applications and consent grants in Entra ID; third-party app access in Google Workspace |
| AI features in existing SaaS | AI add-ons in CRM, design, HR or project tools | SaaS admin consoles, renewal quotes, vendor change notices |
| Desktop apps and code tools | Local AI apps, coding assistants in developer tools | Software inventory from your RMM or endpoint platform |
| Expense claims | Personal AI subscriptions reimbursed as "software" | Expense and card reports |
A five-step process to get control
1. Discover
Pull the reports above for the last 90 days and list every AI tool in use, how many people use it and what it connects to. Then ask people. A short, no-blame survey ("which AI tools help you, and for what?") surfaces usage logs miss and tells you what needs a sanctioned replacement.
2. Assess
For each tool, answer a few questions:
- Does the vendor use our inputs to train models, and can that be turned off contractually, not just in a setting?
- Where is data stored, for how long, and can we delete it?
- Does it support SSO and MFA, and admin controls over sharing?
- What permissions does it request, and are they the minimum needed?
- Does the vendor have a SOC 2 report or ISO 27001 certification? See SOC 2 vs ISO 27001.
- What data would staff realistically put into it?
3. Decide: approve, restrict or block
- Approve tools with business-grade terms, SSO and acceptable data handling, and buy licenses so people use the business tier.
- Restrict tools that are useful but risky: allow them only for non-sensitive data, or only for specific teams.
- Block tools that fail the assessment, and tell people what to use instead.
4. Enforce with technical controls
- Require admin approval for new OAuth app consents in Microsoft 365 or Google Workspace, so users can't grant mailbox access to unknown apps.
- Manage browser extensions with an allow list on company devices.
- Use your secure web gateway or firewall to block unapproved AI services, with a message pointing to the approved one.
- Apply sensitivity labels and data loss prevention to stop labeled data being pasted or uploaded to unapproved destinations where your licensing supports it.
- Review the access an approved AI assistant inherits. Tools such as Microsoft 365 Copilot see what the user can see, so fix oversharing first. See Copilot readiness.
5. Write it down and keep it current
Publish an acceptable use policy that lists approved tools, what data may never be entered into any AI tool, and how to request a new one. Our AI acceptable use policy template and guide to what to include are a fast start. Re-run discovery quarterly, because new AI features ship inside existing tools every month.
Give people a better option
Shadow AI shrinks fastest when the approved tool is easier and better than the workaround. Pick one or two business-grade assistants, license them properly, connect them to SSO, and run short training on safe, productive use. That turns AI from a hidden risk into a managed advantage. Our AI governance guide and AI enablement services cover the wider program.
Frequently Asked Questions
What is shadow AI?
Shadow AI is the use of AI tools, extensions or AI features inside other software without approval or oversight from IT or leadership. It is a subset of shadow IT focused on AI services.
Should we ban AI tools to stop shadow AI?
Usually not. Bans tend to push usage onto personal devices and accounts where you have no visibility. Approving safe tools, restricting risky ones and giving staff a sanctioned alternative is more effective.
How do we find out which AI tools employees are using?
Combine web proxy, DNS or firewall logs, OAuth consent reports in Microsoft 365 or Google Workspace, browser extension inventories, software inventories and expense reports with a short, no-blame staff survey.
Do free AI chat tools use our data for training?
It depends on the provider, plan and settings. Consumer and free tiers are more likely to retain prompts or use them to improve models. Business tiers typically offer contractual commitments not to train on your data. Check the current terms for each tool.
What data should never go into an AI tool?
Unless a tool has been specifically approved for it, avoid passwords and secrets, personal data, health or financial records, client confidential information, controlled or export-restricted data, and non-public financial results.
How often should we review AI tool usage?
At least quarterly, and whenever a major SaaS vendor adds AI features. AI capabilities change quickly, so discovery should be a routine, not a one-off project.
Not sure which AI tools your team is already using? Defensible runs a shadow AI discovery, assesses each tool and sets up the controls and policy to use AI safely. Talk to our team.



