Analysis Cybersecurity AI Compliance

Finding and controlling shadow AI: where it hides and how to respond

Where shadow AI hides beyond ChatGPT, why the type of account changes the risk, and the seven components of a shadow AI response your employees will follow.

10 November 2026
Updated 18 November 2026
6 min read

This page covers finding shadow AI and bringing it under control. For what shadow AI is and the three risks that matter most, see our shadow AI guide. The UK GDPR exposure and what to do after a leak are in shadow AI and UK GDPR.

Where shadow AI hides beyond the obvious

When people think about shadow AI, they usually picture ChatGPT. The real surface is wider and harder to map.

  • Browser extensions: Writing assistants, transcription add-ons and similar extensions can read what the user types or views in the browser, depending on the permissions granted. Many now include generative AI features, and an extension with broad permissions can see text entered into internal systems.
  • AI features inside SaaS tools already in use: Notion AI, Slack AI, HubSpot's AI features, Canva's Magic Write and many others add AI to platforms your staff already use. People often assume that approval of the platform covers the AI feature. Check whether the feature is switched on, which sub-processors it uses, where data is processed and whether your existing contract covers it.
  • Consumer access to enterprise-branded tools: Microsoft 365 Copilot Chat provides enterprise data protection when staff sign in with their Entra ID work account. Signed in with a personal Microsoft account, the same person gets consumer Copilot under consumer terms, even if your company pays for Microsoft 365 Copilot.
  • Personal devices: who owns a device and whether it is managed are different questions. An unmanaged personal phone or laptop sits outside most corporate DLP and web filtering, while a personal device enrolled in your device management (BYOD) may give you some visibility through managed apps and work profiles. Remote and hybrid work makes unmanaged use more common.
  • Voice and meeting tools: AI meeting transcription services such as Fireflies, Otter.ai and similar tools record and process meeting audio, sometimes including commercially sensitive discussions. Check who can switch them on and where recordings are stored.
Important

AI features embedded in approved SaaS tools need their own review. Notion AI, Slack AI or HubSpot's assistant may use different sub-processors, locations or retention periods from the core product. Approval of the parent tool covers the AI feature only once you have reviewed the feature and documented the decision.

The type of account changes the risk

RouteTypical positionWhat to check
Personal consumer accountNo contract with your organisation; the provider's consumer terms apply; conversations may be used to improve models unless the user opts outTreat as unapproved for personal or confidential data
Business or enterprise account under your work identityContract with your organisation; training on your data generally off by default; admin controls and logs, although contract terms, logging, retention and admin controls differ by provider and planData processing terms, retention, location, sub-processors, admin settings
AI feature inside an approved SaaS toolCovered by your existing contract only if the vendor's terms extend to the featureWhether the feature is on, sub-processors, any separate AI terms
API access built into your own systemsGoverned by the API agreement; major providers generally do not train on API data by defaultRetention, logging, transfer safeguards, who maintains the integration
Locally hosted modelData stays in your environment if the deployment is set up that wayModel source and licence, access control, patching, logging

Building a shadow AI response your employees will actually follow

The instinctive response to shadow AI is to block it. Blocking without offering an alternative tends to move the behaviour onto personal devices, beyond the reach of your controls. Staff use these tools because they help with the job, and they will keep doing so if there is no approved option. That does not mean never blocking: balance approved alternatives, monitoring, restrictions and enforcement, and block high-risk uses straight away, such as uploading special category data or source code to an unvetted service.

A workable shadow AI programme has seven components, roughly in order of priority. Give it a named owner and a review date, because the tools change faster than most policies.

  1. AI acceptable use policy: A short, clear document telling staff which tools are approved, what data they can use with each, how to request a new tool or an exception, and what to do when an AI feature appears inside a tool they already use. Keep it to a page or two with a clear approved-tools list.
  2. Approved AI tools list with data-tier guidance: Not all data is equally sensitive. Give staff a simple framework: public information can go into tool X; internal business information can go into tool Y with a work account; personal data, client confidential information and source code cannot go into any AI tool without specific approval. Clear rules remove the guesswork that currently drives poor decisions.
  3. Data classification: Staff cannot make good decisions about what to enter into AI tools if they do not know what counts as sensitive. A lightweight classification scheme with clear examples is a prerequisite for the acceptable use policy to work.
  4. Enterprise-grade alternatives: Give staff access to tools that meet your governance requirements, such as Microsoft 365 Copilot or Copilot Chat used with work accounts, an enterprise Claude or ChatGPT deployment, or other products that process data under a written contract with suitable retention and location terms. That removes much of the reason to use personal accounts.
  5. DLP and monitoring for AI services: Data loss prevention and web filtering tools can detect traffic to known AI services and, where configured, warn or block when sensitive data is involved. Monitoring staff is itself processing of their personal data, so keep it proportionate, tell staff what you monitor and why, and consider a DPIA. The ICO's guidance on monitoring workers sets out what that involves. None of these sources will reveal every agent or AI-enabled SaaS feature: consent logs and web filtering miss AI features inside approved tools, local agents and traffic from unmanaged devices, so combine them with the SaaS audit and staff reporting.
  6. Audit existing SaaS tools for embedded AI features: Review every tool in your approved software estate for AI features that have been added or are on the way. Identify which carry data protection implications and update contracts or acceptable use terms. Vendors add AI features on a rolling basis, so this is an ongoing task.
  7. Regular staff awareness, not a once-a-year tick-box module: Short, relevant communications tied to specific use cases reach staff better than an annual course. Link updates to changes in your approved tools, and make it easy for people to report a mistake without fear of blame.

Frequently asked questions

What should an AI acceptable use policy cover?

An AI acceptable use policy should cover: which tools are approved and for which use cases; what data staff cannot enter into any AI tool (personal data, client confidential information, source code, financial data, M&A information); how to review AI-generated outputs before using them in customer-facing work; how to request a new tool or an exception; and what to do when an AI feature appears inside a tool they already use, or when data has gone somewhere it should not. Keep it short enough to read and specific enough to act on.

Ryland Deakin
About the author
Lead Consultant, Cyvra · CISM · CompTIA Security+ · MCP

Ryland has delivered cybersecurity, compliance, and IT management programmes for regulated organisations across the UK and the Netherlands for over 20 years. View full profile

Talk to Cyvra

Map your shadow AI exposure

We help you identify which AI tools your staff are using, build a policy they will follow, and deploy enterprise tools that give them a compliant alternative.

Disclaimer: This article is for general informational purposes only and does not constitute legal, regulatory, or professional advice. Cyvra makes no warranty as to the accuracy or completeness of this content. Readers should seek independent advice appropriate to their specific circumstances. Cyvra accepts no liability for any loss arising from reliance on this content.