Shadow AI: what it is, how to spot it and how to reduce it

What is shadow AI?

Shadow AI is any use of an AI tool for company work without the knowledge or approval of the people responsible for data and systems. The name comes from shadow IT, meaning software and services employees set up on their own.

The difference is that an AI tool usually receives content: documents, emails, code, meeting recordings. Once that content reaches an external service, it is hard to take back.

Shadow AI rarely comes from bad intent. People want to finish work faster, and the company has not given them a tool or told them what to use.

Increasingly, shadow AI does not even need an employee's decision. Vendors of software the company already uses add AI features in updates. Meeting notes, email summaries or an editor assistant can start sending content to a new subprocessor nobody in the company knows about.

Examples of shadow AI at work

An employee pastes a contract into a personal AI chat account to get a summary. A salesperson uploads a customer list so the AI can write personalised emails. Someone invites a bot to a client meeting that records it and takes notes.

Less obvious cases: a browser extension with an AI assistant that reads every open page, including the HR system. An AI feature switched on automatically in software the company already uses. An AI agent connected to a mailbox or drive, which we cover in AI agent security. A developer pasting a log containing an API key into a chat.

Shadow AI can also sit on your website: an AI chat widget added by an agency or plugin that sends customer conversations to an external service. You can check for free which third parties your site shares data with.

How common is shadow AI?

According to the Microsoft and LinkedIn Work Trend Index 2024 (a survey of 31,000 people in 31 countries), 78% of AI users bring their own AI tools to work, rising to 80% in small and medium businesses.

According to IBM's Cost of a Data Breach Report 2025 (600 organisations, breaches from March 2024 to February 2025), one in five organisations reported a breach involving shadow AI, and only 37% had policies to manage AI or detect shadow AI. Organisations with high levels of shadow AI paid about USD 670,000 more per breach on average.

IBM published the 2026 edition on 29 July 2026. We did not find new shadow AI figures in IBM's own press materials, so we quote the 2025 edition.

Why is shadow AI a risk for companies?

Data leaves without a contract. Personal accounts come with no data processing agreement, which Article 28 of the GDPR requires, and on some individual plans conversations can be used to train models. We cover ChatGPT in detail in can you paste company data into ChatGPT.

There is no trail. The company does not know what was sent and when, so after an incident it cannot assess the scope or notify the people affected.

Decisions rest on unchecked output. Models make mistakes, and a tool that reads web pages and emails can be manipulated by hidden instructions, as explained in what is prompt injection.

It is harder to show compliance with Article 4 of the EU AI Act, which requires companies using AI to support their staff's AI literacy. You cannot train people on tools you do not know exist.

How to detect shadow AI in your company

1. Ask. A short anonymous survey (which AI tools you use, for what, on which account) tells you more than weeks of log analysis. Say upfront that answers will not lead to sanctions.

2. Review app consents in Google Workspace or Microsoft 365. AI tools often request access to mail, calendar and drive through sign-in with a company account.

3. Check browser extensions on company computers and the admin panels of software you already use. AI features are sometimes switched on by default in an update.

4. Look at expenses. AI subscriptions paid with a company card or claimed back show what your team actually needs.

5. If you run DNS filtering or a proxy, check traffic to popular AI service domains. It shows scale, not content, and it does not cover personal phones.

What to do with the AI tools you found

Put every tool you found into one of three groups. Keep: move it to a business plan, sign a data processing agreement and add it to the approved list. Replace: people need the function, but the tool does not meet your conditions, so point them to an approved alternative. Remove: the tool is not needed or the risk is too high, so revoke access and app consents.

Separately, check what has already left. If personal data went into a tool, whoever is responsible for data protection assesses whether it is a breach that must be reported to the supervisory authority within 72 hours (Article 33 GDPR). If keys, passwords or tokens went in, revoke them and issue new ones. Deleting a conversation does not undo the exposure.

Finally, record the outcome: which tools were found, what was decided and when. That record shows the company is in control of the AI systems it uses, which helps in an audit and when a customer asks about your rules.

How to reduce shadow AI without banning it

A ban without an alternative moves usage to personal phones, where you see even less. It works better to give people what they are looking for on the company's terms.

Step 1. Approve one or two tools on a business plan with a data processing agreement. Step 2. Write a short policy: what may be pasted, what may not, who checks output. There is a ready template in AI use policy for companies. Step 3. Set up a fast way to request new tools, with an answer in days, not months.

Step 4. Train the team and record who completed the training. Step 5. Announce that reporting your own mistake, such as pasting customer data, does not lead to sanctions. Without that, nobody mentions a problem until it is too late. Step 6. Repeat the survey after a quarter.

Measure two things: how many people use the approved tools and how many requests for new tools come in. Growing use of approved tools and a steady flow of requests mean shadow AI is shrinking. Silence usually means people are doing it their own way again.