Shadow AI: Risks, Detection and a 6-Step Action Plan

Shadow AI is the use, for work, of AI tools the company has neither approved nor even recorded: a personal account on a consumer assistant, a browser extension that summarises pages, a tool that transcribes meetings, an AI feature switched on inside software you already use. The problem is not that employees use AI. It is that nobody knows which data goes where, or under which terms. The effective response is not a ban: it is a blame-free stocktake, approved tools that meet the same needs, short rules and a quick way to request a new tool.

This article is general information, not legal advice. Data protection and employment law govern both the use of the tools and how you may check on it; involve your data protection officer and, where required, employee representatives.

Why shadow AI appears

Shadow AI is the recent form of an old phenomenon, shadow IT. The UK National Cyber Security Centre (NCSC) explicitly includes unsanctioned AI tools in its shadow IT guidance and points out that staff usually turn to such tools to get their work done, not to break rules. The causes are almost always the same:

• The tool is one click away. No purchase, no installation, just a free account.
• The gain is immediate. Meeting notes in two minutes instead of thirty.
• Nothing is offered internally, or requesting a tool takes weeks.
• The rules are unknown or vague. "Use AI responsibly" does not say what is allowed.
• AI arrives inside existing software. An update adds an assistant to the email client or CRM without anyone deciding it.

The forms it takes

Form	Example	What the company loses sight of
Personal account on a consumer assistant	Pasting a contract to get a summary	The terms of use, retention and reuse of the data
Browser extension	An assistant that reads every open page	Access to webmail, CRM, intranet
Meeting transcription	A bot that joins video calls	Recording of third parties, storage location
AI feature in existing software	An assistant enabled by default in an HR tool	A new, sometimes sensitive use with no assessment
Personal automations	A flow that sends incoming emails to a model	Untracked API keys and data transfers
Connectors	An assistant linked to the team drive	Far broader access than what the user pastes

The real risks

• Leaks of confidential or personal data. The French data protection authority CNIL, for example, recommends not entering personal or confidential data into consumer services, using a business address and, where possible, switching off reuse of the data; see its questions and answers on generative AI (in French).
• No contractual basis. Without a data processing agreement or a check of where data is processed, handling personal data in the tool can conflict with the GDPR.
• Unchecked errors. A wrong result copied into an offer or a letter binds the company.
• Regulated uses nobody noticed. A tool used to screen applications can be high-risk under the EU AI Act without anyone having classified it.
• Lost know-how. Prompts, settings and history sit in personal accounts and leave with the person.
• No record. If a customer asks or an audit comes, the company cannot say what was done.

These risks exist with approved tools too; the difference is that there they are known and managed.

Finding shadow AI without spying on staff

Start with the simplest and least intrusive means.

1. Ask. An anonymous survey with five questions: which tools, for which tasks, with which data, with which account, and what is missing internally. State clearly that no sanctions will follow.
2. Listen to managers and teams. People talk freely about their use when the question is about the need, not the fault.
3. Look at spending. Expense claims and card subscriptions reveal paid tools.
4. Check connections made with company accounts. The admin console of your office suite lists third-party apps users have granted access to.
5. Use the technical means you already have. The NCSC names asset inventories and cloud access security brokers (CASBs) among the detection methods, and notes their limits.

The last two points touch on monitoring employees' activity. Before using them, check with your data protection officer and employment counsel what is allowed and what must be announced, and involve employee representatives where the law requires it.

A six-step plan

1. Take stock

Collect the answers in a table: tool, team, task, type of data, account used. That is the start of your AI inventory.

2. Sort the uses

Finding	Decision	Example
Legitimate need, non-sensitive data	Approve and add to the list	Rewording public marketing copy
Legitimate need, unsuitable tool	Replace with an approved tool	Summaries of internal documents on a personal account
Legitimate need, sensitive data	Put safeguards in place: business plan, contract, settings	Transcribing customer meetings
Use with regulatory risk	Pause and have it assessed	An assistant ranking job applications
No real need	Prohibit, with an explanation	An extension that reads every page

For uses that touch decisions about people, the EU AI Act risk checker gives a first orientation.

3. Offer approved alternatives

This is the step that makes shadow AI disappear. If the approved tool is less convenient than the personal account, hidden use will continue. Choose a business plan whose data terms you have read, with accounts tied to the company; our questions to ask AI vendors can serve as the checklist.

4. Write the rules

A short policy: list of approved tools, data rule, checking outputs, uses that need approval, contact. Our AI acceptable use policy template is ready to adapt.

5. Train

A short session with examples taken from the stocktake: what is allowed, what is not, and why. The EU AI Act asks for measures to support staff AI literacy in any case; see EU AI Act Article 4.

6. Open a fast track for new tools

A form of a few lines and an announced response time, for example five working days. The NCSC stresses this point: a quick request process keeps workarounds from coming back. Repeat the survey after six months to measure the effect.

If sensitive data has already been entered

1. Do not look for someone to blame. A good-faith report must have no consequences, otherwise nobody will report anything again.
2. Establish the facts: which data, which tool, which account, when.
3. Check the tool's terms and use the options to delete the history or the account.
4. Assess with your data protection officer whether it is a personal data breach. Depending on the risk, the GDPR provides for internal documentation or notification to the supervisory authority within 72 hours.
5. Fix the cause: was an approved tool missing, was the rule unknown?

Common mistakes

• Banning without offering. Use continues, on personal phones.
• Punishing the first people who speak up. You lose your best source of information.
• Looking only at chat assistants. Extensions, transcription tools and built-in features stay under the radar.
• Approving a tool without reading the terms of the plan you chose. They often differ between the free and the business plan.
• Dealing with the topic once. New tools appear every month.

Checklist

• An anonymous survey has been run and its results recorded.
• Every use found has a decision: approve, replace, safeguard, pause or prohibit.
• At least one approved tool covers the most frequent needs.
• Accounts are tied to the company.
• A short policy has been published and explained.
• Teams know whom to ask for a new tool and how quickly they will get an answer.
• A blame-free reporting channel exists.
• The data protection officer and, where required, employee representatives have been involved.
• A date is set for repeating the stocktake.

Sources

• NCSC (UK): Shadow IT guidance
• CNIL (France): Questions and answers on using a generative AI system (in French)
• European Commission: AI literacy questions and answers
• EUR-Lex: Regulation (EU) 2016/679 (GDPR)