Shadow AI: your team already uses ChatGPT. How to manage it without banning it

Shadow AI is when employees use artificial intelligence tools (ChatGPT, Claude, Copilot, or Mistral) for their work, with their own accounts and no framework set by the company. It’s already happening in most small businesses. The right answer isn’t a ban: it’s a short AI policy, business accounts under contract, and tools convenient enough that the right way is also the easiest.

I build AI tools for small businesses, and on a first call I often see the same scene play out: the owner finds out that a sales rep pastes their meeting notes into ChatGPT Free, and that the accountant has an online assistant summarize spreadsheets. Nobody did anything wrong: they found a tool that saves them time. What they need is a framework. This article isn’t legal advice: for sensitive data processing, have your setup reviewed by your data protection officer or legal counsel.

What is shadow AI?

The term comes from “shadow IT”: software that teams install without going through the company. Shadow AI is the newer version, and it’s even easier: all it takes is a browser and an email address. Three fictional examples, typical of what I see:

• at Durand Négoce, Alice pastes in an angry customer’s email, with the customer’s name and order number, to get a more diplomatic reply;
• a sales rep has a quote rewritten, prices and discounts included, to make it more persuasive;
• accounting uploads a general ledger export to get a summary before the leadership meeting.

Each time, company data goes to a service the company has no contract with, on an account it doesn’t control.

Shadow AI: what do the numbers say in France?

The studies point the same way, even though several come from vendors selling their own tools: read them as orders of magnitude.

Finally, Insee (France’s national statistics office) ranks data protection and privacy as the second-biggest barrier among companies already using AI (43%). Leaders see the risk; most simply haven’t written the rule yet.

What are the real risks of shadow AI?

Customer data and trade secrets in a personal account

A personal account belongs to the employee. The company doesn’t know what’s been put in it, can’t delete it, and loses it the day the person leaves. Depending on the plan and settings, conversations may also be used to train the vendor’s models (see the table below). France’s cybersecurity agency, ANSSI, is clear: its recommendation R34 calls for banning online generative AI tools for professional use involving sensitive data.

No record for the company, no integration

What happens in a personal account leaves no trace in your software. No one can find out which reply was sent to a customer, or on what basis. And the gain stays individual: copy-pasting replaces retyping, but the manual step is still there.

AI and GDPR: who’s responsible, under what contract?

As soon as personal data (a customer, an employee, a job applicant) goes to an AI service on the company’s behalf, the company is accountable. If you handle EU personal data, the GDPR requires a data processing agreement with the provider (Article 28), safeguards for transfers outside the EU (Chapter V), and an entry in the record of processing activities (Article 30). With a consumer account an employee opened on their own, none of these three is covered.

Inaccurate answers reused without checking

An AI model can be wrong with total confidence. France’s data protection authority, the CNIL, says good practice is to “never reproduce the outputs as is” when using these tools. A made-up figure in a reply to a customer commits the company, not the tool.

Key point The danger isn’t that your team uses AI. It’s that they do it on accounts the company can’t see, with no contract, no record, and no rule about what gets pasted in.

Should you ban ChatGPT to stop shadow AI?

Bans have been tried, often after a data leak. A global Cisco study (January 2024) found that 27% of organizations had banned generative AI at least for a time, and that 48% of respondents admitted entering non-public information into it.

But a ban on its own works poorly: according to the Forvis Mazars survey, shadow AI persists despite a clear ban in 12% of companies, and the most common response is still to tolerate it. An employee who saves an hour a day won’t easily give up their tool; they just stop talking about it. Visible, managed use beats hidden use.

ChatGPT, Claude, Copilot, Mistral: personal account, business plan, or API?

The right question isn’t “which model?” but “which plan?” Here’s what the vendors say, as of this article’s date. Their terms change often, so reread them when you choose.

Two things to keep in mind. First, “no training” doesn’t mean “nothing leaves”: with a hosted model, your data leaves your server while it’s processed, under contract. Second, data location: OpenAI offers a “Europe” region for its API, subject to approval; Anthropic’s direct API doesn’t, so you go through AWS Bedrock or Google Vertex AI in a European region. Without a European region, the data goes to the United States, and the transfer then has to rest on a valid legal framework, which you check in the provider’s contract.

AI acceptable use policy: what should it include?

The CNIL recommends governing use with a charter or internal policy that sets out what’s allowed and what isn’t, banning certain data from certain tools, training users, and involving the data protection officer. A useful AI acceptable use policy fits on one page: any longer and nobody reads it. Here’s the AI policy template I suggest; adapt it to your business, with your counsel. The canton of Geneva, in Switzerland, also publishes a more complete template, useful if you want to go further.

Key point An effective AI policy says three simple things: which tools, which data stays out, and who to call. Point 7 matters more than it looks: if owning up to a mistake gets punished, nobody owns up to anything.

How to manage shadow AI in 5 steps

1. Map current use, without a witch hunt. An anonymous four-question survey: which tools, for which tasks, with what data, how much time saved. You’ll see where the real needs are.
2. Pick the accounts. Business accounts under contract for regular users, rather than a license for everyone.
3. Sign the contracts and keep the records. Data processing agreement, transfers, entry in the record of processing activities.
4. Write the policy and present it. One page, presented at a team meeting with concrete examples from your line of work.
5. Plug AI into your software for the workflows that matter. When a task comes up every day (answering customer questions, preparing quotes, summarizing files), copy-pasting into an assistant is no longer the right answer. A tool connected to the CRM or ERP, in your name, does the work within each person’s access rights and keeps a record. That’s the job of an AI integration or an internal chatbot that answers from your documents.

This last step is the one that really shrinks shadow AI: when the official tool is more convenient than a personal account, habits change on their own. The connection rules I apply (account in your name, data kept to the minimum, access rights inherited from the software, a log) are detailed in my article on the Sage 100 API, with an ERP as the example.

AI and GDPR: 6 points to check

• The data processing agreement (Article 28) with every AI provider that processes personal data for you.
• Transfers outside the EU: a European region when one exists; otherwise, the legal framework for the transfer.
• Data minimization: send the model only the data the task needs, and pseudonymize when possible.
• Inherited access rights: a tool connected to your software only sees what the person using it can see. ANSSI also recommends regularly reviewing AI tools’ permissions on business applications (R35).
• The record and the log: the processing is listed in your record of processing activities (Article 30), and exchanges are logged (ANSSI recommendation R29).
• Informing people: your customers and employees know when an AI processes their data, and a DPIA (data protection impact assessment) is considered for high-risk uses, as the CNIL suggests.

The EU AI Act adds a general AI literacy obligation for companies that use AI (Article 4): giving the people who use it at least basic training is part of it, and a policy presented to the team is a first step. For a chatbot open to your customers, the obligations are more specific: see my article on the EU AI Act and customer service chatbots.

What I don’t recommend

• Banning without an alternative. Usage goes underground, and you lose the one thing that mattered: knowing what’s going on.
• Buying licenses for everyone with no policy or defined use. You pay for accounts that barely get used, and shadow AI carries on alongside them.
• Trusting a “GDPR-certified tool.” No such certification exists for a tool. Ask for the measures: contract, region, retention, access rights, log.
• Making your team feel guilty. The people using AI have often spotted the most tedious tasks. They’re your best allies for choosing what to automate.
• Connecting an assistant with an admin account. An innocent question can then pull up data from another department or another customer.

If you’re unsure what kind of tool should replace personal accounts (an assistant, a chatbot, or an agent that acts in your software), my guide to AI agents, chatbots, and automation will help you decide. Consulting firms and agencies, which handle their own clients’ data, are among the first concerned.