Shadow AI in your business: what it is and how to handle it

Shadow AI is AI that employees use for work without the company choosing or approving the tool. Think of a free chatbot on a personal account, a browser extension that rewrites emails, or an AI feature quietly switched on in software you already use. Usually nobody means any harm. The problem is that business data ends up in tools nobody keeps an eye on.

In this article, you will learn what shadow AI is, why it happens in almost every business, and what risks come with it. You also get a four-step approach that works without banning everything. If you want a list of checks right away, use our checklist for AI and business data.

A grid of dark blocks on a black plinth, with a few cobalt blue blocks glowing softly in the shadow of a tall wall
Shadow AI is rarely meant to cause harm. It simply happens out of the company’s sight.

What is shadow AI?

The term comes from shadow IT: software that employees use without the IT team knowing about it. Shadow AI is the same thing, but with AI tools. The tool itself is not the problem. A chatbot, a translation tool, or a writing assistant can be fine. What matters is use that falls outside the agreed rules and out of the company’s sight.

The difference comes down to three questions. Does your business know which tool it is? Does it know who uses the tool? And does it know which data goes into it? A chatbot your business set up with a business account and clear rules is not shadow AI. The same chatbot on a personal account, with customer data in it, is.

How common is shadow AI?

The numbers are high. In the Alert Online 2026 cybersecurity behavior study by ECP and BehaviorChange Group, 49% of employees in the Netherlands say they sometimes use a personal AI account for work. According to the researchers, that can lead to shadow AI. The survey ran in June and July 2026 among 378 employees, and it reflects what people say they do. The report is in Dutch.

Leaders see it too. In its November 2025 press release, Gartner writes: "A Gartner survey of 302 cybersecurity leaders in March – May 2025 revealed that 69% of organizations suspect or have evidence that employees are using prohibited public GenAI."

Why do employees use AI tools on their own?

Usually to work faster. Rewriting an email, summarizing a long document, or translating a text takes a few seconds with a chatbot. If the company has no approved alternative, employees find one themselves. Sometimes there is a tool, but it is slower, more limited, or hard to find.

A second reason is that nobody knows the rules. There is no policy, or the policy sits in a folder nobody opens. Then a personal account is the fastest route, and it does not even seem forbidden. So shadow AI mostly shows that your team wants to use AI. It is rarely a sign of bad intent.

Shadow AI also takes more forms than a chatbot in the browser. The table shows the forms you run into most often, and the first question to ask about each one.

FormExampleWhat can happen to the dataFirst question

Free chatbot on a personal account

Having a customer email rewritten in a personal account

The provider can store chats and, depending on the settings, use them to train models

Which data must never go in?

Browser extension with AI

An extension that summarizes emails or web pages

The extension can read the pages you open, including your email or your CRM

Who installed the extension, and what can it read?

AI feature in existing software

A writing assistant or summary in your word processor, CRM, or video call

Data goes to an AI service you did not choose separately

Is the feature on, and what do the terms say?

Meeting notes or transcription tool

A tool that listens in on a meeting with a customer

Conversations are recorded and processed outside your business

Does everyone in the meeting know a tool is listening?

Homemade script or integration

A colleague who connects an AI model to a spreadsheet with a personal API key

Keys and data sit in places nobody tracks

Where is the key, and who pays for the usage?

What risks does shadow AI bring?

The biggest risk is data that leaves your business without anyone knowing. The Dutch Data Protection Authority (AP) received several reports of data breaches because employees shared personal data with an AI chatbot. In its August 2024 warning (in Dutch), the AP explains that employees often use these chatbots on their own initiative and against the agreements with their employer. If they enter personal data, that counts as a data breach. In one case, an employee of a general practice entered patients’ medical data into a chatbot.

Belgium has the same European privacy rules. The Belgian Data Protection Authority explains in a brochure on AI systems and the GDPR which rules apply as soon as an AI system processes personal data.

Beyond personal data, there are three more risks. Confidential business information, such as quotes, contracts, or figures, can end up in a tool you do not know. There is no log, so afterward nobody knows which data went where. And when an employee leaves, the personal account with the business data simply stays.

Should you ban AI tools?

A ban seems like the safest choice, but it rarely works. The tasks remain, so the use moves to a personal phone or a home computer. Then you have even less visibility. The AP leaves the choice to you, but it does ask for clear agreements: may employees use chatbots, and if so, which data may they enter and which not?

How do you handle shadow AI?

A better approach is to map what is happening, make agreements, and give your team an approved alternative that works just as well. You do that in four steps.

  1. 01

    Map which AI tools your team uses

    Just ask. A short survey or a conversation in a team meeting tells you more than a technical check. Say upfront that honest answers have no consequences, or you will hear nothing. For each tool, write down who uses it, for what, and which data goes into it. Also look at browser extensions and at AI features already switched on in software you use.

  2. 02

    Make short agreements

    Write down in a few rules which data must never go into an AI tool, such as customer data, personal data, passwords, and figures that are not public yet. Agree that everyone uses business accounts, not personal ones. The Dutch Chamber of Commerce (KVK) describes what belongs in an AI policy (in Dutch). Name one person who maintains the policy and answers questions.

  3. 03

    Give your team an approved alternative

    People only stop using a personal tool when there is something better. Choose tools for the tasks you found in step 1, set up business accounts, and get the settings right, for example so chats are not used to train models. Make sure the alternative is just as easy to use as the tool people already had.

  4. 04

    Follow up and adjust

    You do not deal with shadow AI once. New tools and features appear every month. Repeat the inventory a few times a year and adjust the agreements when the work changes. Also ask what the approved alternative is still missing. That is where the next shadow AI starts.

Where does your own AI agent fit in?

An approved alternative can be a business subscription to a chatbot. If you want more than chat, for example tasks carried out in your CRM or your inbox, your own AI agent can be that alternative. Hermes is such an open-source agent. It runs on a computer you choose and works with the systems you connect to it.

The difference with shadow AI is control. With your own agent, you decide which language model it uses and which data may go to an outside provider. You give it only the permissions a task needs for each system, often read-only at first. Voltti never installs Hermes on a publicly reachable server: the agent sits behind a firewall, and you reach it through your company network or a VPN. Read more about those choices on our security and privacy page.

Your own agent does not make shadow AI disappear by itself. Employees can still open a personal account, so agreements and follow-up remain necessary. But you do give your team a place where AI work is visible and where the rules are already part of the setup. To see which data stays local and which goes to a cloud model, read what stays local in a hybrid AI setup. You can weigh both options in local AI vs cloud AI.

How does Voltti do it?

Our own Hermes installation at Voltti follows the same rules. The agent reads our Search Console data with its own read-only account. Each environment has its own settings and its own access, so work for one project does not end up in another. The agent edits content in our CMS, and a person reviews that work afterward and makes corrections where needed.

Where do you start?

Start with the inventory from step 1. It shows which tasks your team already does with AI, and that is usually where your first approved use case sits. In a pilot, you test that use case with the permissions the task needs, and no more. A pilot usually takes 2 to 4 weeks, from the first conversation to the evaluation. Voltti works with businesses in Belgium and the Netherlands. Read how it works in our approach.

Portrait of Seppe Gadeyne

Seppe Gadeyne is a developer and the owner of Voltti. He sets up AI agents for companies in Belgium and the Netherlands.

Updated on

Frequently asked questions

01What is the difference between shadow AI and shadow IT?

Shadow IT is any software employees use without the company knowing. Shadow AI is part of it: it covers AI tools such as chatbots, writing assistants, or note-taking tools. With AI tools, you often enter text and documents. That is why more business data leaves the company than with a lot of other software.

02Should you ban ChatGPT?

You do not have to. A ban often pushes the use to personal devices, where you have even less visibility. It is better to agree which data must never go in, use business accounts, and get the settings right. The Dutch Data Protection Authority mainly asks for clear agreements.

03Is shadow AI a GDPR problem?

Often, yes. If an employee enters personal data into a chatbot against the agreements, the Dutch Data Protection Authority considers that a data breach. In many cases, you have to report such a breach to the regulator and to the people whose data it is. In Belgium, you report a data breach to the Data Protection Authority.

04How do you find out which AI tools your team uses?

Ask. A short survey or a conversation in a team meeting tells you more than a technical check, especially if you say upfront that honest answers have no consequences. Also look at browser extensions, at AI features in software you already use, and at AI subscriptions that employees submit as expenses.

05Does your own AI agent solve shadow AI?

Not by itself. Your own agent gives your team an approved alternative with permissions per system. That leaves less reason to use a personal tool. Agreements and follow-up remain necessary, because a personal account is always just a few clicks away.

Want to give your team a safe AI alternative?

Discuss your pilot