One Monday morning, a recruiter pastes thirty CVs into her personal ChatGPT account to save an hour. Nobody told her not to. Nobody offered her anything else either. That is Shadow AI: the AI your teams already use, outside any framework.

At AI Sisters, an AI consulting firm, we see this in almost every company we work with. Here is what Shadow AI covers, what the serious studies measure, the real risks, and a five-step plan to deal with it without putting anyone’s back up.

In brief

  • Shadow AI refers to the use of AI tools that the company has neither provided nor approved: personal accounts, browser extensions, meeting recorders.
  • In France, 67% of employees who use AI rely on tools not provided by their employer, according to OpinionWay for Cegid (September 2026).
  • The main risk is data leakage: customer information, unpublished figures, internal documents.
  • Banning does not work: usage moves to personal phones and becomes invisible.
  • What works: an approved tool, a one-page policy and training. The AI Act has also required companies to train their teams since February 2025.

Read also: AI Act: the obligation to train your teams in AI and our guide to bringing AI into a company.

Shadow AI: what are we talking about?

Shadow AI is the use by employees of artificial intelligence tools that the company does not know about or has not authorised. The term comes from “shadow IT”, software installed without going through the IT department.

It takes three forms. The most common: a personal ChatGPT, Claude or Gemini account used for work tasks. Next, browser extensions that read the open page, including a CRM or an email inbox. Finally, tools that plug themselves into your systems: a meeting recorder connected to the calendar, an agent linked to the mailbox.

In the vast majority of cases, there is no ill intent. The employee has a task to finish, a free tool does it in two minutes, and the company has given them nothing equivalent.

How Shadow AI starts. A real need: summarise a file, write a follow-up, analyse a spreadsheet. An unapproved tool, because none was provided: personal AI account, browser extension, meeting recorder. What leaves, without anyone knowing: customer data, unpublished figures, internal documents. Nobody decided it: Shadow AI mostly comes from the lack of an approved tool and a clear rule.
Shadow AI almost always starts from a legitimate need.
To put it simply. Pasting a document into a personal AI account is like emailing it to your personal address to read at home. The intention is good, but the document leaves the company and you no longer know who keeps it, or for how long.

What the studies say

Figures on Shadow AI circulate widely, and not always with their source. A “73%” often comes up with no identifiable study behind it. Here are the ones that can be traced to a published method.

Shadow AI: the verifiable figures in 2026
StudyWho was surveyedWhat was measured
OpinionWay for Cegid, September 20262,006 employees in Europe, including 500 in France67% of French employees who use AI rely on tools not provided by their company. Only 39% have tools provided
WatchGuard, July 2026684 employees of companies with 50 to 500 staff, 8 countries64% admit using unauthorised AI tools
IBM, Cost of a Data Breach 2025Organisations that suffered a data breach1 in 5 had a breach linked to Shadow AI. High usage adds an average of $670,000 to the cost
IBM, same reportSame organisationsOnly 37% have a policy to govern AI or detect Shadow AI

Two caveats. The WatchGuard and OpinionWay surveys rely on what employees report, so the real figures are more likely higher than lower. And the IBM figure does not mean that every breach costs $670,000 more: it is the average gap between organisations where Shadow AI is widespread and those where it is rare.

The underlying finding is consistent from one study to the next: around two out of three employees who use AI do so outside the intended framework, and most companies have neither an official tool nor a rule that everyone knows.

Why banning is not enough

Management’s first reflex is often to block ChatGPT on the network. The most cited example dates from May 2023: after engineers pasted confidential source code into ChatGPT, Samsung restricted generative AI tools on its devices, according to Bloomberg.

Blocking protects the company network, not the employee’s phone. On Reddit, system administrators say it again and again in threads on the subject: when management blocks without offering an alternative, usage continues, but on devices nobody can see any more. Visible Shadow AI becomes invisible Shadow AI.

The other effect is less visible: teams that respect the ban fall behind those that work around it. Your most conscientious employees are penalised.

To put it simply. Banning AI without offering anything is like shutting down the coffee machine on your floor. Nobody stops drinking coffee: everyone goes to the café across the street, and office conversations happen there out loud.

The real risks, without the drama

Shadow AI exposes the company to specific risks. It is better to name them one by one.

Shadow AI risks and what protects you
RiskConcrete exampleWhat protects you
Data leakageA customer file pasted into a free account whose conversations can be used to train the modelA business account that excludes your data from training
Personal data and GDPRCVs or employee records processed outside any registerA clear rule on HR data and an approved tool
Trade secretsFinancial results analysed before they are publishedA list of data that never goes into any AI
Unchecked errorsA figure invented by AI copied into a management memoTraining in checking outputs
AI Act complianceTeams using AI without training, even though Article 4 requires itA documented training programme

The last point is often overlooked. Since 2 February 2025, Article 4 of the AI Act has required companies that use AI systems to ensure their teams have a sufficient level of AI literacy. Widespread, ungoverned use also means a training obligation that is not being met. We covered this in detail in our article on the AI Act training obligation.

Give your teams a tool and a method

Training on the tool you have chosen, with your own documents and a policy adapted to your data.

Our training by toolBook a call

Three common situations, and what to do

1. HR and CVs

Screening applications takes hours, and an AI reads thirty CVs in a minute. The problem: this is personal data, often with address, age and career history. In a free account, it falls outside any GDPR framework. It is better to do this screening in an approved tool, after removing information that is irrelevant to the decision. A good first task to suggest: writing job ads, which contain no personal data.

2. Finance and unpublished figures

A financial controller pastes the month’s export into an AI to get a variance commentary. The time saving is real, but this file sometimes contains figures that no competitor should see before the results are announced. Here, the rule must be clear: this data only goes into an approved tool, never into a personal account. Our AI training for finance starts from this kind of case.

3. Sales teams and meeting recorders

An automatic note-taking tool connects to the calendar, joins video calls and sends a summary to every participant, clients included. This is the most discreet form of Shadow AI, because the tool installs in one click and acts on its own. You need to decide which tools are allowed to connect to email and calendars, and say so clearly. For teams that want to go further with this kind of tool, our AI agents training teaches how to keep them under control.

The five-step plan

The goal: make the official tool more convenient than the hidden one. Here is the order we follow with our clients.

The five-step plan against Shadow AI. 1, assessment: an anonymous questionnaire, with no sanctions. 2, approved tool: a business account for each team. 3, one-page policy: three levels of data, concrete examples. 4, training: required by the AI Act since February 2025. 5, follow-up: usage and tools reviewed every quarter. The order matters: without an approved tool, the policy and the training remain theoretical.
The five steps, in the order in which they have an effect.
  1. Take stock, with no sanctions. An anonymous five-question survey: which tools, for which tasks, with which data. If the answers can cost those who give them dearly, nobody will tell the truth.
  2. Provide an approved tool. ChatGPT Enterprise, Claude or Copilot, depending on your existing tools. Business plans exclude your data from model training by default. Our AI tools for business page compares the options.
  3. Write a one-page policy. Three levels of data, with examples taken from your business (see the diagram below). A twenty-page rulebook will not be read.
  4. Train your teams. On the chosen tool, with their real documents. Training is what shifts usage to the official tool, and it is also what the AI Act requires. Our training is eligible for OPCO funding (French training funds).
  5. Monitor and adjust every quarter. Usage rate of the official tool, new requests, newly spotted tools. To put a figure on the gains, see our method for calculating the ROI of AI training.
The policy in three levels of data. Anywhere, no risk if it leaks: texts already public, blank templates, ideas, outlines, rewording. Approved tool only, business account and protected data: internal documents, meeting minutes, everyday customer data. Never, whatever the tool: health data, figures before announcement, passwords and access codes. One page is enough: every employee should know in ten seconds where each type of data can go.
A policy you can read in ten seconds beats a complete rulebook.
To put it simply. The three-level policy works like recycling: nobody is asked to understand the chemistry of recycling, only to know which bin to use. Three colours, examples on display, and the right habit becomes automatic.

Where should your company start?

A 30-minute call to review how your teams use AI and choose the right tool.

Book a callFree AI assessment

Key takeaways

  • Shadow AI is the use of AI tools that the company has neither provided nor approved. It concerns around two out of three AI users.
  • It mostly comes from a real need and the lack of an official tool, rarely from bad intent.
  • Blocking tools moves usage to personal phones, where it becomes invisible.
  • The risks to address first: customer and HR data, unpublished figures, tools that connect themselves to your systems.
  • The plan that works: assessment, approved tool, one-page policy, training, quarterly follow-up.
  • Training your teams is not optional: the AI Act has required it since February 2025.

Sources