No items found.
November 19, 2026
November 18, 2026
Share :

AI Shadow IT: the risk that 47% of organizations have not yet formalized

Table of contents

Key takeaways

‍

  • The problem: Employees are using AI tools (ChatGPT, Gemini, unconfigured Copilot, etc.) without IT department approval. This is Shadow AI, and it affects almost every organization.
    ‍
  • Key statistic: 47% of organizations score 0 or 1 out of 5 in terms of preparedness for this risk (2025 IT & AI Data Barometer).
    ‍
  • Real risks: Sensitive data leaks, GDPR non-compliance, and loss of control over outbound data flows to third-party LLMs.
    ‍
  • Priority action: There is no need to ban these tools. You must provide a framework, raise awareness, and offer secure, sovereign alternatives.

‍

What is Shadow AI?

‍

Shadow AI refers to the use of artificial intelligence tools by employees without the validation, knowledge, or oversight of the IT department.

It is not sabotage. It is not even deliberate negligence. It is a shortcut. A salesperson copies meeting notes into ChatGPT to rephrase them. An HR professional uses Gemini to draft a job posting. A legal expert pastes a contract into an AI assistant to extract key clauses. In each of these cases, potentially sensitive data leaves the organization's perimeter, transmitted to a third-party language model with zero traceability.

That is Shadow AI.

‍

The difference between Shadow IT and Shadow AI

The question comes up often: "What is the difference between traditional Shadow IT and Shadow AI?" It is real and fundamental.

‍

Indicator Result
Overall average readiness score 1.82 / 5
Average score — public sector 1.23 / 5
Average score — private sector 2.09 / 5
Organisations scoring 0 or 1 / 5 47%
Organisations scoring maximum (5 / 5) 3%

‍

Traditional Shadow IT existed when an employee used Dropbox to share files or Slack when the company had approved Teams. Shadow AI is different in nature: data is no longer just stored elsewhere—it potentially feeds a model. And that changes everything.

‍

Why Shadow AI is spreading so fast

‍

The mechanism is simple. Brutally simple.

Getting a response from ChatGPT takes 30 seconds. Getting an internal AI project approved can take six months. When an employee has a problem to solve right now, the choice is easy—and you can't blame them.

‍

"What is AI Shadow IT?" is a question mostly asked by CIOs after the fact, once they realize the scale of the phenomenon. For employees, there is no question. There is a free, efficient tool available in two clicks. And an internal process that takes time.

‍

Three factors are fueling its spread:

  1. The accessibility of consumer LLMs : a browser and a personal email address are all it takes. No network administrator sees the request go through.
  2. Productivity pressure : teams are under strain. Shortcuts that work quickly become habits.
  3. The lack of internal alternatives : when the organization doesn't offer a vetted AI tool, employees take matters into their own hands.

‍

The result: Shadow AI isn't just the work of a handful of tech-savvy employees. It is systemic. According to Cisco, 80% of employees in a company use some form of shadow IT. With the widespread adoption of consumer LLMs since 2023, this figure has only grown—and the AI component along with it.

‍

‍

What the 2025 Datas AI CIO Barometer says

‍

"Why is AI Shadow IT dangerous?" The question deserves a factual answer. The 2025 Datas AI CIO Barometer provides one.

‍

The numbers that stand out

‍

Classic Shadow IT AI Shadow IT (Shadow AI)
Object SaaS application, software, personal device LLM, generative assistant, no-code AI tool
Risk vector Data leakage, non-compliance Data leakage + model training + uncontrolled outputs
Spread speed Weeks / months Hours (a browser + an email address is all it takes)
Detection Possible via network audit Very difficult (usage via standard browser)
GDPR exposure Moderate High (data potentially ingested by third-party servers)

‍

47% of organizations have implemented neither a formal policy on employee AI usage, nor a proxy to monitor traffic to LLMs, nor a structured awareness program. Virtually nothing at all.

‍

‍

A nuanced look at public vs. private sectors

‍

The gap between the public sector (1.23/5) and the private sector (2.09/5) is worth noting. It does not mean the public sector is more exposed to risk. It reflects a different structural reality.

The public sector has compliance mechanisms and a culture of access control that mechanically limit individual Shadow IT: a public servant has less leeway than a private-sector employee to spontaneously sign up for an AI service using their work email. However, when Shadow AI does take hold in the public sector, it is often less visible—and therefore harder to address.

The private sector shows a better "preparedness" score because it has a greater need to act: the risk exposure is higher, and the consequences (contractual, reputational, competitive) are more immediate.

‍

‍

The 4 concrete risks for your organization

‍

"Examples of AI Shadow IT in the workplace" — here is what that looks like in practice, and why each case poses a real problem.

‍

1. Leakage of sensitive data to third-party LLMs
‍
An employee pastes a supplier contract, a financial statement, or a patient file into ChatGPT. The terms of service for consumer-grade LLMs vary—and do not guarantee that data won't be used to train future models. The result: confidential information leaves the organization's perimeter without a trace.

‍

2. GDPR non-compliance
‍
GDPR requires that all processing of personal data be regulated, documented, and localized. An unapproved AI tool that processes HR data, customer information, or user files constitutes undeclared processing. In the event of a regulatory audit, the organization is in violation—even if the employee acted in good faith.

‍

3. Loss of control over outputs
‍
An LLM can hallucinate. It can generate incorrect information presented with confidence. If an employee uses the output of an unvalidated AI tool to draft an official document, a summary note, or a response to a user, the organization assumes liability based on unverified information.

‍

4. Impossibility of audit and traceability
‍
Who used what? When? With which data? In an uncontrolled environment, these questions remain unanswered. This is a problem in the event of a security incident, as well as during litigation or internal audits.

‍

‍

How can you regulate AI usage in the workplace without stifling innovation?

‍

This is the question most CIOs and executive teams are asking today: "Corporate AI governance — where to start?"

The short answer: not with a ban.

Blocking ChatGPT at the network proxy level is a technical response that doesn't solve the problem. Employees will switch to their smartphones, personal connections, or other tools. You have blocked visible access, but you haven't reduced usage.

"AI usage policy for employees" — that is indeed the starting point. But a policy alone is not enough. Here is a 5-step approach.

‍

5 actions to manage Shadow AI

A progressive approach — without blocking innovation

1

Step 1 — Understand

Map actual usage

Before deciding, you need to know. Which AI tools are your employees already using? For which tasks? A quick audit often reveals widespread, undeclared usage.

2

Step 2 — Formalise

Draft an AI usage policy

A short, clear, practical document. What employees can do, what they cannot, and why. Signed off by the executive committee and communicated to everyone.

3

Step 3 — Equip

Provide approved AI alternatives

If the organisation does not offer an approved AI tool, employees will use whatever is available. Sovereign solutions, hosted in France, meet the need without the risks.

4

Step 4 — Monitor

Deploy a traffic control proxy

Monitor what is being sent to third-party LLMs. Not to penalise — but to detect, alert, and measure the true scale of Shadow AI within the organisation.

5

Step 5 — Train

Raise awareness — for real

Not a mandatory e-learning nobody reads. Short, concrete awareness sessions, with real-world examples and approved alternatives presented at the same time.

‍

‍

Shadow AI is not inevitable — it is a project to be managed

‍

The real risk isn't that your employees are using AI. The real risk is that they are doing so without a framework, without alternatives, and without your knowledge.

A national score of 1.82/5 is not a condemnation. It is a starting point. Organizations that take action now—mapping, policy, tooling—gain a decisive head start. Those that wait for an incident to occur will be managing a crisis instead of preventing one.

AI governance is not an IT project. It is an organizational project.

‍

FAQ

‍

Frequently asked questions

The real questions decision-makers ask about AI Shadow IT

Yes — as soon as its use has not been approved by IT. ChatGPT used with a personal account, on a standard browser, without a formalised usage policy: that is Shadow AI. What matters is not the tool itself, but the context in which it is used. A ChatGPT Enterprise instance configured and approved by IT is not Shadow AI. A free ChatGPT account used to rephrase a client contract is.
No — and organisations that have done so have regretted it. Blocking access at the network level does not eliminate usage: employees simply switch to their smartphones or personal connections. You lose visibility without reducing risk. The right response is a clear framework: what is allowed, with which data, on which tools — backed by approved, sovereign alternatives.
AI Shadow IT — also known as Shadow AI — refers to employees using artificial intelligence tools without IT validation or oversight. It is rarely a malicious act: it is most often a productivity shortcut. A sales rep pasting a meeting summary into ChatGPT, an HR manager using Gemini to draft a job posting… In each case, potentially sensitive data leaves the organisation's perimeter with no traceability.
Any processing of personal data through an unapproved tool constitutes an undeclared processing activity under GDPR. If HR data, client records or user files pass through a third-party LLM, the organisation may be in breach — even without any malicious intent. In the event of a supervisory authority audit, the organisation is liable, including when a single employee acted alone.
The effective 5-step sequence: map actual usage → formalise a clear policy → provide sovereign AI alternatives → monitor outbound traffic → deliver concrete training. Banning does not work. What works: replacing the void with a framework and approved tools that employees actually want to use.

‍

‍

👉 Would you like to assess your level of preparedness for AI Shadow IT?
[Contact our teams for a maturity audit.]

‍

Further articles you maybe interested in

No items found.
October 20, 2026

AI, LLMs, and agents: how to make the right choices for your business?

LLMs, AI agents, RAG: the promises are plentiful, but the majority of generative AI projects never make it to production. Why is this? The problem isn’t the model; it’s the data. Find out how to distinguish between LLMs and AI agents, choose between cloud and on-premises solutions, and prioritise structuring your data to achieve measurable results: +30% relevance, -40% latency.
IT management
October 27, 2026

AI Hallucinations: Understanding and Mitigating Errors

96 % des entreprises françaises ne sont pas prêtes pour l'IA : données non structurées, RAG mal alimenté, manque de gouvernance. Découvrez les trois principales causes des « hallucinations » de l'IA et les cinq piliers pour améliorer la pertinence.
No items found.
No items found.
No items found.
No items found.