Shadow AI: definition, risks and how to overcome it in business (2026)

Article published on December 14, 2024, rewritten and updated on October 24, 2026 based on reports from IBM, Microsoft, Netskope and Gartner, the Datacraft and Inria study, Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744 and the NIS2 Directive, consulted on that day.

Shadow AI refers to the use by employees of artificial intelligence tools that their organization has neither chosen, authorized, nor regulated: a personal ChatGPT account to summarize a contract, a browser extension that rephrases emails, an image generator for a client presentation. It's rarely malicious; it's a sign that teams have discovered a time-saving feature that the company doesn't yet offer. The risk, however, is very real: according to IBM, one in five organizations will have already experienced a data breach related to shadow AI by 2025. Here's how to recognize it, what it costs, what the law says, and how to get out of it without stifling innovation.

The essentials in five lines

  • Definition: any use of an AI tool for professional purposes outside of tools validated by the IT department, on a personal account or an unapproved application.
  • Scope: 78% of AI users at work bring their own tools (Microsoft, 2024), and 47% of generative AI users still use personal accounts (Netskope, 2026).
  • Cost : A leak involving a lot of shadow AI costs on average $670,000 more than another (IBM, 2025).
  • Legal framework : The GDPR, trade secrets and the NIS2 directive fully apply; Article 4 of the AI Act requires training for teams, including on these practices.
  • Remedy: Banning doesn't work; we need to measure, offer an authorized and practical tool, set simple rules and provide training.

What is Shadow AI?

Shadow AI covers three situations, from the most visible to the most discreet:

  1. Consumer assistants used with a personal account : ChatGPT, Claude, Gemini, Le Chat (now Vibe), Perplexity or DeepSeek, opened in a tab of the professional browser or on a phone, to write, translate, summarize or analyze.
  2. AI functions activated without the IT department's knowledge : browser extensions, note-taking assistants that join video conferences, AI functions added by a publisher in already used software (CRM, design tool, recruitment platform).
  3. The tools built by the teams themselves : a ChatGPT project or plugin shared across an entire service, an agent connected to a messaging system, a small application developed using vibe coding with real data.

The common point is not the tool, it is the lack of framework: no one has checked where the data goes, whether it is used to train a model, who can access it, or what happens when the employee leaves the company.

Shadow AI and Shadow IT: what's the difference?

For the past fifteen years, shadow IT has referred to cloud software and services adopted without the IT department's involvement: instant messaging, personal storage, project management tools. Shadow AI is an extension of this, with three key differences that make it more complex.

Shadow ITShadow AI
What comes out of the companyFiles stored elsewhereContent copied into a prompt, sometimes reused to train a model
Speed ​​of adoptionMonths, tool by toolA few days: a free account is all you need, no installation required
DetectionMade possible through the inventory of applications and flowsDifficult: a browser tab, a personal phone, a hidden function in authorized software
Effect on workStorage or exchange changesThe content itself is changing: texts, analyses, and decisions are based on unverified sources.
Specific regulationsNoThe AI Act is in addition to the GDPR and NIS2

Shadow AI in numbers

Studies converge: the phenomenon is massive, it barely recedes where companies offer an official tool, and it is costly when it goes wrong.

FigureWhat it measuresSource
75% and 78%Percentage of knowledge workers who use AI at work, and percentage of these users who bring their own toolsMicrosoft and LinkedIn, Work Trend Index, May 2024, 31,000 people in 31 countries
37%Percentage of employees using AI who have not discussed it with their superiors in large French organizationsDatacraft and Inria, 2025
47%Percentage of generative AI users at work accessing it through personal accounts, compared to 78% a year earlierNetskope, Cloud and Threat Report 2026
223 per monthAverage number of data protection violations related to generative AI per organization, a number that has doubled in one yearNetskope, Cloud and Threat Report 2026
20%Percentage of organizations that have experienced a data breach due to shadow AIIBM, Cost of a Data Breach 2025
670 000 dollarsAverage additional cost of a leak in organizations with high levels of shadow AIIBM, Cost of a Data Breach 2025
65% and 40%The proportion of these leaks that expose customers' personal data, and the proportion that expose intellectual property (53% and 33% on average)IBM, Cost of a Data Breach 2025
34%Percentage of organizations with an AI policy that regularly audit unauthorized usesIBM, Cost of a Data Breach 2025
Over 40%Percentage of organizations that will experience a security or compliance incident related to unauthorized AI tools by 2030Gartner's prediction for November 2025

The decline in personal accounts measured by Netskope, from 78% to 47% in one year, is the most encouraging statistic: it coincides with the massive deployment of official tools such as Copilot Chat, ChatGPT Enterprise, and Claude for Work. Where companies offer a credible alternative, the shadow banking phenomenon recedes. Where they don't, it thrives.

The concrete risks of shadow AI

The leak of confidential data

This is the most immediate risk. A salesperson pastes a sales proposal into a free assistant to rephrase it, a lawyer uploads a draft contract, an engineer copies source code to find a bug. Depending on the plan and settings, this content can be stored by the publisher, reviewed by their teams for moderation purposes, or used to train subsequent models. Free, individual versions of major assistants generally enable training by default; enterprise offerings exclude it. The 2023 incident at a major Korean industrial company, where engineers copied proprietary code and meeting notes into a public assistant, remains the most frequently cited example: it resulted in a temporary ban, followed by the development of an internal tool.

The GDPR and personal data

As soon as a prompt contains data about individuals (a CV, a client file, an annual performance review, a report with names), it constitutes the processing of personal data. With an unvalidated tool, the company cannot guarantee the legal basis, informing individuals, the existence of a data processing agreement, the location of the data, or its deletion. However, it remains responsible for the processing. For information on how to reconcile GDPR obligations with an AI project, see our article. AI and GDPR: Do we really need a second register?

Intellectual property and trade secrets

Trade secrets are protected only if the company takes "reasonable protective measures." Allowing strategic documents to circulate through uncontrolled tools can weaken this protection in court. Conversely, content produced by an unvalidated tool can raise copyright issues: texts, images, or code generated without knowing the underlying model or the applicable terms of use.

The quality of decisions

A less visible risk: analyses, summaries, or recommendations produced by a tool that no one has evaluated become part of real-world decisions, without review or any record. A numerical error in a dashboard, a fabricated legal precedent in a legal brief, or a bias in the screening of applications goes undetected when the involvement of AI is unknown.

Cybersecurity

Browser extensions and agents connected to email or document storage create access points that security measures often overlook. IBM notes that 97% of organizations that experienced an AI-related data breach lacked appropriate access controls for their AI tools. The NIS2 directive, applicable to critical and important entities, mandates the management of these risks: Articles 20 (management responsibility), 21 (risk management measures), and 23 (incident notification) apply to unauthorized tools just as they do to others.

What the AI Act says

The European regulation on AI does not target shadow AI as such, but two of its obligations directly concern it.

Article 4, on AI literacy. Since February 2, 2025, any organization using AI systems must take steps to develop its staff's AI proficiency, taking into account roles and contexts of use. Omnibus Regulation (EU) 2026/1744, in force since July 27, 2026, has made this an obligation of means, without a guaranteed level for each individual. Enforcement has been the responsibility of national authorities since August 2, 2026. An employee who uses AI in their work, even with an unauthorized tool, falls under this requirement: widespread shadow AI signals that these measures have not been taken or have not been effective.

Article 50, on transparency. Since August 2, 2026, highly manipulated and generated texts published to inform the public must be flagged as such. Content produced with an unknown tool is unlikely to receive the correct label.

The most serious risk, however, lies in the high-risk category, postponed to December 2, 2027: a general-purpose assistant used to sort applications or evaluate employees becomes a high-risk system, and the company that repurposed it bears the associated obligations. The complete timeline is detailed in our article. AI Act: What your company needs to do.

Why doesn't banning work?

Blocking ChatGPT at the proxy level seems reassuring on paper. In practice, however, usage shifts to personal phones, other unblocked assistants, or AI functions integrated into authorized software. The ban doesn't eliminate the need; it eliminates visibility: the organization no longer knows what's being used, and employees stop talking about it. The 37% of users who didn't inform their superiors in the Datacraft and Inria study illustrate this silence.

Shadow AI is primarily a useful signal. It shows where AI saves time, which professions are ahead of the curve, and what official tools don't yet offer. An organization that listens to it learns faster than one that tries to ignore it.

How to break free from shadow AI: the six-step plan

  1. Measure without monitoring. A short, anonymous survey (which tools, for which tasks, with which data, and why not the official tool?) reveals more than a network traffic audit, which doesn't examine phones or integrated functions. Cross-reference this with proxy logs and SaaS application inventory, without targeting individuals.
  2. To offer an authorized tool that is at least as practical. This is the measure that pushes back the shadow. A corporate assistant (Copilot Chat, ChatGPT Enterprise, Claude, Gemini, Vibe by Mistral, or an internal tool), with single sign-on, data excluded from training, logging, and a subcontracting agreement. If it's slower or more limited than the free version, employees will go back to it.
  3. Set short and concrete rules. A one- or two-page charter: authorized tools, data that is never included (health data, trade secrets, personally identifiable HR data), what is reviewed before sending, what is flagged, and who to ask questions. AI charter template It provides a foundation to be adapted; our support aims to co-create it with the various professions and to bring it to life. draft a code of conduct.
  4. To train, profession by profession. The rule is not enough if no one knows why it exists or how to achieve the same results with the official tool. Practical training on each team's real-world cases transforms shadow users into ambassadors. It is also the simplest measure to document under Article 4; the AI training plan template helps to structure it.
  5. Open a service point for new tools. A shorter process (one week rather than six months) for a team to request an evaluation of an AI tool or function. Without this streamlined approach, every new need creates unnecessary delays.
  6. To regularize without sanctioning. When unauthorized use is discovered, the risk is assessed, an alternative is proposed, and assistance is provided for migration. Systematic sanctions ensure that future use remains hidden.
Usage observedRisk levelResponse,
Rewrite a public text or come up with ideasLowSwitch to the official tool, without any formalities.
Summarize non-sensitive internal documentsMediumOfficial tool required, reminder of the rules
Processing customer, HR or financial dataHighOfficial tool with data excluded from training, proofreading, DPO information
Health data, trade secrets, decisions about peopleCriticalProhibited outside designated environments; legal classification of use
Agent or application built by a teamVariableCheck-in at the counter: inventory, access rights, designated manager

Three examples of shadow AI in business

The legal department. A lawyer reviews a 40-page contract with a free assistant to identify unusual clauses. The time saved is significant: one hour instead of half a day. However, the contract contains a client's financial details and is sent to a personal account. The right approach isn't to ban assisted review, but to offer it within the official tool, with a rule to verify cited references.

Human resources. A recruitment officer asks an assistant to rank 80 CVs based on job suitability. This practice involves personal data and falls within a domain that the AI Act classifies as high-risk. The response here is a precise framework: no automatic ranking of candidates without qualified tools, human supervision, and notification of candidates.

Communication. A team generates campaign visuals using an image tool found on a social network. Questions arise regarding image usage rights, adherence to brand guidelines, and the required attribution for realistic content. A short list of approved tools, along with their terms of use, addresses most situations.

F.A.Q

What is Shadow AI?

This is the use of artificial intelligence tools for professional purposes without the agreement or control of the organization: consumer assistants on a personal account, browser extensions, AI functions enabled in software, or agents created by the teams themselves.

Shadow AI, shadow IA, AI of the shadows: is it the same thing?

Yes. Shadow AI is the English term, shadow IA is its French equivalent, and IA de l'ombre or IA fantôme are its translations. We also talk about BYOAI ("bring your own AI") when employees bring their own tools.

Is shadow AI illegal?

Not in itself. It becomes unlawful when it leads to a violation of the GDPR, professional secrecy, a confidentiality agreement, or internal regulations, or when a tool is used for a prohibited or high-risk purpose as defined by the AI Act. The company remains responsible to the individuals concerned and the authorities.

Should ChatGPT be blocked in the company?

Blocking alone simply shifts usage to personal phones and other devices, and eliminates all visibility. It only makes sense if accompanied by an official alternative that is at least as practical, clear rules, and training.

How to detect shadow AI?

By combining an anonymous employee survey, analysis of proxy logs and SaaS application inventory, and a review of AI features enabled in existing software, the survey remains essential, as technical tools cannot detect personal phones or integrated functions.

What is the link between shadow AI and Article 4 of the AI Act?

Article 4, in effect since February 2025, mandates measures to develop AI proficiency among those who use it in their work, including with unauthorized tools. Training teams on the risks of shadow AI, providing them with validated tools, and documenting these actions are among the expected measures.

Sources

For further : our article on digital responsibility and the war for talent et our guide to designing responsible generative AI.

Back to top
All NextStart.AI offersEight consulting missions, seven pillars, around twenty formats, eighteen professions, five tools
Around thirty major organizations were supported.Large groups, public agencies and networks: from acculturation to production

Insurance and social protection

Public sector and research

Consumption and retail

The NextStart.AI collectiveConsultants, trainers and industry experts: more than 5 350 trained professionals since 2023
All free resourcesDiagnostics, Finders, prompt library, kits by profession, by tool, and for managing an AI program
The NextStart.AI blogGuides, prices and comparisons of Copilot, Claude, ChatGPT, Gemini and Mistral, and field analyses
Let's talk about your AI projectResponse within 24 business hours
AI agents in businessUnderstanding, choosing the tool, training the teams, deploying
Delegate entire tasksClaude Cowork, Copilot Cowork, ChatGPT Work, Vibe Work