The company states that it does not use artificial intelligence. Meanwhile, however, sales staff summarise contracts with a chatbot, marketing generates campaigns through a personal account, HR transcribes interviews with a browser extension and a project manager connects an AI assistant to shared documents.
None of these tools appears in official systems, yet AI has already entered the processes. And that is a problem, because nothing governs it.
This invisible adoption is known as Shadow AI: the professional use of artificial intelligence tools, models and features that the organisation has not authorised, recorded or brought under its own rules.
Such a practice may stem from an individual initiative, a subscription bought with a company card or an AI feature activated within software that has already been approved. The point is the same: the company is using systems whose presence, data processed, suppliers and responsibilities it does not know.
Let us look at what Shadow AI is, and the risks it entails.
What is Shadow AI?
Shadow AI is the evolution of Shadow IT: the set of technologies adopted outside the control of the responsible departments. With artificial intelligence the boundary becomes harder to observe, because adoption does not necessarily require installing software. A website, a free account, a plugin or a built-in feature is enough to create potential conflicts with the AI Act.
As early as 2024, the Microsoft and LinkedIn Work Trend Index found that 78% of AI users were bringing tools of their own choosing to work. Shadow AI therefore stems from the gap between very rapid operational demand and a still slow organisational response.
The phenomenon includes public chatbots, writing assistants, transcription systems, image generation platforms, coding copilots and agents connected to CRMs, CMSs, archives or email inboxes. The tool may be reliable and its use perfectly legitimate. The risk stems from the lack of a conscious decision about what it may receive, produce and modify.
Shadow AI and the risks for the company
Beyond the simple question of whether a model is authorised, Shadow AI brings serious risks within the corporate ecosystem.
The first concerns data. An employee may enter personal information, draft contracts, financial data or confidential documents without knowing the retention periods, data location, terms of use or whether the model may be trained on them. Even when the provider offers adequate guarantees, a personal account may not have the same protections as the enterprise version.
The second risk concerns quality. Models can produce content that is plausible but inaccurate, incomplete or unsupported by sources. Our guide to AI hallucinations shows why checking outputs must be part of the process rather than an isolated act.
Next come intellectual property, security and reputation. Generated content may incorporate unverified elements, an extension may access more information than necessary, and an agent connected to internal systems may take actions without adequate limits. OWASP lists among the main risks of applications based on language models the disclosure of sensitive information, prompt injection, improper output handling and excessive autonomy granted to the system.
Finally, there is a brand risk. When different departments use different tools and prompts, the tone of voice, sources and approval criteria change from one piece of content to the next. AI speeds up production, but it can multiply inconsistencies that the organisation can no longer trace.
How to recognise the signs of Shadow AI
| Signal | What it may indicate | First check |
| Much faster output with no declared process | Use of chatbots or personal automations | Ask which tools support the workflow |
| AI subscriptions in expense claims | Unrecorded adoption by individual teams | Check the account, plan and data processed |
| Text or images of inconsistent quality and tone | Differing prompts, models and revisions | Reconstruct generation and approval |
| Plugins and extensions with broad access | Unassessed integrations | Check permissions and connections |
| Difficulty tracing the source of data | Output used without traceability | Define verification and recording obligations |
Why banning everything does not work
An outright ban may reduce some uses, but it does not remove the need that generated them. When people perceive a concrete benefit and find no authorised alternative, use tends to become less visible. The organisation thus loses the ability to train, monitor and correct.
The most effective response combines clear boundaries with workable tools. Some data must remain off-limits, some use cases require approval, and others can be handled with authorised solutions and proportionate controls. The rule must start from the activity and the risk, not merely from the name of the software.
How to regain control
The first step is to map actual use cases. Simply asking “Which AI tools do you use?” yields partial answers. It is more useful to reconstruct where AI comes into play: writing, analysis, screening, translation, programming, customer support, decisions or automation. The same tool may be acceptable for an internal draft yet unsuitable for handling sensitive data.
The inventory gives rise to a classification. For each use, record the department, purpose, data entered, outputs produced, supplier, integrations, owner, human review and authorisation status. At that point the company can define an AI Policy aligned with its processes, select approved tools and establish how to handle exceptions and incidents.
Training completes the system. Article 4 of the European AI Act requires dedicated measures for AI literacy among staff and those who use systems on the organisation’s behalf. In 2026, then, the issue is the practical ability to recognise limits, prohibited data, output risks and the conditions for human intervention.
Monitoring must be continuous. New tools, built-in features and agents can change the perimeter within a few weeks. The NIST AI Risk Management Framework proposes risk management across the entire life cycle, which also helps prevent the inventory from becoming a snapshot that quickly goes out of date.
From Shadow AI to AI Governance
Shadow AI signals that adoption has already taken place ahead of governance. Treating it merely as a disciplinary problem prevents you from seeing the demand for innovation it contains. The organisation’s task is to turn scattered experimentation into uses that are visible, assessable and compatible with its own objectives.
Corporate governance provides the underlying principle: roles, information and responsibilities must be clear before a decision has consequences. Applied to AI, this means knowing who can authorise a tool, who controls the data, who checks the output and who steps in when the system fails.
Control means creating a perimeter within which teams can experiment without data, risks and decisions becoming invisible. When this is missing, the company keeps using AI but loses the ability to govern its real value.
New connections (FAQ)
How do you find out which AI tools employees actually use?
What is needed is a census built around use cases, supported by departmental interviews, subscription checks, analysis of browser extensions and a review of integrations. A punitive questionnaire tends to drive the phenomenon underground. A process-oriented mapping, by contrast, helps bring the real tools, data and needs to the surface.
Does an AI Policy eliminate Shadow AI?
It reduces it when it offers understandable rules, authorised alternatives and fast procedures for assessing new use cases. A generic document, or one based solely on prohibitions, can push adoption outside official channels. The policy must be updated alongside the inventory and training.
Where can a company start if it has not yet mapped its use of AI?
The starting point is a light-touch audit of the most exposed processes, followed by an inventory of tools and a risk classification. Bliss supports management and departments in building this perimeter, connecting AI Governance, data rules, output control and brand consistency without blocking the uses that create value.
BLISS®
© 2026. ALL RIGHTS RESERVED