Shadow AI: Identifying, Evaluating, and Managing Shadow AI in the Enterprise
An employee uses a chatbot to shorten a proposal, while a developer uses one to explain third-party code—both via private accounts. For the IT department, these processes do not exist. This is precisely what Shadow AI refers to: AI tools utilized for work purposes despite not being provided by the employer. While a ban might seem like the obvious response, it is often insufficient. Effectively managing Shadow AI requires evaluating each use case individually to decide whether to terminate, redirect, or adopt it. This article demonstrates how organizations can gain the necessary oversight to do so.
Defining Shadow AI and Its Prevalence
The German Economic Institute (IW) describes Shadow AI as the professional use of AI applications not provided by the employer, such as private chatbot accounts. This is not automatically clandestine or prohibited; however, it means that no one within the organization knows which data is flowing where.
The prevalence is significant. In the IW employee survey from Spring 2026, involving nearly 5,000 participants, 29.1% of professional AI users reported using such tools. Among those under 30, the figure was 38.4%, and in companies with fewer than 50 employees, it reached 33.5%.
A similar picture emerges from the corporate perspective. In a representative Bitkom survey of 604 companies with 20 or more employees, the use of private AI tools is widespread in 8% of companies, while isolated cases exist in 17%. A further 17% are not certain but assume it is occurring. Consequently, around four out of ten companies expect Shadow AI to be present in their operations.
Why Employees Turn to Their Own Tools
One might assume that Shadow AI primarily arises where the employer neglects the topic. However, IW data suggests the opposite. Among those who credit their employer with a strong commitment to building AI skills, 44.5% use non-provided tools. Without perceived support, the figure is 21.3%. The IW interprets this as a provisioning gap: employees learn what AI is capable of but cannot find the appropriate tools within the company. The authors themselves point out that their data proves a correlation, but not a direction of causality.
Bitkom’s figures align with this. Only 26% of surveyed companies provide their employees with access to generative AI; for companies with 20 to 99 employees, this figure is 23%. Barely a quarter have established rules for the use of AI.
This leads to an uncomfortable conclusion for management. Shadow AI is frequently not a sign of a lack of discipline, but rather of a need that the company has not yet met. Those who merely block these tools lose the very information regarding which tasks the workforce intends to perform with AI.
The Risks Associated with Shadow AI
The core problem is the loss of control over data. Most freely accessible AI services operate as cloud offerings from the provider. The Data Protection Conference (DSK) refers to such applications as open systems: inputs leave the protected environment, may be reused depending on the provider, and are frequently transmitted to third countries. Furthermore, any processing of personal data requires a legal basis under the GDPR. In the case of a privately created account, this assessment has never taken place.
The speed at which an input becomes personal is often underestimated. Removing names is insufficient if the context reveals who is involved. A request to draft a warning letter for the only warehouse worker at a specific branch identifies a specific person, even if their name is not mentioned. In addition to personal data, trade secrets are at stake: calculations, draft contracts, and proposals that are processed daily in sales and projects.
A second level of risk arises when AI tools do not merely respond but act. Many assistants can be connected to mailboxes, calendars, or file storage with just a few clicks. In such cases, the issue is no longer just about what data reaches the outside world, but also what a tool can modify or send in an employee’s name.
Identifying Shadow AI Without Sowing Mistrust
Shadow AI cannot be fully measured, as private smartphones and accounts lie outside the company’s own infrastructure. Nevertheless, two approaches complement each other well.
The more important approach is the open inquiry. A short, anonymous survey clarifies which tools are used for which tasks and what data is involved. This only works if it is established beforehand that honest answers will not result in consequences. Those who threaten sanctions will receive silence rather than an accurate assessment.
The second approach involves technical data that is generated anyway, such as proxy or DNS logs showing access to known AI services. These indicate which services are being accessed from the corporate network, but not what is being entered there. Caution is advised regarding evaluations at the individual level: the introduction of technical facilities designed to monitor the behavior or performance of employees is subject to co-determination under Section 87 (1) No. 6 of the Works Constitution Act (BetrVG), provided a works council exists. Therefore, the works council and data protection officer should be involved from the outset. For management purposes, aggregated figures are sufficient in any case.
Categorizing Every Use Case: Terminate, Redirect, or Adopt
The assessment provides a list of specific uses. Rather than allowing or forbidding them across the board, it is worthwhile to make a decision per use case. Two questions carry the bulk of the evaluation: What data is flowing in? And does the tool run as an open system at the provider or in a shielded, contractually regulated environment?
This results in three paths:
- Terminate: Personnel data, customer data, or trade secrets are moving into a private account with an open service. There is no room for maneuver here; use must stop immediately.
- Redirect: The task is meaningful, but an already approved tool performs it just as well. In this case, it is sufficient to make the alternative known and ensure the transition is easy.
- Adopt: The use case brings recognizable benefits, and there is no official offering for it yet. It is then a candidate for a regular rollout, including an examination of the legal basis, a contract with the provider, and, if necessary, a data protection impact assessment (DPIA) according to Art. 35 GDPR. The DSK assumes that such an impact assessment will frequently be required when using AI applications.
The advantage of this classification is that a diffuse risk situation is transformed into a prioritized task list. Cases in the first category are shut down immediately, while the others are addressed in an order based on their utility.
Setting the Framework: Rules, Tools, and Competence
To prevent new Shadow AI from emerging in the first place, three building blocks are required.
A written usage policy. This defines which AI tools are permitted for which purposes and which data must never be entered. The urgency of this is highlighted by the DSK’s assessment that unauthorized and uncontrolled AI use is already a reality in many companies and authorities without such rules. A single page featuring typical internal cases is more helpful in daily operations than many pages of general principles.
Official tools with corporate accounts. The DSK recommends providing employees with devices and accounts for professional AI use so they do not have to rely on private access. This is the most effective lever, as a rule without an alternative will be bypassed. The IW reaches the same conclusion: pure bans primarily affect engaged employees.
Competence. Since February 2, 2025, Article 4 of the AI Act has obliged providers and operators of AI systems to do their best to ensure sufficient AI literacy among their staff. The Federal Network Agency points out that a lack of AI competence can be viewed as a breach of the duty of care, especially if damage results, and recommends documenting the measures taken. Training that explains which data is permitted in which tool fulfills this obligation while simultaneously reducing the risk from Shadow AI. Further obligations arising from the AI Act are summarized in the article Corporate Obligations under the AI Act.
The First Step
Shadow AI does not disappear through a ban, nor can it be ignored. A manageable roadmap is sufficient for getting started:
- An anonymous short survey on tools, tasks, and data used, with a clear commitment that openness will have no negative consequences.
- The categorization of each use case: terminate, redirect, or adopt.
- A concise usage policy with internal examples.
- An official tool with corporate accounts for the most common tasks.
- Documented AI literacy training.
Repeating the survey after one year will show whether the official offering meets the demand. If the share of private tools rises again, it is not a discipline problem, but an indication of the next gap. The article Data Protection 2026 explains how this topic fits into the remaining data protection requirements surrounding AI.
