AI Law, Risk & Governance

Shadow AI: How to Detect and Govern Unsanctioned Use

Marin Niehues6 January 202511 min read

Fountain pen on documents

Shadow AI is the use of unapproved AI tools in daily work, and it is detected through visibility and governed through an approved offering. A large share of employees who use AI work with their own tools and private accounts that neither procurement nor IT has ever seen. This article shows how to detect such use, assess it and bring it into governed operation.

What is shadow AI and why does it arise?

Shadow AI is the work-related use of AI tools without the company's knowledge or approval. It is a special case of shadow IT, meaning any software that business units deploy without IT. The difference lies in the type of risk: with shadow IT, data sit in an outside system; with shadow AI, data are entered into a model, processed and reused as a new output. This article provides an overview and does not constitute legal advice.

Shadow AI arises where work pressure grows faster than the approved offering. A clerk who drafts a customer letter with a public chatbot acts rationally: the tool is at hand, it saves time and the official tool catalog offers nothing comparable. The main cause lies in what the organization delivers: where the approved offering is missing or worse than a private account, employees turn to the private account.

In practice, three constellations recur:

  • Private accounts used for work tasks, such as summarizing documents or analyzing spreadsheets.
  • Subscriptions bought by business units themselves, charged to expenses or departmental budgets.
  • AI features in existing software that a vendor has switched on with an update.

How many employees use AI without their employer's knowledge?

In around one in three companies in Germany, employees use private AI accounts, either known or at least suspected. This is the finding of a representative 2024 survey of around 600 companies with 20 or more employees. In 17 percent of companies private use is known, and a further 17 percent assume it. Only 37 percent rule it out with certainty, down from 51 percent a year earlier. At the same time, only one company in seven had rules for the use of generative AI.

The figures reflect the companies' own view. Actual use is likely higher, because private accounts on private devices remain invisible to management. For steering, this leads to a clear assumption: a company that knows of no use has not yet measured it.

What are the risks of shadow AI?

Shadow AI chiefly endangers the protection of confidential data and the company's ability to account for its use of AI. Confidential information leaves the organization's control, personal data flow out without a data processing agreement, and trade secrets lose their legal protection where no reasonable confidentiality measures are in place. Add to this unchecked outputs reaching customers and a use of AI the company cannot account for.

The EU AI Act adds a further point. It requires companies that deploy AI to ensure sufficient AI literacy among their staff. Training that covers only approved tools does not reach use in the shadows. Without an overview of actual use, the obligation is therefore met only in part.

Detection: Four Lenses

No single data stream shows all use, so detection works with several sources at once. Looking along four lenses has proved effective:

  • Procurement: expense claims and credit card statements reveal subscriptions to AI providers.
  • Network: proxy and DNS logs show which AI services departments access and to what extent.
  • Software inventory: browser extensions and plug-ins, plus newly enabled AI features in existing applications.
  • Survey: a short, anonymous survey names purposes that no log can reveal.

The survey delivers the most valuable part, namely the purpose. A network log shows that a service is used. Only the conversation shows which task lies behind it and how much time it saves. The result of the four lenses is an inventory per tool, with purpose and department as well as the type of data processed.

Analyzing usage data touches data protection and co-determination. Coordination with the data protection officer and the works council therefore belongs at the start.

Assessment: Risk by Data Class and Impact

Every finding needs a classification before any reaction follows. Two questions suffice: which data leave the organization, and what impact does the output have on customers, decisions or contracts? This produces a matrix with four fields.

The matrix classifies each finding in a few minutes and prevents blanket reactions. It also sets the order in which the approved offering must close gaps. Where the legal assessment of a case is open, it belongs with specialist lawyers.

Governance: The Offering Comes before the Ban

A ban without an alternative shifts use to private devices, where it becomes invisible. A fixed sequence works: first the offering, then the rule, finally the control.

The offering consists of one or two approved assistants with an enterprise license, sign-in through the central account and contractually regulated data processing. It must work better in daily use than the private account, or it will remain unused. The rule links data classes to tools and sits in a clear policy. Content and rollout of such a policy are covered in AI policy for employees. Control consists of regularly repeating the four lenses and tracking metrics.

A named owner decides on new tools, and does so within a few days. An approval procedure that takes weeks creates the next wave of shadow AI. Assigning this role is part of how the organization governs AI overall. Building committees and policies is described in Building AI governance, and the place of governance in overall AI management in the article AI management: what executive leadership must steer.

What may employers do, and what do employees risk?

An employer may prohibit the use of unapproved AI tools for work tasks by instruction or restrict use to specific tools. The rule takes effect once it is clearly communicated, ideally in the AI policy. In Germany, anyone who wants to check compliance through technical controls needs the works council's involvement where one exists.

For employees, the consequences of a breach depend on severity and intent. For a first breach of a known rule, a formal warning is the usual measure. Where someone deliberately enters confidential customer data or trade secrets into a public service, more serious employment consequences come into play. For damage, employees in Germany are liable on a graded basis according to their degree of fault, and as a rule not at all for slight negligence.

For steering, sequence matters. Sanctions presuppose a clear rule and a usable offering. Without both, a warning hits people who have solved a problem for the organization. The assessment of individual cases belongs with HR and legal advisers.

Shadow AI as a Delivery Task: The Role of Niehues Consulting

Niehues Consulting supports organizations in the transition from individual use to governed operation. In AI management, the inventory, the requirements and the order of applications the approved offering must cover are established. Prototypes prove the benefit in the real process before a license is rolled out widely. Implementation sets up processes and responsibilities and enables business units to adopt the offering. Delivery management ensures that the approval route and responsibilities work reliably in day-to-day business.

The approach follows the Niehues Delivery Model with the phases Understand, Build and Implement. Three review questions matter most for shadow AI. In Understand: what is the baseline, and how do we measure success? In Implement: who is accountable for operation from day one? Also in Implement: are users enabled and processes adapted? Marin Niehues brings experience from projects at VC-funded scale-ups and DAX 40 groups in 20+ countries. The entry point is a free 30-minute initial consultation via the contact page.

Frequently asked questions

What is shadow AI?

Shadow AI refers to AI tools that employees use for work tasks without management or IT knowing about or approving them. Typical cases are private accounts with public language models and self-purchased subscriptions, plus AI features in existing software switched on unnoticed.

What is the difference between shadow IT and shadow AI?

Shadow AI is a special case of shadow IT in which data are entered into an AI model and processed into new outputs. With shadow IT, data sit in an unapproved system. Shadow AI adds unchecked outputs that flow into customer letters or decisions.

How do you detect shadow AI?

A company detects shadow AI through four sources: procurement data, network logs, the software inventory and an anonymous survey. Together they produce an inventory per tool with purpose and data class. Analysis of usage data is agreed in advance with the data protection officer and the works council.

How many employees use AI without their employer knowing?

In around one in three German companies, employees use private AI accounts or management suspects so, according to a representative 2024 company survey. Only 37 percent of companies rule out private use with certainty. Actual prevalence is likely higher, because private devices remain invisible.

Can an employer ban the use of ChatGPT?

An employer may prohibit work-related use of ChatGPT or other AI services by instruction or restrict it to approved accounts. The ban only becomes effective with a usable approved offering. In Germany, technical controls of compliance require works council involvement where one exists.

What happens to employees who use AI without permission?

For a breach of a known AI rule, a formal warning is the usual first measure, and deliberate disclosure of confidential data can lead to more serious consequences. Employees are liable for damage on a graded basis according to their degree of fault. Individual cases are assessed by HR and legal advisers.

Does banning private AI tools make sense?

A ban only works together with an approved offering that covers employees' tasks. Without an alternative, use migrates to private devices and becomes invisible. For confidential data with high impact, the ban remains in place.

What are the risks of shadow AI?

Shadow AI endangers confidential data and trade secrets, which lose their legal protection without reasonable confidentiality measures. Personal data flow out without a data processing agreement, and unchecked outputs reach customers. In addition, the required AI literacy of staff can only partly be demonstrated without an overview of actual use.

Let’s talk about your initiative.

In 30 minutes we clarify where your organization stands and which step will pay off most now.

Request a consultation

30 minutes · free · no obligation