A Guide to Managing Shadow AI: Making Invisible AI Risks Visible and Manageable

History is repeating itself: a few years ago, the issue of ‘shadow IT’ was one of the biggest challenges facing IT departments. Cloud services or software were being used without the approval of the relevant corporate bodies – usually not out of malice, but because the official processes were too slow and the tools made work easier. Data leaks, security breaches and a loss of efficiency were often the result.

Today, we are seeing the same phenomenon in the context of artificial intelligence. We refer to ‘Shadow AI’ when AI tools are used that have neither been officially authorised nor are subject to IT governance controls. The users’ aim is greater flexibility, higher productivity and a better user experience. However, whilst the individual benefits increase, the risk to the organisation as a whole grows in the background.

The hidden dangers: Why shadow AI is a strategic ris  

Shadow AI is not merely a technological issue, but a matter of business risk. Anyone who is unaware of its current use within the organisation is flying blind. This gives rise to four key risk clusters:

• Data protection and information security: Confidential or sensitive company or customer data is leaking unnoticed into public AI models. In the worst-case scenario, AI systems could derive competitive product advantages from this data and pass this knowledge on to third parties.

• Compliance and Legal Affairs: The use of unauthorised AI tools may breach regulatory requirements. Examples include AI-driven candidate selection or the monitoring of staff. Such practices may contravene the EU AI Act (the European AI Regulation) and thus result in significant legal and financial consequences.
  
  
• Wrong decisions caused by ‘hallucinations’: Many users are too quick to trust AI results without understanding the technology’s limitations. Those who accept unverified results risk making potentially incorrect strategic or operational decisions that could cause significant damage.
  
  
• Damage to reputation: Incorrect AI results that become public undermine the trust of partners and customers. Such a loss of trust often leads to long-term losses in turnover and makes it more difficult to attract new customers.

The current volatility of the market is exacerbating this problem: AI tools are emerging and evolving faster than companies can adapt their policies.

Of course, the business risks arising from the use of shadow AI can be assessed and consciously accepted as residual risk. However, if known legal breaches – such as those of the GDPR – are knowingly tolerated, this may be construed as wilful misconduct. Depending on the incident, this can significantly increase the severity of sanctions and ultimately force organisations to take action. Similarly, the reputational consequences of improper AI use are incalculable. In practice, inaction is not an option.

According to a 2025 IBM study, the additional costs of security incidents were, on average, USD 670,000 higher for organisations with high levels of shadow AI usage than for those with low or no shadow AI usage.

Furthermore, Gartner predicts that by 2030, more than 40 per cent of organisations will be dealing with security or compliance incidents related to shadow AI.

Such figures show that the issue of shadow AI must not be underestimated. Its detection can be broken down into several steps.

Before you start your search: Laying the foundations for the visibility of the AI tools you use

Before you start actively looking for Shadow AI, you should establish a sensible framework.

1. Clarifying terminology
Start by clearly distinguishing the concept of AI from conventional process automation. Avoid ‘AI washing’ – that is, labelling tools as AI when they merely use simple algorithms. A shared basic understanding forms the foundation for any further action.

2. Clarifying responsibilities
Once the definition has been established, the next step is to assign responsibilities and roles within AI governance. We recommend using a RACI chart (a matrix for defining responsibilities) for this purpose. An example of the process for regularly identifying AI usage: the Head of IT is the owner and bears the duty of oversight (Accountable), whilst the Information Security Officer (ISO), together with the IT department, is responsible for implementation (Responsible); the Data Protection Officer and the Works Council are involved in an advisory capacity (Consulted) during the collection and evaluation of data.

3. Using guidelines as a search tool
Once responsibilities have been clarified, it is advisable to publish some initial, simple AI guidelines. These often serve as a search tool at first: staff begin to ask, ‘Is this application permitted under our rules?’. This allows you to identify shadow AI before you even need to actively search for it.

4. Turning those affected into active participants
To ensure that Step 3 works: make it clear that you will not impose sanctions and encourage people to ask questions. Make it clear that the aim is to gain insight and develop a joint strategy, not to criticise or cause personal harm. This will increase the likelihood of honest answers.

Targeted search for and identification of shadow AI

Once the foundations are in place, the next step is identification. When doing so, focus on the areas posing the greatest risks.

Technical measures
Make use of the tools already available within your IT infrastructure. Microsoft Defender XDR and Purview already offer extensive capabilities for detecting the use of publicly available generative AI. SIEM (Security Information and Event Management) systems, using logs from proxies or firewalls, also help to identify unusual data flows.

The methodological approach
Start on the defensive. Focus initially on the ‘what’ (which tools are being used) and treat the ‘who’ as a lower priority. Anonymise the results so as not to judge individuals, but simply to analyse usage and the associated risks.

Be prepared to accept ‘blind spots’ – such as unmanaged mobile devices. It is more effective to start quickly with an initial, incomplete analysis than to hold up the process by striving for perfection.

Create a data-driven framework for the use of AI

1. Share: The tool is secure and will be officially integrated.

2. Prohibit: The risk is too high and there is no acceptable alternative.

3. Replace: This is often the most effective approach. There is a genuine need behind almost every instance of Shadow AI use. Simply banning it results in a loss of visibility and drives users back underground. 

And last but not least: creating a secure framework for the use of AI within the organisation

Effective governance does not consist of bans, but of providing safe, efficient and user-friendly alternatives.

Recommendations for a secure AI environment:

• Create whitelists: Define permitted tools and corporate accounts that meet data protection and compliance requirements.

• Check contracts: Ensure that prompts in your business accounts are not used for the provider’s model training.
    
• Data classification: Clearly define which data is sensitive and introduce a clear labelling system (labels or tags).
  
  
• Ongoing training: Keep your staff regularly informed about the possibilities and limitations of AI in order to improve media literacy within the organisation.

Check these points at regular intervals to ensure they are still applicable, up to date and relevant, e.g. every 3 months.

Conclusion: Finding the path to strategic AI governance through Shadow AI

Shadow AI is not an adversary, but a guide to the real needs of your workforce. The challenge for businesses is to strike a balance between the necessary security and the pace of innovation. By fostering transparency and offering secure alternatives, organisations not only reduce risks but also actively boost productivity.