Shadow AI: The Growing AI Governance Risk Nobody Owns

There are several ways we frequently see shadow AI develop in organisations, such as:

• Using personal AI accounts or licences for work-related activities
• Using approved enterprise tools and services, but forgetting to sign in to an approved corporate account
• Entering business information into tools and services that have not been approved
• Different teams adopting their own AI tools and services without a consistent approach across the organisation.

Shadow AI does not necessarily mean employees are deliberately breaking the rules.  In our experience, it more often arises when access to AI outpaces organisational guidance and controls, or when expectations are unclear.  Many AI tools are available free of charge, making it easy for employees to start using them without following procurement processes.

Readers may already be familiar with shadow IT, where employees use technology, software or services that have not been approved or managed by the organisation.  

In our work with organisations, we are increasingly seeing a familiar challenge: AI use is moving faster than the guidance and controls around it.  Shadow AI follows the same principle as shadow IT, although the term is not yet as widely recognised.  As publicly available AI tools and services become more widely used, the challenges around AI governance, visibility and control are becoming harder to ignore.

In short, the concept of ‘shadow’ technology is not new, but its application to AI is relatively recent, and the term ‘shadow AI’ is not yet widely appreciated.

The central risk is simple: your organisation cannot adequately protect information or manage activity it does not know about.  Some examples of the risks associated with shadow AI include:

• Data leakage.  Employees may enter corporate information into AI tools and services that have not been approved.  We strongly recommend using approved enterprise AI tools through a corporate account, with the relevant contractual safeguards and settings checked.  Whether information is retained, used for model training or potentially disclosed depends on the service, account type, terms and configuration.  A corporate account alone does not guarantee protection, and personal accounts may not offer the safeguards your organisation requires.
• Data protection concerns. Personal data may be processed without appropriate consideration of UK General Data Protection Regulation (GDPR) requirements.
• Data residency and processing location. Some organisations must process customer information within specified national borders to meet legal or contractual requirements.  Approved enterprise AI services may offer controls over where information is stored and processed, but organisations need to verify the locations covered by the service and its contractual commitments.  Personal AI accounts may offer fewer controls, potentially resulting in information being processed elsewhere in the world.
• Increased attack surface.  Unmanaged applications, accounts, browser extensions or integrations can introduce additional exposure, without adequate protection.
• Loss of control and oversight.  Your organisation may not know which tools and services are being used, what information is being entered or how outputs are subsequently used.

As stated previously, these risks most often materialise due to a lack of staff understanding or organisational guidance on secure AI usage, not through malicious intent, and the simplest mistakes can have a serious impact.  

Imagine a colleague trying to finish a piece of customer work.  They have not been told which AI tools are approved, or whether AI is suitable for the task, so they enter the customer’s intellectual property into an AI service.  The customer’s contract requires that information to remain within the UK, but the service processes it in a US data centre.  Without understanding the service’s processing and retention arrangements, the colleague has inadvertently breached that requirement.

Approving an AI tool is only part of the process. Staff also need to understand how to access it and which safeguards must remain in place.  If an employee enters work information without signing in with their approved corporate account, the organisation’s intended protections may not apply.  Depending on the service’s terms and settings, that information could be retained or used for model training, creating a risk of unintended disclosure.

Staff need structured training to use AI safely and effectively, with more specialised training for some roles.  For example, a technical team member might use AI-enabled security-testing features within an existing tool to assess the company’s website.  Without training in the tool’s capabilities, configuration and scope restrictions, they could allow testing to extend to other organisations’ websites.  A legitimate activity could then create substantial legal, contractual and reputational consequences for the business.

‍

Our starting point is not simply to block AI.  It is to understand how people are using it, give them an approved way to benefit from it and make responsibilities clear.  The following practical steps help organisations gain visibility of AI use and manage the risks of shadow AI.

1. Establish your AI landscape.  Interview and collect evidence on which AI tools and services are already being used across your organisation.
2. Assign clear ownership.  Give a named role or governance group responsibility for coordinating AI oversight, with input from IT, information security, data protection, procurement and business teams. Make clear who approves AI use, who checks safeguards and who responds when unapproved use is identified.
3. Define approved AI tools and services.  Establish processes to evaluate and approve AI tools and services for use, as well as clearly communicating to employees which tools can be used and for what purposes.
4. Link to project and change management.  Build checkpoints into development and change management processes to identify proposed AI products, integrations and new AI features within existing systems before they are introduced.  Connect these checkpoints to AI approval and procurement processes so that the necessary reviews, approvals and purchases take place before use.
5. Provide protected AI access.  Use approved enterprise AI tools and services that your organisation can control and configure with protections.
6. Control unapproved services.  Where appropriate, restrict access to unapproved AI tools and services.  We often see organisations achieving this using tools they already have, such as web filtering to block access to online AI services that haven’t been approved.
7. Make the need for authentication clear.  Explain how employees should access approved AI services and why they must use the organisation’s approved corporate account rather than a personal account or unauthenticated service. Include these requirements in relevant policies, such as the acceptable use policy, and reinforce them when access is provided and whenever requirements change, as well as during annual policy reviews.
8. Set clear rules around information.  Establish an AI policy that clearly outlines what information employees can and cannot put into AI tools and services.
9. Train your organisation.  Train employees on how to use AI effectively and safely.  The most successful approaches we see integrate AI into established information security training and awareness campaigns, similar to topics such as phishing.
10. Monitor AI usage.  Use network monitoring and data loss prevention tools to help identify unapproved AI services, unusual activity and possible policy violations.  Depending on their configuration and the traffic they can inspect, next-generation firewalls and deep packet inspection may provide additional application-level visibility. These measures support oversight but do not guarantee visibility of every AI interaction.

Our advice is to make safe AI use easier, not simply to tell people what they must not do.  Shadow AI often arises when employees try to work more efficiently without the guidance or tools they need, rather than from deliberate rule-breaking.  In our work helping organisations strengthen their AI risk management and governance, we have found that approved AI services, clear policies and effective training put businesses in a stronger position to benefit from AI while maintaining control over their data, security and compliance obligations.