It’s a scenario that plays out across workplaces. In the race to meet a deadline, a marketing manager copies last quarter’s sales data into ChatGPT for analysis. A developer pastes proprietary code into an AI assistant to debug faster. Or a customer rep uploads a transcript of client complaints to generate response templates. Sounds familiar? asks Steve Timothy, pictured, Cyber Security Specialist Director, Ricoh UK.
The threat here isn’t a potential cyber-attack, but the routine, well-intentioned transfer of business information outside the organisation’s governance and security controls. Every prompt can become a potential data leak. This growing challenge of uncontrolled AI use in the workplace has sparked a new, growing trend known as shadow AI. The term refers to the unauthorised, ad hoc use of AI tools within an organisation, often outside IT governance. Unlike shadow IT, this isn’t primarily a technical problem, but an adoption and governance challenge that demands a different approach.
Why shadow AI has become a material security risk
Employees use shadow AI for much the same reason as shadow IT: adopting unsanctioned technology because it’s accessible, fast, and helps them get their work done. The difference is that AI creates a more immediate exposure risk. Rather than simply accessing an unapproved application, users may actively paste, prompt or upload business information into a third-party service, often without knowing where that data is processed, stored, or reused.
The data exposure risks are significant. Confidentiality and privacy may be compromised when personal data, HR records or customer information are entered into public tools. Intellectual property can leave the organisation through source code, product roadmaps and commercial proposals. Regulatory and contractual obligations may be breached where data-processing terms, retention policies and cross-border transfers are unclear. And if a preventable data-handling error becomes public, the reputational impact can be significant. This differs from attackers using AI to automate phishing, reconnaissance, or malware development. That threat also matters, but it calls for threat detection and resilience. Shadow AI is primarily an internal governance issue, best addressed with clear guardrails and secure alternatives rather than blanket bans.
Why banning AI will not solve the problem
Shadow AI is rarely the result of negligence or malicious intent. It has grown because AI is accessible, fast and increasingly embedded in the tools people already use. It is now part of everyday work. And everyday habits will not disappear because of a policy email. A blanket ban may appear decisive, but it can simply push usage underground. When staff lack an approved option that meets their role’s needs, security teams lose visibility while the underlying demand remains.
The more effective approach is to make the secure option the easiest option. A secure approach must provide approved AI tools that fit real workflows, have clear data-handling guidelines and deliver comparable convenience. In other words, give people a better tool than the one they can find for themselves.
From shadow AI to secure adoption
Addressing shadow AI effectively requires a balanced mix of robust governance frameworks, technical safeguards, employee support and ongoing oversight. Here are three strategies to mitigate the risks:
- Start with visibility
You cannot govern what you cannot see. Begin by identifying which AI tools and services employees access, whether through browsers, embedded workplace applications, or personal accounts. Map the data flows and prioritise high-risk use cases, especially those involving customer data, source code, financial information, or sensitive internal documents.
Speak with business functions to understand why people use particular tools and what problem they are trying to solve. This helps distinguish genuinely valuable use cases from avoidable risk. Use the findings to create a clear inventory of approved, tolerated, restricted and prohibited uses.
- Put proportionate guardrails around data
Rules should be clear, practical and tied to data sensitivity. Define the information that employees must never enter into public AI tools, such as confidential customer data, unreleased financial results and proprietary code. Avoid an unrealistic blanket policy. Different roles will have different needs and risk profiles, so create guidance that reflects both the type of data involved and the intended use. Where appropriate, deploy AI-aware data-loss prevention controls that can identify sensitive content, block or redact it in real time, and produce an auditable alert trail.
Establish a consistent assessment process for new AI tools, integrations, and autonomous agents before introducing them into business workflows.
- Make secure adoption sustainable
Controls will only hold if the approved route is usable. Give employees easy access to capable, approved AI tools that support the tasks they actually need to complete. If the secure alternative is slower or less effective than a public tool, workarounds are inevitable. Training should use recognisable examples rather than generic warnings. Staff also need a simple way to ask whether a proposed use case is safe, without navigating a lengthy approval process.
Finally, treat AI governance as an ongoing operating model, not a one-off policy project. Review tools, controls and training as capabilities and risks evolve. Ownership must be cross-functional, bringing together security, IT, privacy, legal, HR and business leadership.
Enable innovation responsibly
The goal is not to eliminate AI, but to prevent its uncontrolled use. Employees will keep seeking faster ways to automate routine tasks, and organisations that ignore that demand risk pushing it beyond their line of sight.
The more constructive approach is to understand current exposure, set clear and proportionate boundaries, and provide secure tools that support real work. Governance must also be treated as an ongoing discipline, evolving alongside new tools, use cases and regulatory expectations. Organisations that take this approach can capture AI’s productivity benefits while ensuring confidential data is not treated as the price of innovation.
