What is shadow AI?

Shadow AI is the use of artificial intelligence tools by employees without the approval, knowledge or oversight of the organisation. It is the AI version of shadow IT: people reach for whatever helps them get the work done, and the tools arrive faster than any policy can keep up. The intent is usually good. The exposure it creates is not.

It shows up in three main ways:

• Public chatbots. Staff pasting work into personal ChatGPT, Gemini or Claude accounts to draft emails, summarise documents or write code.
• Browser extensions. Unvetted AI plug-ins that can read and process whatever is on screen, granted broad permissions in a couple of clicks.
• Hidden features. AI quietly switched on inside everyday SaaS products that nobody assessed for security before it went live.

Searches for shadow AI have risen sharply over the past year, which tells you the problem is outrunning most organisations’ awareness of it. The tools are in the building whether leadership has noticed or not.

Why shadow AI is a real risk, not a tech scare

Because these tools sit outside any governance, they create exposures that only become visible after something has gone wrong:

• Data leakage. Sensitive information, customer records, source code or financial data pasted into a public model can be retained and used to train future versions. Once it is out, it does not come back.
• Regulatory breach. Processing personal or customer data through an unsanctioned tool can breach UK GDPR and the EU AI Act, and the accountability sits with the organisation, not the employee.
• Intellectual property loss. Proprietary methods and strategy shared with an external model can leak into the public domain.
• Unverified decisions. Acting on unchecked AI output, including confident hallucinations, puts errors straight into business-critical work.
• No evidence. When an auditor, customer or regulator asks what AI you use and how you control it, “we are not sure” is its own failure.

Why banning AI does not work

The reflex is to ban the tools. It is the worst available option. Bans do not stop the behaviour, they drive it underground: people move from a chatbot you might have monitored to a personal phone you cannot see at all. You lose the visibility that was the only thing keeping the risk manageable, and you forfeit the productivity your competitors are busy capturing.

The organisations handling this well are doing the opposite. They treat it as a governance problem, not a security crackdown: make the safe path the easy path. Give people sanctioned tools that are good enough to use, draw clear lines around what data can go where, and watch what is actually happening rather than pretending it is not.

How to govern shadow AI

Governing shadow AI is a leadership job before it is a technical one. The response that works has four parts:

• Get visibility. Find out which AI tools are actually in use, through endpoint and SaaS discovery, before you write a single rule. You cannot govern what you cannot see.
• Set clear boundaries. A short, usable AI policy that says plainly what data can and cannot go into which tools, in language people will actually follow.
• Provide sanctioned alternatives. Enterprise-grade AI with proper data controls, so the approved option is the convenient one and the shadow option loses its appeal.
• Put it on the board agenda. AI use, and the risk attached to it, reported to the board the way any other material risk is, with someone accountable for it.

This is the heart of AI governance: not a one-off policy document, but an owned, reported, improving programme. For a mid-market or PE-backed business it is the same discipline a CISO as a service or fractional CIO and CISO brings to any other enterprise risk, alongside broader cyber security consulting.

Why this is a board accountability

AI is now a governance and compliance matter, and the cost of getting data protection wrong is a board-level number. The global average cost of a data breach reached USD 4.44 million in 2025, according to IBM’s Cost of a Data Breach Report, and breaches that involve poorly governed AI tend to sit at the worse end of that range. UK GDPR already applies to any personal data processed through an AI tool, and the EU AI Act adds obligations for organisations operating in or selling into the EU. The board cannot delegate that accountability to whoever happened to install the plug-in.

Why Starkhorn

Starkhorn is led by Daniel J. Jacobs, who has spent over 20 years in technology and security, 15 of them in leadership roles, including Interim Group Technology Director at VetPartners, the BC Partners-backed veterinary group, and CIO and CISO at Jardine Motors Group. He is the author of The Strategy Bridge and holds PRINCE2, ITIL Foundation and full membership of the Institute of Interim Management.

What that means for shadow AI: a leader who has governed enterprise risk and run technology at scale, who treats AI as something to harness safely rather than ban, and who can get you from “we have no idea what is being used” to a governed, board-ready position. One senior practitioner, not a tool you still have to run yourself.

Frequently asked questions