Shadow AI: Governing What You Cannot See

Executive Summary

Shadow AI is not primarily a policy problem. It is a visibility problem created by a growing gap between how people work and how organisations govern technology.

Every organisation that has ever tried to prohibit a useful technology has eventually learned the same lesson: employees do not stop needing the outcome simply because access to the tool is restricted. They find another route.

The same pattern that produced shadow IT is now repeating itself with artificial intelligence, but faster, across more workflows and with potentially more sensitive information involved.

The challenge is also becoming harder to define. Shadow AI is no longer limited to employees opening an unapproved chatbot. Increasingly, AI capabilities are appearing inside software that organisations already use and have already approved: productivity suites, CRM platforms, meeting tools, development environments and collaboration systems.

That means the governance gap is widening from both directions. Employees are adopting unsanctioned tools outside formal approval processes, while sanctioned software is gaining AI capability without always triggering a new governance decision.

This report examines what shadow AI actually is, why it persists, why embedded AI makes the problem structurally harder, what happens when sensitive information enters systems an organisation cannot see, and the governance principles most likely to remain useful regardless of which tools or vendors dominate next.

02 · The Scale of the Blind Spot

The problem is no longer whether shadow AI exists. It is whether organisations can see enough of it to govern it.

Industry research consistently points in the same direction: AI use across organisations is more widespread than formal governance systems can reliably observe.

Employees are using public AI tools, personal accounts, browser extensions and AI assistants outside approved procurement routes. At the same time, authorised SaaS platforms are introducing their own AI features, creating another layer of activity that may not appear in traditional software inventories.

That gap between exposure and visibility matters because unobserved use is difficult to assess, difficult to investigate and difficult to control.

When an incident involves an approved enterprise system, security teams generally know which platform was involved, what data it processes, which contractual controls apply and where to begin the investigation.

When an incident involves an unsanctioned AI tool, those basic facts may be unknown. The organisation may not know which model received the information, what data was submitted, whether the content was retained or whether a contractual relationship exists with the provider at all.

The central governance risk is therefore not simply that employees may use the wrong AI tool. It is that organisations may be unable to answer basic questions about where AI is being used, what information is entering those systems and who is accountable when something goes wrong.

03 · AI That Arrives Without Asking

The harder governance problem is increasingly not the tool employees choose to adopt, but the AI capability that appears inside software already approved.

Much of enterprise AI policy was initially designed around standalone applications. Should employees use ChatGPT? Should a particular generative AI platform be approved? Which public tools are prohibited?

Those questions still matter, but they address only part of the exposure.

AI capabilities are increasingly being embedded directly into productivity platforms, CRM systems, meeting software, collaboration tools and other SaaS products that organisations already use.

That changes the governance problem because the normal moment of decision disappears.

An employee who opens an unapproved chatbot has made a visible choice to use a new tool. An employee whose existing CRM begins generating AI summaries, or whose meeting platform begins processing calls through a new AI feature, may have made no comparable decision at all.

The software was already approved. The vendor was already contracted. The capability changed underneath the existing relationship.

This is why shadow AI cannot be governed solely through procurement controls for new tools. A growing share of the exposure is arriving through software that is already inside the organisation.

05 · What It Costs When It Goes Wrong

The defining risk of shadow AI is not simply that sensitive information may leave the organisation. It is that the organisation may have no reliable record that it happened.

One of the earliest widely cited examples emerged in 2023, when employees at a major electronics manufacturer reportedly entered proprietary source code and internal meeting material into a public generative AI service while trying to debug and summarise their work.

The incident became notable because it was visible. Many equivalent interactions may never be detected at all.

That is the critical difference between shadow AI and a conventional data incident involving a known corporate system.

With an approved platform, investigators can usually identify the system involved, retrieve logs, understand contractual data-handling terms and reconstruct at least part of the event.

With an unsanctioned AI tool, the only meaningful record of the interaction may sit on infrastructure the organisation does not control and with a provider it has no direct contractual relationship with.

The resulting exposure extends beyond personal data. Source code, commercially sensitive documents, intellectual property, unreleased financial information, customer records and privileged material may all be entered into systems that were never assessed for those data categories.

The cost of shadow AI therefore includes the incident itself and the organisation's reduced ability to investigate, evidence and contain what happened afterwards.

07 · The Principle That Will Not Expire

You cannot govern what you refuse to see, and you cannot close a capability gap with policy alone.

Strip away the individual tools, vendors and statistics and the underlying pattern is remarkably stable.

Employees adopt unsanctioned technology when it helps them accomplish something the sanctioned environment makes difficult.

Organisations lose control when their governance model focuses on prohibition before understanding the real behaviour taking place.

And visibility declines further when AI capability enters through software that has already passed the normal procurement gate.

The durable response is therefore not to predict which tool should be banned next.

It is to build continuous discovery, clear data rules, strong vendor governance and a sanctioned AI path that employees actually want to use. That combination is far more likely to survive the next wave of AI capability than any list of approved and prohibited products ever will.

References & Sources

Research and evidence informing this report.

This report draws on publicly available 2026 industry research, enterprise security analysis and AI governance guidance. Statistics vary by study population and methodology and should be verified against their originating source before being cited independently.

Statistics in this report are drawn from publicly available 2026 industry research and should be treated as directional indicators rather than directly comparable measures unless stated otherwise. Specific figures, vendors and product capabilities will change quickly; the durable findings of this report are the visibility gap, the growth of embedded AI and the need to align discovery, data classification, vendor governance and sanctioned AI access.