Shadow AI: From Hidden Use to Managed Reality

Shadow AI is not a discipline problem. It is a visibility problem that prohibition cannot solve.

A 2026 PagerDuty survey of 1,250 office professionals, conducted by Wakefield Research, found that two-thirds (66%) had used an AI tool at work despite believing it was against company policy, and 88% had shared some form of work-related information with a public AI tool such as ChatGPT, Claude or Gemini, including 31% who had shared financial information or confidential company documents and strategies.

This is happening largely out of sight. Bitdefender's 2026 Cybersecurity Assessment of 1,200 global security professionals found that 47% lack full visibility into the AI tools their own employees use, and that leadership is systematically overconfident about the gap: 58% of IT and security managers claimed full AI visibility, but only 46% of the practitioners actually doing the work agreed. IBM's 2025 Cost of a Data Breach Report found that 20% of breached organisations had a shadow AI component, adding an average of $670,000 to the cost of the breach, and that 97% of organisations reporting an AI-related breach lacked proper access controls.

As PagerDuty's own chief technology officer, Tim Armandpour, put it on releasing the survey: "When over 30% of employees are putting confidential company data into public models, 'Shadow AI' becomes a massive enterprise liability... The goal for any executive today should not be to slow down AI adoption, but to redirect that energy into proven platforms that offer governance and automation at scale."

This whitepaper sets out why blanket bans reliably fail, what genuine visibility into everyday AI use actually requires, and how organisations can build governed access that manages the risk without smothering the experimentation driving real productivity gains.

What employees are actually pasting into public AI tools is more sensitive than most policies assume.

PagerDuty's data breaks down exactly what is being shared with tools such as ChatGPT, Claude and Gemini: 43% of respondents have shared emails or other correspondence, 40% have shared meeting notes or summaries, 34% have input customer data, and 31% have shared financial information or confidential company documents and strategies. None of this requires a sophisticated attack. It requires an employee trying to work faster and a public tool with no visibility wrapped around it.

IBM's 2025 Cost of a Data Breach Report quantifies what that exposure costs once it goes wrong. Organisations with a shadow AI component to their breach paid $4.63 million on average, against $3.96 million for breaches with no shadow AI involvement, a premium of roughly $670,000. The nature of the exposure was also more severe: 65% of shadow-AI-linked breaches involved customer personally identifiable information, against a 53% baseline across all breaches studied, and 40% involved intellectual property, against a 33% baseline. Among organisations that reported an AI-related breach specifically, 97% said they lacked proper access controls.

As the Bitdefender report examined by TechTarget put it: shadow AI is the successor to shadow IT, but "it's harder to detect, and the potential for data leakage is orders of magnitude greater." A forgotten SaaS subscription is a discovery problem. A public chatbot fed customer records is a data-protection incident the moment it happens, whether or not anyone in security ever finds out.

A policy that only says no pushes the exact behaviour it is trying to stop further out of view.

Chase Cunningham, a zero-trust security expert and chief strategy officer at Demo-Force, is blunt about the limits of prohibition: "A policy that simply says, 'Do not use generative AI,' is not a strategy. It frequently pushes usage further underground, where security teams have even less visibility." His conclusion follows directly: "Organizations can't govern what they can't see."

PagerDuty's data shows why enforcement alone does not change the underlying behaviour. Among employees who used AI tools they believed were not permitted, 53% received informal feedback or guidance to stop, and 48% faced formal consequences such as a warning or disciplinary action, and usage remained widespread regardless. The same survey found real costs on the other side of the ledger: 77% of office professionals believe their company's AI restrictions are limiting their professional growth, and 75% say they would consider leaving for a role that offered better AI skills development, a figure that rises to 80% at companies with $1 billion or more in revenue.

A ban, in other words, does not remove the risk. It removes the organisation's visibility into the risk, while simultaneously creating a retention problem among the employees most inclined to use AI productively in the first place.

The goal is not to stop people using AI. It is to stop them using it invisibly, indiscriminately and with more access than the task requires.

Chase Cunningham's own framing of the alternative to a ban is the clearest available summary of what a managed approach actually looks like in practice: "Don't try to stop employees from using AI. Stop them from using AI invisibly, indiscriminately and with more access than the task requires." Each clause maps to a concrete design choice.

Invisibly is addressed by the discovery layer described in the previous section, paired with a genuinely fast approval path for new tools, matched to how quickly employees actually adopt them in practice, given that 89% first encounter a tool personally, long before any request could reach a review committee. Indiscriminately is addressed by a hard data-classification boundary that travels with the tool rather than the policy: certain categories of information, customer data, financial detail, source code, unreleased strategy, should never enter a public model, regardless of which tool an employee prefers. With more access than the task requires is addressed by scoping what any sanctioned AI integration can actually see and do, rather than granting broad access by default because it is operationally simpler.

None of this works without removing the penalty for disclosure. PagerDuty's data shows that formal consequences did not stop unauthorised use; they simply added a reason to hide it. A channel through which employees can surface the tools they are already using, without punishment, converts information the organisation currently cannot see into information it can actually govern, which is the entire point.

Shadow AI is not going away. Whether it stays hidden is a choice organisations are still making.

Two-thirds of employees are already using AI tools their organisation has not sanctioned. Nearly half of security professionals cannot see it happening, and the organisations that experience a breach with a shadow AI component pay materially more for it. None of that changes by writing a stricter policy, the evidence is consistent that prohibition drives the behaviour further out of view rather than ending it.

What changes the picture is treating shadow AI as a visibility problem to be solved rather than a compliance breach to be punished: discover what is actually in use, draw a firm line around the data that must never leave the organisation regardless of tool, make the sanctioned path faster than the unsanctioned one, and remove the penalty that currently gives employees every reason to keep their AI use hidden.

As PagerDuty's own leadership put it, the goal was never to slow AI adoption down. It is to redirect the energy already driving that adoption into tools the organisation can actually see, turning shadow AI from a hidden liability into a managed, and genuinely productive, reality.