top of page
HCL Review
nexus institue transparent.png
Catalyst Center Transparent.png
Adaptive Lab Transparent.png
Foundations of Leadership
DEIB
Purpose-Driven Workplace
Creating a Dynamic Organizational Culture
Strategic People Management Capstone

Out of the Shadows: Why Employees Are Using Unsanctioned AI at Work, and How Leaders Can Turn Hidden Experimentation Into Governed Capability

2 hours ago
30 min read

Download the article here:


Listen to a review of this article:


Abstract: Employees are adopting generative artificial intelligence faster than their organizations can approve, secure, or support it, producing widespread "shadow AI": the use of unsanctioned AI tools for work. This article synthesizes recent survey evidence from 2024 to 2026, breach-cost research, and established scholarship on shadow IT, workarounds, proactive behavior, technology acceptance, and psychological safety to explain why unsanctioned use persists and what it costs. It argues that shadow AI is less a compliance failure than a signal of unmet demand, weak enablement, and low trust, and that blanket prohibition tends to drive usage underground rather than eliminate it. Drawing on documented examples from technology, banking, professional services, law, retail, biotechnology, health care, and government, the article sets out five evidence-based responses: providing secure sanctioned alternatives, designing risk-tiered and usable governance, investing in role-based AI literacy, creating psychological safety for disclosure, and harvesting grassroots experimentation into shared capability. It closes with three pillars for building durable, governed AI capability.

 

A few years ago, a manager who discovered that a team member was quietly using an unapproved software tool would probably have treated it as a minor policy breach. Today the same discovery tells a much bigger story. Generative AI tools are free or cheap, available in any browser, and capable of drafting, summarizing, coding, and analyzing in seconds. Employees have found them, learned them at home, and brought them to work, often without asking anyone. The organizations they work for are, in many cases, still drafting the policy.

 

The numbers are striking even allowing for the limits of vendor-sponsored surveys. In a June 2026 study of 1,250 office professionals at large organizations in the United States, United Kingdom, Australia, and Japan, two-thirds (66%) said they had used AI tools at work while believing company policy did not allow it, and 88% said they had shared work-related information with public AI tools (PagerDuty, 2026). Research from Ricoh reported in October 2026 found that 45% of workers use non-approved tools such as shadow AI at least monthly, even as 81% of decision-makers admitted they are investing in new technology faster than they are supporting worker adoption (Hale, 2026). Two years earlier, Microsoft and LinkedIn's global survey of 31,000 knowledge workers found that 78% of AI users were bringing their own AI tools to work (Microsoft & LinkedIn, 2024).

 

I have spent much of my career advising organizations on how people actually adopt new ways of working, as opposed to how strategy documents say they will. The pattern with generative AI is familiar, but the speed and stakes are new. Employees are not waiting. When the sanctioned path is slow, clumsy, or absent, they build their own. And because the tools sit outside the organization's view, so do the risks: confidential documents pasted into consumer chatbots, customer data processed on servers no one has vetted, and outputs used in client work without anyone checking their accuracy.

 

The instinctive leadership response is prohibition. Several prominent firms took exactly that route in 2023, and some had good reason to. But bans alone rarely solve the underlying problem. They remove the visible behavior while leaving the demand intact, and the demand finds other channels: personal phones, home laptops, or tools embedded in other software. Worse, a culture of prohibition teaches employees to hide what they are doing, which deprives leaders of precisely the information they need to manage risk and spread good practice.

 

This article makes the case for a different stance. Shadow AI is best understood as a diagnostic signal. It tells leaders where work is painful, where approved tools are falling short, where training is missing, and where trust between employees and the organization has frayed. The goal is not to tolerate ungoverned use, which carries real costs, but to bring experimentation into the light and convert it into governed capability. That shift requires secure alternatives that people actually want to use, governance that is proportionate and comprehensible, serious investment in AI literacy, a climate in which people can admit what they have been doing without fear, and mechanisms that turn individual workarounds into shared organizational assets.

 

The sections that follow define the key terms, review what we know about the prevalence and drivers of shadow AI, assess its consequences for organizations and individuals, and then set out five evidence-based responses illustrated with documented organizational examples. The final sections describe how to build long-term capability and offer practical takeaways for leaders who want to move from policing AI use to leading it.

 

The Shadow AI Landscape

 

Defining Shadow AI, BYOAI, and Sanctioned Experimentation

Shadow AI refers to the use of artificial intelligence tools, models, or features for work purposes without the knowledge, approval, or oversight of the organization's IT, security, or governance functions. The term borrows from the older concept of shadow IT, which information systems researchers define as hardware, software, or services that employees use for work without formal approval from the IT function (Haag & Eckhardt, 2017; Silic & Back, 2014). Shadow IT research has long noted that such systems are typically adopted with good intentions, to get work done faster or better, yet sit outside the organization's security controls and architectural standards. Silic and Back (2014), analyzing one organization's shadow software, found that communication tools, content applications, and small utilities dominated, and that most were adopted because they were convenient rather than malicious.

 

Bring your own AI (BYOAI) is the more neutral, practitioner term popularized by Microsoft and LinkedIn's 2024 Work Trend Index. It describes employees using AI tools their employer did not provide (Microsoft & LinkedIn, 2024). BYOAI overlaps heavily with shadow AI but is not identical. An employee might bring a personal subscription to a tool that the organization actually permits, which is BYOAI but not shadow AI. Conversely, an employee might use an AI feature quietly switched on inside an approved application, which is shadow AI without any personal tool at all.

 

Two further distinctions matter for leaders. The first is between unsanctioned and prohibited use. Much shadow AI occurs not because a tool has been explicitly banned but because no one has said anything at all. When policy is silent or unclear, employees fill the gap with their own judgment. The second distinction is between the tool and the data. Using a public chatbot to brainstorm headlines is very different from pasting a confidential merger document into it. Governance frameworks that treat all AI use as equally risky tend to be ignored because employees can see that they are disproportionate.

 

Against these, I use the term sanctioned experimentation to describe the alternative this article advocates: an approach in which the organization provides approved, secure AI tools, sets clear and risk-proportionate rules about data and use cases, and actively invites employees to explore, share, and improve how AI is used in their work. Sanctioned experimentation is not a free-for-all. It is a deliberate design choice that channels employee initiative into visible, governable forms.

 

A useful theoretical lens here is Alter's (2014) theory of workarounds. In Alter's account, a workaround is a purposeful change that people make to an existing work system so they can get around obstacles, constraints, or breakdowns that stand between them and the results they are trying to achieve. Crucially, his theory treats workarounds as neither inherently good nor bad. They can be harmful, but they can also reveal flaws in the formal system and become the seed of larger, planned improvements. Shadow AI fits this description almost perfectly.

 

Prevalence, Drivers, and Distribution

Estimating the true prevalence of shadow AI is difficult, partly because definitions vary and partly because people who are breaking rules do not always admit it in surveys. Nevertheless, the convergence of findings from independent sources is notable. They are not directly comparable, because each survey asks a different question of a different population, but together they show that unsanctioned AI use is a mainstream phenomenon rather than a fringe one.

 

At the high end, 78% of AI users in Microsoft and LinkedIn's 2024 survey brought their own AI tools to work, with the share rising to 85% among Generation Z and remaining high at 73% among the oldest workers (Microsoft & LinkedIn, 2024). This was not a story about young employees alone. PagerDuty's 2026 survey of office professionals at organizations with more than $500 million in revenue found that 66% had used AI while believing it was not permitted (PagerDuty, 2026). Ricoh's research found 45% of workers using non-approved tools at least monthly (Hale, 2026). A 1Password survey of more than 5,000 knowledge workers in six countries, reported in October 2025, found a lower figure of 27% who said they had worked with AI tools their company had not authorized, though 37% admitted they do not always follow company AI policies (Poireault, 2025). Lenovo's Work Reborn research, which surveyed 6,000 employees at large organizations in 12 countries between December 2025 and January 2026, found that between one-fifth and one-third of employees use AI outside IT governance (Cueto, 2026).

 

Health care, a sector with strict data protections, shows lower but still meaningful figures. A December 2025 survey of 518 hospital and health system providers and administrators, commissioned by Wolters Kluwer Health, found that 17% admitted using unauthorized AI tools at work (Wolters Kluwer Health, 2026). Among providers who had done so, the most common primary reason was a faster workflow (45%), followed by better functionality than approved tools (24%).

 

Why do employees do it? The evidence points to four overlapping drivers.

 

The first is simple performance expectancy: people believe AI makes them better at their jobs. The Unified Theory of Acceptance and Use of Technology identifies performance expectancy, effort expectancy, social influence, and facilitating conditions as core predictors of technology use (Venkatesh et al., 2003). Generative AI scores unusually high on the first two. It is easy to try, and the payoff is often immediate. In the 1Password data, 45% of those using unauthorized AI cited convenience and 43% said they feel more productive (Poireault, 2025). Lenovo's respondents largely agreed that AI improves productivity (71%), work quality (72%), and creativity (73%) (Cueto, 2026). These beliefs are not baseless. Controlled field studies show substantial gains on suitable tasks, a point I return to below.

 

The second driver is a provision gap. Employees often lack sanctioned tools that are as capable or convenient as consumer alternatives. In Lenovo's survey, 22% of AI users said their employer provides no AI tools for work and 31% said it provides no AI training (Cueto, 2026). Deloitte's survey of 3,235 senior leaders across 24 countries, fielded in August and September 2025, found that worker access to sanctioned AI tools had grown by roughly half in one year, from under 40% to around 60% of workers (Deloitte AI Institute, 2026). That is real progress, but it still implies that a large minority of the workforce has no approved option. Ricoh's finding that 53% of workers say AI tools add friction or extra work suggests that even where tools exist, they do not always fit the way people work (Hale, 2026).

 

The third driver is a strategy and communication gap. Gallup's tracking of U.S. employees shows AI use rising steadily, with 52% using AI at work at least a few times a year by May 2026, up from 21% in May 2023. Yet only about a quarter of employees (25% to 26% across 2026 readings) say their organization has communicated a clear plan or strategy for integrating AI (Gallup, 2026). In the 2024 Work Trend Index, 60% of leaders worried that their organization's leadership lacked a plan and vision for implementing AI (Microsoft & LinkedIn, 2024). When the official direction is unclear, employees improvise.

 

The fourth driver is proactivity and career self-interest. Organizational research has long shown that employees with high role breadth self-efficacy and a flexible role orientation are more likely to take initiative to improve work methods without being asked (Parker et al., 2006). Many shadow AI users fit this profile. PagerDuty's respondents provide a vivid illustration: 77% believed their company's AI restrictions limit their professional growth or career mobility, and 75% said they would likely look for a new job offering better AI skills development (PagerDuty, 2026). Seventy-two percent believed they know AI better than their own technology teams. Whether or not that belief is accurate, it signals that many employees see AI fluency as a career asset they cannot afford to wait for.

 

Distribution matters too. Gallup's data show daily AI use is far more common in technology (42%), finance (27%), and professional services (22%) than in most other industries (Gallup, 2026). Shadow AI is also not confined to junior staff. In PagerDuty's survey, senior leaders were more likely than mid-level managers and below to believe they understand AI better than their technology teams (77% versus 66%), and 81% of respondents believed that leadership follows different AI rules from everyone else (PagerDuty, 2026). Perhaps the most telling finding is that 89% of workplace AI users first encountered the tools in their personal lives (PagerDuty, 2026). Consumer adoption is setting the pace, and organizations are trying to catch up.

 

Organizational and Individual Consequences

 

Organizational Performance Impacts

The most concrete organizational risk of shadow AI is data exposure. When employees paste content into a consumer AI service, the organization loses control over where that data is stored, who can access it, and whether it may be used to train future models, depending on the service's terms. PagerDuty's respondents reported sharing emails or correspondence (43%), meeting notes (40%), customer data (34%), and financial information or confidential documents and strategies (31%) with public AI tools (PagerDuty, 2026).

 

The 2025 Cost of a Data Breach study, conducted by the Ponemon Institute and analyzed by IBM across 600 breached organizations, was the first in the series to examine AI security and governance. It found that about one in five organizations reported a breach involving shadow AI, and that organizations with high levels of shadow AI faced an average of $670,000 more in breach costs than those with low levels or none (IBM, 2025). Shadow AI incidents were also more likely to compromise sensitive data. Customer personally identifiable information was compromised in 65% of shadow AI incidents compared with 53% across all breaches, and intellectual property in 40% compared with 33% (IBM, 2025).

 

The governance picture is sobering. IBM (2025) reported that 63% of breached organizations either had no AI governance policy or were still developing one. Of those that did have policies, only 34% regularly audited for unsanctioned AI. Among organizations that experienced breaches of AI models or applications, 97% lacked proper AI access controls. These are correlational findings from organizations that had already suffered breaches, so they cannot tell us that shadow AI causes breaches in a simple way. But they do suggest that ungoverned AI use widens the attack surface and makes incidents more damaging when they occur.

 

Beyond security, shadow AI creates quality and reliability risks. Generative AI performs impressively on some tasks and poorly on others that look superficially similar. Dell'Acqua et al. (2023) captured this unevenness in a field experiment with 758 Boston Consulting Group consultants. For tasks inside what they called the jagged technological frontier, consultants using GPT-4 completed 12.2% more tasks, worked 25.1% faster, and produced results rated more than 40% higher in quality. For a task deliberately selected to fall outside the frontier, however, consultants using AI were 19 percentage points less likely to reach a correct answer than those working without it. When employees use AI in the shadows, no one is helping them learn where that frontier lies for their work, and no one is checking whether outputs that reach clients, patients, or regulators are accurate.

 

There is also an opportunity cost that rarely shows up in risk registers. Brynjolfsson et al. (2025) studied the staggered introduction of a generative AI assistant among 5,179 customer support agents and found that it increased issues resolved per hour by about 14% on average and by 34% for novice and lower-skilled workers. Part of the mechanism appeared to be the diffusion of the practices of top performers to less experienced colleagues through the tool. Shadow AI forfeits this benefit. When individuals discover effective uses privately and keep them private, the organization captures little of the learning. Ricoh's finding that 91% of decision-makers believe their AI investments are paying off, while 53% of workers say AI tools create friction, suggests that leaders may also be misreading where value is actually being created (Hale, 2026).

 

Finally, shadow AI fragments the technology estate. Shadow IT research has documented how unsanctioned systems create integration problems, duplicate spending, and compliance gaps (Haag & Eckhardt, 2017). With AI, the problem is compounded because the tools are proliferating rapidly, often embedded in other products, and frequently updated by vendors in ways the organization does not control.

 

Individual Wellbeing and Stakeholder Impacts

The individual consequences of shadow AI are less visible but no less important. The most corrosive is concealment. Microsoft and LinkedIn (2024) found that 52% of AI users were reluctant to admit using AI for their most important tasks, and 53% worried that using AI on important tasks made them look replaceable. Employees in this position carry a double burden. They are doing extra work to learn and apply new tools while also managing the anxiety of being found out, either for breaking a rule or for appearing less essential.

 

Sanctions add to the strain. Among PagerDuty's respondents who had used AI despite believing it was not allowed, 53% had received informal feedback to stop and 48% had faced formal consequences such as a warning or discipline (PagerDuty, 2026). Discipline may sometimes be warranted, particularly where confidential data has been mishandled. But when large numbers of well-intentioned employees are punished for trying to do their jobs better, the message they hear is that initiative is dangerous. Research on employee voice and silence shows that people withhold information and suggestions when they believe speaking up is futile or personally risky (Morrison, 2014). Shadow AI policies built only on enforcement are likely to amplify exactly this silence.

 

Perceptions of fairness matter as well. When 81% of employees believe leaders operate under different AI rules (PagerDuty, 2026), governance loses legitimacy. Rules that appear to apply only to the rank and file are more likely to be circumvented, and the resentment can spill over into broader distrust.

 

There are also wellbeing upsides that organizations risk losing. Lenovo's respondents said they would enjoy their jobs more if they could focus on business-relevant work instead of routine administrative tasks (79%) (Cueto, 2026). In the UK government's large trial of Microsoft 365 Copilot, 63% of surveyed users reported increased job satisfaction (Government Digital Service, 2025). Employees who turn to shadow AI are often seeking relief from tedious work. Driving that use underground does not remove the tedium; it simply makes the relief riskier.

 

External stakeholders bear consequences too. Customers whose data is pasted into public tools have not consented to that use. Patients may be affected if unvetted AI influences clinical documentation. Clients may receive work products that contain AI errors no one has checked. Regulators are increasingly attentive. Under Article 4 of the European Union's Artificial Intelligence Act, providers and deployers of AI systems must take measures to ensure, to the best of their ability, a sufficient level of AI literacy among staff and others who operate or use AI systems on their behalf, an obligation that has applied since 2 February 2025 (European Parliament and Council of the European Union, 2024). Commentators note that national supervisory enforcement of the Act's provisions begins in August 2026 and that weak AI training may be treated as an aggravating factor in other breaches (Travers Smith, 2025). An organization that does not know which AI tools its employees are using will struggle to show that those employees are AI literate.

 

Evidence-Based Organizational Responses


If shadow AI is a symptom of unmet demand, weak enablement, and low trust, the remedy lies in addressing those causes rather than only suppressing the symptom. Table 1 summarizes five complementary responses, the evidence behind each, key practices, and documented organizational examples. No single response is sufficient; they work best as a system.

 

Table 1: Evidence-based organizational responses to shadow AI

Response

Evidence Base

Key Practices

Illustrative Examples

Provide secure, sanctioned alternatives

Shadow IT and technology acceptance research; survey evidence on provision gaps

Enterprise-grade tools with data protections; fast, opt-in access; parity with consumer tools

JPMorgan Chase LLM Suite; PwC ChatGPT Enterprise; Walmart My Assistant

Design risk-tiered, usable governance

NIST AI RMF; ISO/IEC 42001; IBM breach data on governance gaps

Data classification rules; use-case tiers; evaluation before deployment; human review

Morgan Stanley evals; Allen & Overy lawyer review of Harvey; Kaiser Permanente consent and review

Invest in role-based AI literacy

EU AI Act Article 4; jagged frontier evidence; training gaps in surveys

Segmented curricula; prompt and verification skills; data-handling training

JPMorgan Chase AI Made Easy; UK government Copilot trial

Create psychological safety and amnesty

Psychological safety and employee voice research

Disclosure amnesty; leader role-modeling; consistent rules for all levels

Moderna champions and office hours; lessons from 2023 bans

Harvest grassroots experimentation

Workaround theory; diffusion of best practice in field studies

Champion networks; shared libraries; pathways from pilot to platform

Moderna custom GPTs; JPMorgan Chase prompt libraries; A&O Harvey beta

 

Replace Prohibition With Secure, Sanctioned Alternatives

The core insight from shadow IT research is that unsanctioned systems persist because they meet needs that official systems do not (Haag & Eckhardt, 2017; Silic & Back, 2014). Technology acceptance research adds that use follows perceived usefulness and ease (Venkatesh et al., 2003). Put together, these imply that the most reliable way to reduce shadow AI is to offer something better through official channels: tools that are at least as capable and convenient as consumer alternatives, with the data protections the organization needs. Survey evidence supports the logic. Where employees lack approved tools, as 22% of Lenovo's AI users reported (Cueto, 2026), they have little choice but to look elsewhere.

 

Effective approaches:

 

  • Enterprise parity: Provide sanctioned tools whose capabilities and user experience come close to the consumer tools employees already know, so the approved path is also the easy path.

  • Contractual data protection: Choose enterprise agreements that prevent the use of company data for model training and that support data residency, retention, and access requirements.

  • Speed of access: Make getting a license a matter of days, not months, and use opt-in rollouts that let enthusiastic early adopters move first.

  • Integration with work: Connect sanctioned tools to internal knowledge and workflows, so they do things public tools cannot.

  • Retire bans deliberately: When a restriction was a stopgap, announce its replacement clearly and explain what is now permitted.

 

Samsung Electronics banned employees from using ChatGPT and other generative AI chatbots in May 2023 after an engineer accidentally uploaded sensitive internal source code to ChatGPT (Ray, 2023). The company's concern was that data shared with such services is stored on external servers, is hard to retrieve or delete, and could be exposed to other users. That same month, Apple reportedly restricted internal use of ChatGPT and GitHub Copilot over concerns that confidential data could reach developers who train models on user data, and reporting at the time noted that JPMorgan Chase, Verizon, and Walmart had also restricted staff access to ChatGPT (Mehta, 2023). These were understandable first responses to a technology that had arrived suddenly and without enterprise safeguards.

 

What is instructive is what some of these organizations did next. JPMorgan Chase rolled out LLM Suite, its generative AI platform built on third-party large language models, to employees in 2024. According to chief analytics officer Derek Waldron, the bank deliberately avoided mandating use, instead onboarding employees in phases as they requested access, and took time to address risk and control issues before launch. More than 100,000 employees were onboarded within months, nearly 250,000 now have access, and roughly half of employees use generative AI tools daily (McKinsey & Company, 2025). Walmart, for its part, announced in August 2023 that it would give about 50,000 non-store employees a generative AI tool called My Assistant, built into its employee app and combining Walmart data with third-party language models, to help summarize documents and create content (Fried, 2023).

 

Professional services firms moved in the same direction. In May 2024, PwC agreed to provide ChatGPT Enterprise to about 100,000 employees in the United States, United Kingdom, and Middle East, becoming OpenAI's largest enterprise customer at the time and its first reseller (Lunden, 2024). The pattern across these cases is a shift from blocking consumer tools to licensing enterprise-grade alternatives with stronger data protections, on the assumption that demand will not disappear.

 

A common pitfall is to treat the sanctioned tool as a box to tick. If the approved tool is noticeably weaker, harder to reach, or wrapped in so many restrictions that it cannot do useful work, employees will quietly keep using their personal subscriptions. Ricoh's finding that 53% of workers experience AI tools as adding friction (Hale, 2026) is a warning that provision without usability simply relocates the problem.

 

Design Risk-Tiered, Usable Governance

Governance remains essential. IBM's (2025) finding that 63% of breached organizations lacked AI governance policies, and that only 34% of those with policies audited for unsanctioned AI, shows how far many organizations have to go. Established frameworks provide a foundation. The NIST AI Risk Management Framework organizes AI risk management around four functions (govern, map, measure, and manage) and emphasizes that risk management should be proportionate to context and impact (National Institute of Standards and Technology, 2023). ISO/IEC 42001:2023 specifies requirements for an AI management system, giving organizations a certifiable structure for policies, roles, risk assessment, and continual improvement (ISO/IEC, 2023). The practical challenge is translating these frameworks into rules employees can understand and follow in the moment of use.

 

Effective approaches:

 

  • Classify data, not just tools: Give employees a simple rule set about which categories of information may go into which tools, using the organization's existing data classification scheme.

  • Tier use cases by risk: Allow low-risk uses (drafting internal emails, brainstorming) freely in sanctioned tools, require review for medium-risk uses, and reserve formal approval for high-risk uses affecting customers, patients, employees, or legal obligations.

  • Evaluate before scaling: Test AI applications against representative tasks before deployment and keep testing afterward.

  • Keep humans accountable: Specify where human review is mandatory and who is accountable for final outputs.

  • Make governance visible and fast: Publish an approved-tools list and offer a quick intake route for new tool requests, with a target turnaround time.

  • Monitor without surveillance theater: Use network and software discovery tools to understand which AI services are in use, and treat the findings as demand data rather than only as violations.

 

Morgan Stanley offers a clear example of evaluation-led governance. In building its AI @ Morgan Stanley Assistant, a GPT-4-based tool that helps financial advisors find information in the firm's knowledge base, the firm tested every AI use case before deployment, had advisors and prompt engineers grade outputs for accuracy and coherence, and ran daily regression tests using a suite of sample questions to catch weaknesses and support compliance (OpenAI, n.d.-a). According to OpenAI's account, more than 98% of advisor teams now actively use the Assistant. Its meeting-summary tool, Debrief, produces draft notes and follow-ups that advisors review and adjust before finalizing.

 

In law, Allen & Overy announced in February 2023 that it would deploy Harvey, a generative AI platform for legal work, to more than 3,500 lawyers across 43 offices after a beta trial in which about 3,500 lawyers submitted around 40,000 queries in day-to-day client work. The firm was explicit that outputs required careful review by an A&O lawyer (Allen & Overy, 2023). In health care, Kaiser Permanente announced in August 2024 that it was rolling out Abridge's ambient AI documentation tool across 40 hospitals and more than 600 medical offices, after a year of testing in one market. The tool requires patient consent, and clinicians review AI-drafted notes before they enter the medical record (Landi, 2024). Each of these organizations chose a sanctioned platform, set clear boundaries, and built human review into the workflow.

 

A common pitfall is to write a single, long AI policy that treats every use as high risk. Employees who read it, if they read it at all, quickly conclude that compliance would make AI useless and go back to their own tools. Governance that cannot be summarized on a page, and applied in a few seconds, is unlikely to change behavior.

 

Invest in Role-Based AI Literacy

AI literacy is now both a performance issue and, in the European Union, a legal one. The jagged frontier evidence shows why. Dell'Acqua et al. (2023) found that AI improved performance on suitable tasks but degraded it on an unsuitable one, which means the critical skill is knowing when and how to rely on AI and when not to. Yet training lags usage. Only 39% of AI users in the 2024 Work Trend Index had received AI training from their company (Microsoft & LinkedIn, 2024). In Lenovo's 2026 data, 31% of AI users reported no employer training, and among those who did receive training, 51% said it was not regular or ongoing and 42% said it was not effective (Cueto, 2026). Article 4 of the EU AI Act requires deployers to take measures to ensure a sufficient level of AI literacy among staff, with the European Commission's guidance indicating that measures should reflect people's roles, existing knowledge, and the context of use (European Parliament and Council of the European Union, 2024; Travers Smith, 2025).

 

Effective approaches:

 

  • Segment by role: Design different learning paths for general staff, managers, technical specialists, and executives rather than one generic course.

  • Teach verification, not just prompting: Train people to check outputs, spot plausible errors, and recognize tasks that fall outside the frontier.

  • Make data handling concrete: Use realistic examples of what can and cannot be shared, tied to the organization's own data categories.

  • Learn in the flow of work: Supplement formal courses with short videos, office hours, prompt libraries, and peer coaching.

  • Record and refresh: Keep records of training to demonstrate compliance and update content as tools and risks change.

 

JPMorgan Chase paired its LLM Suite rollout with a training program called AI Made Easy, which tens of thousands of employees have completed. Waldron describes a progression from understanding what large language models can and cannot do, to effective prompting, to advanced techniques such as having a model switch between maker and checker roles. The bank tailors training to general staff, technologists, data scientists, and executives, and supplements it with peer learning through prompt libraries and prompt-of-the-week emails (McKinsey & Company, 2025).

 

The UK government's cross-government Microsoft 365 Copilot experiment, run by the Government Digital Service from October to December 2024 with 20,000 licenses across 12 organizations, offers public sector evidence. Participants reported saving an average of 26 minutes per day, though about 17% noticed no clear time savings, and the report cautions that these were self-reported estimates. It described training as essential to realizing benefits, noted a learning curve in prompting effectively, found a link between users' AI familiarity and confidence and the time they saved, and stressed the need for human oversight, particularly for decisions requiring judgment (Government Digital Service, 2025).

 

A common pitfall is to treat AI literacy as a one-off compliance module completed once and forgotten. Tools change every few months, and so do their failure modes. Literacy programs that are not refreshed quickly become irrelevant, and employees who find the official training out of date will turn to online forums and personal experimentation instead.

 

Create Psychological Safety and Amnesty for Disclosure

None of the previous responses works if employees are afraid to tell the organization what they are actually doing. Edmondson (1999) defined psychological safety as a shared belief that the team is safe for interpersonal risk-taking and showed that it predicts learning behavior, including seeking feedback, discussing errors, and experimenting. Morrison's (2014) review of employee voice and silence shows that people speak up when they believe it will make a difference and will not harm them. Shadow AI, by definition, is behavior that employees have chosen not to disclose. Bringing it into the open requires deliberately lowering the cost of disclosure.

 

Effective approaches:

 

  • Declare an amnesty: Invite employees to share which AI tools they have been using and for what, with an explicit commitment that honest disclosure of past use, outside of serious misconduct, will not be punished.

  • Ask about needs, not just tools: Use disclosure conversations to learn which tasks employees are trying to improve and what approved tools are missing.

  • Role-model openly: Have senior leaders talk about how they use AI, including mistakes, and follow the same rules as everyone else.

  • Reward sharing: Recognize employees who surface risks, document good practices, or help colleagues learn.

  • Reserve sanctions for genuine harm: Distinguish clearly between deliberate misuse of sensitive data and good-faith experimentation.

 

Moderna's approach to generative AI adoption shows how organizations can make experimentation visible and social rather than hidden. According to OpenAI's account of the partnership, Moderna's change program combined individual training with collective initiatives, including an AI prompt contest that identified 100 power users who became internal Generative AI Champions, local office hours in every business line and geography, and an internal AI forum with about 2,000 weekly active participants. The program had visible engagement from the chief executive and executive committee (OpenAI, n.d.-b). By turning AI use into something people were recognized for, rather than something they had to hide, the company created strong incentives for disclosure.

 

The early bans of 2023 illustrate the other side of the ledger. Restrictions such as Samsung's were adopted for sound security reasons (Ray, 2023), but in isolation, prohibition tends to push demand toward personal devices and accounts, as the persistent survey evidence of unsanctioned use suggests. PagerDuty's 2026 data, in which nearly half of those who used AI despite believing it was not allowed had faced formal consequences, show how enforcement-heavy approaches can coexist with very high levels of continued use (PagerDuty, 2026).

 

A common pitfall is announcing an amnesty while leaving punitive policies untouched, or allowing leaders to operate under different rules. When 81% of employees believe leadership follows different AI rules (PagerDuty, 2026), any promise of safety will be met with skepticism unless leaders visibly hold themselves to the same standard.

 

Harvest Grassroots Experimentation Into Shared Capability

The final response treats employees' private discoveries as raw material for organizational learning. Alter's (2014) theory of workarounds explicitly notes that workarounds can lead to larger planned changes in work systems, and shadow IT researchers have argued that user-driven systems can be sources of innovation if organizations create pathways to absorb them (Haag & Eckhardt, 2017). The field evidence on AI points the same way: Brynjolfsson et al. (2025) found that the largest gains accrued to less experienced workers, consistent with AI helping to spread the tacit knowledge of top performers. That diffusion happens only if good practice is captured and shared.

 

Effective approaches:

 

  • Build champion networks: Identify skilled users in each function and give them time and recognition for helping colleagues.

  • Create shared libraries: Maintain searchable collections of prompts, custom assistants, and worked examples that employees can reuse and improve.

  • Run structured pilots: Give teams a sanctioned sandbox, a time limit, and clear success criteria for testing new uses.

  • Define a path to scale: Set out how a successful team-level use case becomes an approved, supported enterprise tool.

  • Close the loop: Feed insights from discovery tools and disclosure conversations into procurement and platform roadmaps.

 

Moderna again provides a documented example. Within two months of adopting ChatGPT Enterprise, its employees had created 750 custom GPTs, 40% of weekly active users had built at least one, and users averaged 120 conversations a week. Examples included Dose ID, a pilot GPT to support clinical dose selection analysis, Contract Companion for summarizing contracts, and Policy Bot for answering questions about internal policies (OpenAI, n.d.-b). The company's earlier internal chatbot, mChat, built on OpenAI's application programming interface, had been adopted by more than 80% of employees. These figures come from a vendor account and should be read with that in mind, but they illustrate how a sanctioned platform can convert individual ingenuity into reusable, governed assets.

 

JPMorgan Chase's use of prompt libraries and regular prompt-sharing emails is a smaller-scale version of the same idea (McKinsey & Company, 2025), and Allen & Overy's beta trial of Harvey, in which thousands of lawyers tested the tool on real work before firmwide deployment, shows how structured experimentation can precede scaled adoption (Allen & Overy, 2023).

 

A common pitfall is to celebrate grassroots innovation without ever investing in scaling it. If useful team-built assistants remain orphaned, unsupported, and unknown outside one team, employees learn that experimentation earns praise but not resources, and the energy drains away.

 

Building Long-Term Governed AI Capability


The five responses above can reduce shadow AI quickly. Sustaining the shift requires building capability that endures as tools, regulations, and workforce expectations change. Three pillars stand out.

 

Governance as a Living System

AI governance cannot be a document written once and filed. The tools change too quickly, and so does the regulatory environment. The EU AI Act's obligations are phasing in over several years, and frameworks such as the NIST AI RMF and ISO/IEC 42001 are designed around continual monitoring and improvement rather than one-time compliance (National Institute of Standards and Technology, 2023; ISO/IEC, 2023). In practice, this means a cross-functional governance group with representatives from technology, security, legal, HR, and the business, meeting regularly with authority to approve tools and use cases quickly. It means periodic reviews of which AI services are actually in use, informed by discovery tools and disclosure conversations. And it means treating each shadow AI discovery as a data point about unmet needs.

 

IBM's (2025) finding that only about a third of organizations with AI policies audit for unsanctioned use shows how often governance stops at the policy stage. A living system closes this gap by connecting policy to observation and observation to action. Deloitte's finding that only 30% of leaders rate their risk and governance preparedness as high, and only 21% report a mature governance model for autonomous agents, suggests that most organizations are still early on this journey (Deloitte AI Institute, 2026). As AI agents that take actions on employees' behalf become more common, the next wave of shadow AI may involve agents rather than chatbots, making a responsive governance system more important, not less.

 

Leadership Consistency and Trust

Shadow AI is, at root, a trust problem running in both directions. Leaders worry that employees will mishandle data; employees worry that leaders will punish them for trying to do their jobs better or will use AI to replace them. Both concerns have some basis, and neither is resolved by policy alone.

 

Three leadership behaviors matter most. The first is clarity: Gallup's finding that only about a quarter of employees say their organization has communicated a clear AI plan (Gallup, 2026) points to a gap leaders can close directly. Employees need to know why the organization is adopting AI, what it means for their roles, and what is expected of them. The second is consistency: the same rules should apply to executives and frontline staff, and leaders should be seen following them. The third is support. Gallup's data show that employees who strongly agree their manager supports their team's AI use are substantially more likely to use AI frequently (Gallup, 2026). Managers are the point at which policy becomes practice, and they need guidance and permission to encourage sanctioned experimentation.

 

Leaders also need to address the fear of replacement honestly. When more than half of AI users worry that using AI makes them look replaceable (Microsoft & LinkedIn, 2024), concealment is a rational response. Organizations that explain how productivity gains will be shared, invest in reskilling, and redesign roles around human strengths give employees a reason to bring their AI use into the open.

 

Measuring What Matters

Organizations tend to measure what is easy: licenses issued, training completions, policy acknowledgments. These metrics say little about whether shadow AI is declining or whether sanctioned AI is creating value. Ricoh's contrast between decision-makers' confidence in AI returns and workers' experience of friction (Hale, 2026) shows how far leadership perception can drift from employee reality.

 

A better measurement approach combines several lenses. Visibility metrics track the share of AI use occurring through sanctioned channels, informed by discovery tools and anonymous surveys. Capability metrics assess whether employees can use AI well, including their ability to recognize errors and tasks outside the frontier. Value metrics look at outcomes in specific workflows, such as cycle time, quality, or customer satisfaction, rather than generic time-saved estimates, which the UK government's report itself cautions should be interpreted alongside other evidence (Government Digital Service, 2025). Trust metrics, gathered through pulse surveys, ask whether employees feel safe disclosing AI use and whether they believe the rules are fair. Tracking these over time turns shadow AI from an invisible risk into a managed indicator of organizational health.

 

Conclusion


Shadow AI is not a passing compliance headache. It is the visible edge of a profound shift in how work gets done, driven by employees who have discovered that generative AI can make their jobs easier and their output better, and who are not willing to wait for permission. The survey evidence from 2024 through 2026 is consistent: large shares of employees use AI tools their organizations have not provided or approved, they often share sensitive information with those tools, and many hide their use from managers. The breach-cost evidence shows that ungoverned AI use is associated with higher costs and more exposure of sensitive data. The field experiments show that AI can deliver substantial gains, but also that it can lead people astray on the wrong tasks.

 

Prohibition was an understandable first reaction in 2023, and narrow restrictions on specific data or uses remain sensible. But as a primary strategy, prohibition fails on its own terms. It does not eliminate demand, it drives use underground, and it deprives organizations of the visibility and learning they need. The organizations that have moved furthest have done something different. They have provided secure, capable alternatives; built governance that is proportionate and comprehensible; invested in serious, role-based AI literacy; made it safe for employees to disclose and discuss what they are doing; and created pathways for good ideas to scale.

 

The underlying shift is one of stance. Instead of asking how to stop employees from using AI, leaders should ask what employees' unsanctioned use is telling them about unmet needs, missing tools, inadequate training, and fragile trust, and then respond to those signals. Handled well, the energy now flowing through the shadows can become one of the organization's most valuable sources of capability.

 

  • Treat shadow AI as a signal: Use discovery tools and honest conversations to learn what tasks employees are trying to improve, and respond to that demand.

  • Make the sanctioned path the easy path: Provide enterprise AI tools with strong data protections, fast access, and capabilities close to consumer alternatives.

  • Govern by data and risk tier: Give employees simple rules about which information can go into which tools, and reserve heavy approval processes for genuinely high-risk uses.

  • Build literacy that lasts: Offer role-specific, regularly refreshed training that emphasizes verification and the limits of AI, and keep records to meet obligations such as the EU AI Act's Article 4.

  • Make disclosure safe: Offer amnesty for good-faith past use, apply the same rules to leaders and staff, and reserve sanctions for genuine harm.

  • Harvest and scale what works: Build champion networks and shared libraries, and create clear routes from team experiments to supported enterprise tools.

  • Measure visibility, capability, value, and trust: Track the share of AI use in sanctioned channels and whether employees feel safe and competent, not just license counts.

 

Research Infographic




References


  1. Allen & Overy. (2023, February 15). A&O announces exclusive launch partnership with Harvey [Press release]. A&O Shearman.

  2. Alter, S. (2014). Theory of workarounds. Communications of the Association for Information Systems, 34, Article 55.

  3. Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942.

  4. Cueto, J. (2026, April 28). Shadow AI rises as employees outpace workplace controls: Survey. Human Resources Director America.

  5. Dell'Acqua, F., McFowland, E., III, Mollick, E., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality (Working Paper No. 24-013). Harvard Business School.

  6. Deloitte AI Institute. (2026). State of AI in the enterprise: The untapped edge. Deloitte.

  7. Edmondson, A. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44(2), 350–383.

  8. European Parliament and Council of the European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union.

  9. Fried, I. (2023, August 30). Walmart will give 50,000 office workers a generative AI app. Axios.

  10. Gallup. (2026). Indicator: Artificial intelligence. Gallup.

  11. Government Digital Service. (2025). Microsoft 365 Copilot experiment: Cross-government findings report. UK Government.

  12. Haag, S., & Eckhardt, A. (2017). Shadow IT. Business & Information Systems Engineering, 59(6), 469–473.

  13. Hale, C. (2026, October 22). Companies are bringing in more AI than ever, but employees are increasingly just working around it. TechRadar Pro.

  14. IBM. (2025, July 30). IBM report: 13% of organizations reported breaches of AI models or applications, 97% of which reported lacking proper AI access controls [Press release].

  15. ISO/IEC. (2023). ISO/IEC 42001:2023 Information technology: Artificial intelligence: Management system. International Organization for Standardization.

  16. Landi, H. (2024, August 14). Kaiser Permanente rolls out Abridge's gen AI clinical tech across 40-hospital system. Fierce Healthcare.

  17. Lunden, I. (2024, May 29). OpenAI signs 100K PwC workers to ChatGPT's enterprise tier as PwC becomes its first resale partner. TechCrunch.

  18. McKinsey & Company. (2025, October 29). JPMorgan Chase's Derek Waldron on building an AI-first bank culture [Interview].

  19. Mehta, I. (2023, May 19). Apple reportedly limits internal use of AI-powered tools like ChatGPT and GitHub Copilot. TechCrunch.

  20. Microsoft & LinkedIn. (2024). 2024 Work Trend Index annual report: AI at work is here. Now comes the hard part. Microsoft.

  21. Morrison, E. W. (2014). Employee voice and silence. Annual Review of Organizational Psychology and Organizational Behavior, 1, 173–197.

  22. National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). U.S. Department of Commerce.

  23. OpenAI. (n.d.-a). Morgan Stanley uses AI evals to shape the future of financial services. OpenAI.

  24. OpenAI. (n.d.-b). Accelerating the development of life-saving treatments [Moderna customer story]. OpenAI.

  25. PagerDuty. (2026, June 11). PagerDuty report finds two-thirds (66%) of office professionals have used unauthorized AI tools at work [Press release].

  26. Parker, S. K., Williams, H. M., & Turner, N. (2006). Modeling the antecedents of proactive behavior at work. Journal of Applied Psychology, 91(3), 636–652.

  27. Poireault, K. (2025, October 30). Shadow AI: One in four employees use unapproved AI tools, research finds. Infosecurity Magazine.

  28. Ray, S. (2023, May 2). Samsung bans ChatGPT among employees after sensitive code leak. Forbes.

  29. Silic, M., & Back, A. (2014). Shadow IT: A view from behind the curtain. Computers & Security, 45, 274–283.

  30. Travers Smith. (2025). The EU AI Act's AI literacy requirement: Key considerations. Travers Smith.

  31. Venkatesh, V., Morris, M. G., Davis, G. B., & Davis, F. D. (2003). User acceptance of information technology: Toward a unified view. MIS Quarterly, 27(3), 425–478.

  32. Wolters Kluwer Health. (2026, January 22). Shadow AI: Providers are using unapproved tools to improve workflow. Wolters Kluwer.

Jonathan H. Westover, PhD, Chief Research Officer (Nexus Institute for Work and AI); Co-Founder & Chief Workforce and Learning Officer (Future State University); Founder & CEO (Human Capital Innovations); Professor of Organizational Leadership & Change (UVU). Read Jonathan Westover's executive profile here.

Suggested Citation: Westover, J. H. (2026). Out of the Shadows: Why Employees Are Using Unsanctioned AI at Work, and How Leaders Can Turn Hidden Experimentation Into Governed Capability. Human Capital Leadership Review, 41(1). doi.org/10.70175/hclreview.2020.41.1.1



Human Capital Leadership Review

eISSN 2693-9452 (online)

future of work collective transparent.png
Renaissance Project transparent.png

Subscription Form

HCI Academy Logo
Effective Teams in the Workplace
Employee Well being
Fostering Change Agility
Servant Leadership
Strategic Organizational Leadership Capstone
bottom of page