Artificial intelligence has quietly become part of the working day. Employees are using digital assistants to summarize meetings, polish emails, analyze reports and turn rough ideas into presentable work—sometimes before their employers have approved the tools or decided what information can safely be entered into them.
That gap between workplace demand and corporate oversight is creating what security specialists call “shadow AI.” Recent findings suggest the practice is not confined to a small group of rule-breakers. Depending on the occupation and how unauthorized use is defined, studies have placed the share anywhere from roughly one-third to one-half of professionals. The challenge for employers is no longer deciding whether workers will use AI. It is determining how to make that use productive, visible and safe.
The Headline Number May Understate the Scale
The “one in three” description is best understood as a conservative summary of several overlapping workplace trends rather than a universal rate across every industry. Ivanti’s 2025 workplace research, based on more than 6,000 office workers and 1,200 IT and cybersecurity professionals, found that 42% of office workers were using generative AI at work, up from 26% a year earlier. Among those using generative AI, 32% said they kept that use secret from their employer. The research also found that 46% of office workers used at least some AI tools that were not provided by their employer, while 38% of IT professionals acknowledged using unauthorized tools. Those figures describe slightly different behaviours—secrecy, outside-tool adoption and explicit lack of authorization—but together they reveal a workplace where AI usage is frequently occurring beyond formal oversight. A communications employee may use a personal chatbot to soften the tone of an email, while a developer may rely on an outside coding assistant because the approved system cannot solve a particular problem. Both actions can fall into the shadow-AI category, even though their technical and legal risks are very different.
Other findings suggest the Ivanti numbers may not capture the full extent of the practice in professional services. Intapp questioned 820 professionals working in accounting, consulting, finance and law and found that 72% were using AI at work, compared with 48% in its previous annual findings. Half said they had used a work-related AI tool that their firm had not provided or recommended. Of the total, 24% reported doing so many times and 26% said they had tried it once or twice. A separate ManageEngine study conducted by Censuswide questioned 700 full-time professionals and IT decision-makers at larger organizations in the United States and Canada. Seventy percent of the IT leaders said they had identified unauthorized AI usage, while 60% of employees said their use of unapproved tools had increased during the previous year. The studies should not be combined into a single worldwide rate because their samples, questions and definitions differ. Still, their direction is remarkably consistent: employee adoption is moving faster than procurement, security reviews and workplace policies. The phrase “one in three” therefore reflects the lower end of a broader pattern, not the outer limit of the problem.
Productivity Pressure Is Driving Workers Outside the Rules
Most shadow-AI use appears to begin with an ordinary workplace frustration rather than an intention to expose company information. In the ManageEngine findings, the most common unauthorized uses included summarizing meetings or calls, cited by 56% of respondents; brainstorming ideas or content, cited by 55%; analyzing reports and drafting or editing documents, both at 47%; and creating client-facing material, at 34%. These are not obscure technical experiments. They are routine assignments that can consume hours of an employee’s week. Picture a consultant facing an afternoon deadline and a lengthy set of meeting notes. The company-approved assistant may be unavailable, slow or restricted to a narrow set of tasks, while a familiar public tool can produce a usable outline in seconds. From the employee’s perspective, the choice may feel less like breaking a security rule and more like using a calculator that happens not to be on the approved list. The problem is that the notes may contain client names, financial assumptions or strategic information that should never leave the organization’s controlled systems.
Employees also report emotional and organizational reasons for keeping AI usage out of sight. Ivanti found that 36% of workers who concealed their AI use liked having a “secret advantage,” while 30% worried their job could be eliminated and 27% did not want colleagues to question their ability. More than half of office workers agreed that becoming more efficient often results in being assigned more work. That creates an uncomfortable incentive: an employee who completes a three-hour assignment in one hour may decide there is little personal benefit in explaining how it was done. At the same time, Intapp’s professional-services findings help explain why workers are reluctant to abandon the technology. Sixty-two percent of AI users described it as highly useful. Among professionals saving time with AI, 42% said they redirected some of that time toward higher-level client work, 33% toward strategy and planning, and 24% toward increasing billable hours. Shadow AI is therefore not simply a story about careless employees. It also reflects a mismatch between what organizations officially provide and what workers believe they need to meet deadlines, maintain their performance and remain competitive. A policy that only says “do not use AI” does not remove those pressures; it can merely push the behaviour onto personal accounts and devices where the employer has even less visibility.
The Biggest Danger Is What Employees Put Into the Tools
The central risk is not that an employee asks an outside chatbot to improve a generic sentence. It is that the prompt, uploaded document or connected application may contain information that the organization has a legal or commercial duty to protect. ManageEngine found that 37% of surveyed employees had shared internal documents such as strategies or financial material with unauthorized AI tools. Thirty-three percent reported sharing confidential client information, 32% had entered non-public product information and 37% had included information about colleagues or team members. The same findings exposed a confidence problem: 90% of employees said they trusted unauthorized AI tools to protect their data, while half believed there was little or no risk in using them. Yet the organization may not know where the information is processed, how long it is retained, whether it is used to improve the service or what contractual protections apply. There is also the possibility that an AI-generated answer will be inaccurate but convincing. In a low-stakes brainstorming session, that may create an awkward sentence. In financial, legal, health or employment-related work, it can influence a consequential decision or place incorrect information in front of a client.
The potential cost becomes clearer when examining organizations that have already experienced data breaches. IBM and the Ponemon Institute studied breaches at 600 organizations around the world between March 2024 and February 2025. One in five of those breached organizations reported an incident linked to shadow AI, while organizations with high levels of shadow-AI activity recorded average breach costs that were $670,000 higher than those with little or none. The report also found that 63% of the breached organizations either lacked an AI-governance policy or were still developing one. The solution, however, is not necessarily a blanket ban. Canadian privacy authorities advise organizations using generative AI to establish a valid basis for handling personal information, use anonymized or de-identified information where possible, assess privacy impacts, evaluate accuracy and apply safeguards suited to the sensitivity of the data. A workable employer response would translate those principles into everyday choices: a short list of approved tools, clear examples of prohibited inputs, secure enterprise accounts, human review for consequential outputs and a fast process for requesting new capabilities. Workers should know that asking for help will not automatically trigger discipline. When approved technology is practical and policies are understandable, employees have fewer reasons to hide the tools they have already made part of their jobs.
