Shadow AI: 78% of AI users bring their own tools to work

Shadow AI is the use of AI tools at work without the company’s approval or visibility, and the data shows it has become the norm. In the 2024 Work Trend Index by Microsoft and LinkedIn, a vendor survey run by Edelman Data & Intelligence with about 31,000 knowledge workers in 31 countries, Brazil included, between February and March 2024, 75% said they used generative AI at work, and 78% of those users brought their own tools, a practice Microsoft named BYOAI (bring your own AI). A larger academic study, from the University of Melbourne with KPMG, found about half of employees admitting they had put sensitive company information into public AI tools, with the behavior more frequent in organizations that had banned generative AI (67%) than in those with no rule at all (33%). My reading is that a ban pushes usage out of leadership’s sight, and what lowers the risk is the combination of a short policy, a corporate tool and training for the people who use it.

At Iasbeck&Co we implement CRM and AI agents in B2B operations and train the teams who work with them, and hidden AI use shows up in the diagnosis of almost every project. I gathered here the surveys with published samples, the Samsung case from the original news sources and what companies are doing about it. The step by step for writing the rules is in AI use policy, and what Brazilian data protection law requires when the data is personal is in AI and LGPD.

What are shadow AI and BYOAI?

The term comes from shadow IT, the software a business unit buys or installs without going through IT. The most common case is the professional who opens an AI assistant on a personal account and pastes work text into it, which is Microsoft’s BYOAI and what the surveys below measure. The definition also covers the AI feature that an existing vendor switches on in an update without anyone reviewing the contract, and the agent an employee builds on their own and connects to a company spreadsheet. Each form calls for a different answer: the personal account is solved with a corporate tool and a data rule, the vendor feature with a contract review, and the employee-built agent with a registry and an owner, the subject of AI agents.

How many employees use AI on their own?

Large-sample surveys, run by different organizations with different questions, arrive at a similar picture: AI use at work has passed half of professionals, and most of it happens in tools the company did not provide.

Work Trend Index 2024: 75% use AI, 78% bring their own

Microsoft and LinkedIn collected the data between February 15 and March 28, 2024, among full-time knowledge workers. Beyond the 75% usage and 78% BYOAI figures, the study reports that the practice reaches 80% in small and medium-sized companies and that 52% of people who use AI at work are reluctant to admit using it for their most important tasks. Only 39% of users had received AI training from their company. Microsoft sells AI tools to companies and has an interest in the conclusion that employees need a corporate tool; I use the data because the sample, the method and the research firm are published.

University of Melbourne and KPMG: 70% use free tools

The 2025 global study by Gillespie and colleagues at the University of Melbourne, with KPMG surveyed 48,340 people in 47 countries between November 2024 and January 2025 through an anonymous online questionnaire, and 32,352 of them answered the section about work. Three in five employees (58%) said they use AI regularly at work. Among those who use it, 70% rely on free public tools and 42% on tools provided by their employer. Only two in five said their organization has a policy or guidance on generative AI, and almost one in five did not know whether there was one.

What security and IT leaders see

Gartner surveyed 302 cybersecurity leaders between March and May 2025, and 69% said their organization suspects or has evidence that employees use prohibited public generative AI. The figure refers to organizations, as seen by the people in charge of security, and should not be quoted as a share of employees. In the May 2026 survey by ISACA, the international association of audit, governance and information security professionals, with more than 3,400 respondents, 90% believe employees use AI in their organization and only 38% say the company has a formal, comprehensive AI policy. Another 30% have a limited policy, and 25% have no active policy.

What data do employees put into public AI?

The Melbourne study asked how often each employee had done a series of things at work. About half (48% to 49%) said they had uploaded sensitive company information to public AI, such as financial, sales or customer data, or copyrighted material. More than half (56%) had used AI without knowing whether it was allowed, and 57% admitted non-transparent use, such as presenting AI-generated content as their own. The authors warn that, even with an anonymous questionnaire, the answers may understate the behavior.

Behavior at work (Melbourne and KPMG, 2025)	Never	Rarely	Sometimes or more
Uploaded company information into a public AI tool	52%	14%	34%
Used AI in ways that contravene policies or guidelines	56%	13%	31%
Used AI without knowing whether it is allowed	44%	18%	38%
Avoided revealing they used AI in their work	39%	19%	42%

Two vendor surveys detail the type of data. TELUS Digital, a technology and AI data services company, surveyed in January 2025, through the Pollfish platform, 1,000 employees of US companies with 5,000 or more staff who use AI assistants. Of these, 68% access public assistants through personal accounts and 57% admitted entering sensitive information into them: personal data such as names, emails and phone numbers (31%), unreleased product or project details (29%), customer information, including chat logs and recorded calls (21%), and financial data such as revenue, margins and forecasts (11%). In Cisco’s privacy study published in January 2024, with 2,600 privacy and security professionals in 12 countries, 48% admitted entering non-public company information into generative AI, and the sample is made up of the very people who should know the risk.

In a commercial operation, the TELUS list looks familiar. The marketing analyst uploads the spreadsheet of event registrants, with name, title and phone number, to ask for a segmentation, and the SDR pastes into a personal chat the call transcript produced by a free app. At month end, the manager asks the assistant on a personal account to analyze the quarter’s forecast, with the value of each deal. All of them wanted to deliver faster, and none of them read the tool’s terms of use.

What happened at Samsung in 2023?

The most cited case of a leak by employees has two primary news sources. On March 30, 2023, the South Korean magazine Economist Korea published as an exclusive that Samsung Electronics’ semiconductor division had allowed ChatGPT with a notice telling employees to be careful with information security, and that in less than 20 days the company identified three incidents. Two involved source code, pasted into the chat to fix an error and to optimize a program that identifies defects, and the third was the transcript of an internal meeting, pasted to produce the minutes. The company limited each prompt to 1,024 bytes, opened an inquiry with possible disciplinary action and began studying its own AI service. On May 2, Bloomberg reported, based on an internal memo, that Samsung had banned generative AI tools for staff at one of its largest divisions, concerned about data stored on external servers and hard to delete, and that it was preparing internal tools.

The case is usually told as engineers being careless, and the timeline tells another story. The company released a public tool with a generic notice, without a corporate tool, without a rule on what data could go in and without training, and responded to the incident with a ban. The employees were debugging code and writing minutes with the tool they had at hand. This sequence of releasing without structure and then banning is the one I see repeated most often in companies.

Does banning AI at work work?

The strongest data against a ban comes from the Melbourne study. Uploading company information to public AI and using AI against the rules were more common among employees of organizations that had banned generative AI (67%) or had a policy guiding its use (56%) than in organizations without a policy (33%) or among those unsure whether there was a rule (38%). The authors conclude that outright bans may be ineffective, that having a policy does not guarantee compliance and that clear guidance and education on responsible use are needed.

The comparison is a correlation and allows other explanations. Companies that ban AI may be those with more sensitive data and more use, and an employee in a company with a rule knows when they broke it, while one in a company without a rule may not notice. Even so, the study found no sign that banning reduces the behavior, and the other numbers point the same way. In the Cisco study, 27% of organizations had banned generative AI, at least for a while, while almost half of the professionals admitted entering non-public information into it. In the TELUS survey, 57% entered sensitive information into public assistants, in a sample where 29% acknowledged their company prohibits that practice.

What does shadow AI cost when things go wrong?

The 2025 Cost of a Data Breach report from IBM, conducted by the Ponemon Institute and sponsored and analyzed by IBM, studied 600 organizations that suffered a breach between March 2024 and February 2025. One in five reported a breach caused by shadow AI, and those with high levels of shadow AI faced on average US$670,000 more in breach costs than those with low levels or none. Among these organizations, 63% had no AI governance policy or were still developing one. The sample includes only breached companies, the cost is an estimate and IBM sells security, so I read the number as an order of magnitude. Gartner, in the release cited above, predicts that by 2030 more than 40% of enterprises will experience security or compliance incidents linked to shadow AI.

What are companies doing about shadow AI?

The measures in the surveys fall into three groups: rules, tools and training. The numbers come from different surveys and audiences, so they cannot be added up or compared as a series.

Measure	Data	Source
Formal, comprehensive AI policy	38% of organizations (28% in 2025)	ISACA, 2026
Limits on what data can be entered into AI	63% of organizations	Cisco, 2024
Limits on which tools employees can use	61% of organizations	Cisco, 2024
Ban on generative AI, at least temporary	27% of organizations	Cisco, 2024
Tool provided by the employer	42% of employees who use AI	Melbourne and KPMG, 2025
AI training for all employees	33% of organizations (22% in 2025)	ISACA, 2026
AI training received from the company	39% of AI users	Work Trend Index, 2024

Each measure fails when it arrives alone. In the TELUS survey, 22% of employees who have a company-provided assistant still use personal accounts, and in the Melbourne study risky behavior shows up in 56% of employees at organizations with a policy. Training without an approved tool teaches people to use their personal account better. That is why I implement the three together, in this order: the tool the company buys, the rule that says what goes into it and the training that teaches people to use both in their daily work.

Summary table of the surveys cited

Source	Year	Sample and method	Main finding
Microsoft and LinkedIn	2024	~31,000 knowledge workers, 31 countries; vendor survey	78% of users bring their own AI
Melbourne and KPMG	2025	48,340 people, 47 countries, anonymous questionnaire	Half uploaded sensitive data to public AI
Gartner	2025	302 cybersecurity leaders	69% of organizations suspect prohibited use
ISACA	2026	3,400+ governance and security professionals	38% have a formal AI policy
TELUS Digital	2025	1,000 employees of large US companies; vendor	57% entered sensitive information
Cisco	2024	2,600 privacy professionals, 12 countries; vendor	27% of organizations banned it
IBM and Ponemon	2025	600 breached organizations; vendor	US$670,000 more with high shadow AI
Economist Korea and Bloomberg	2023	News reports, the second based on an internal memo	Samsung: three incidents and a ban

What this data does not tell you

Almost all of these numbers are self-reported: people say whether they use AI, what data they entered and whether they knew the rule. The surveys of managers measure perception, since Gartner asks whether the organization suspects prohibited use and ISACA asks whether the professional believes employees use AI. Four sources come from companies that sell AI, data services or security, and each has a commercial interest in solutions to the problem it measures. Definitions also vary: “sensitive information” in the TELUS survey includes names and phone numbers, Cisco asks about non-public company information, and the Work Trend Index covers only knowledge workers.

The data on bans and on breach costs shows association, and none of the studies was designed to prove cause. Brazil-specific data is also missing. The Work Trend Index and the Melbourne study include Brazil in their samples, but the numbers cited here are global, and Cetic.br’s TIC Empresas survey, which I discuss in AI in Brazilian companies, measures adoption by the company and only asks firms with an IT department or specialist, leaving out employees’ informal use.

What leadership does with these numbers

1. Assume the use already exists and measure before deciding. An anonymous survey, with a guarantee that declared use will not be punished, shows which tools the team uses, with which account and for which tasks.
2. Buy the corporate tool before publishing the rule. Choose it for the tasks that came up in the survey, with login through the company email and, when the vendor offers it, a setting that prevents the content from being used to train the model. Banning personal accounts without offering an alternative repeats the Samsung case.
3. Write a short rule about data. Say what type of information can go into each tool, with examples from the commercial routine such as the event registrant list and the quarterly forecast.
4. Bring the use that touches customer data into the CRM. History, transcripts and tickets yield more in an agent with defined permissions, an owner and a registry than in a loose chat, and the company can see what is running.
5. Train people by the role they play. Sales reps, marketing analysts and CS need to learn to ask, to check and to recognize the data that must not leave.
6. Review every quarter. Track how much of the use moved to corporate accounts, how fast requests for new tools get answered and which AI features vendors switched on in the systems you already pay for.

What HR and L&D take from this data

For HR, the number that matters most is about transparency. In the Work Trend Index, 52% are reluctant to admit using AI for important tasks, and in the Melbourne study 57% admitted hiding their use or presenting generated content as their own. As long as people believe using AI counts against them, the use stays hidden, and no policy reaches what the company cannot see. HR helps when it treats declared use as a skill, puts the data rule into the onboarding of employees and contractors and organizes training by role, as I describe in AI training for companies. In the Melbourne study, perceived organizational support for responsible use was the factor most associated with benefits from AI at work, with a correlation of 0.56, above training (0.24) and knowledge (0.42).

Shadow AI shrinks when the company offers a tool, says what data can go into it and teaches people to use both in the work they already do. That third part is the work of the in-company applied AI training, with groups for leadership, sales, marketing and customer service built on the company’s processes and rules. It delivers the prompt library by area and the agent architecture, with what each agent reads, what it executes and who approves.

Frequently asked questions

What is shadow AI?

Shadow AI is the use of AI tools at work without the company’s approval or visibility, such as an assistant on a personal account, a browser extension or an AI feature a vendor switched on without review. BYOAI, short for bring your own AI, is the name Microsoft gave to the habit of bringing your own tool to work.

How many employees use AI without company authorization?

In the 2024 Work Trend Index, 78% of professionals who use AI at work brought their own tools. In the University of Melbourne study with KPMG, 70% of employees who use AI rely on free public tools, and 56% said they had used AI without knowing whether it was allowed.

Does banning ChatGPT at the company solve the problem?

The available data suggests it does not. In the Melbourne and KPMG study, uploading company information to public AI was more common among employees of organizations that banned generative AI (67%) than in organizations without a policy (33%). The result shows association and does not prove cause, but none of the studies cited here found a protective effect from a ban.

What kind of data do employees put into public AI?

In the TELUS Digital survey of 1,000 employees at large US companies, the most cited were personal data (31%), unreleased product or project details (29%), customer information (21%) and confidential financial data (11%). In a commercial operation, that includes lead lists, call transcripts and revenue forecasts.

What happened with Samsung and ChatGPT?

In March 2023, Economist Korea reported three incidents in the semiconductor division in less than 20 days after ChatGPT was allowed, two involving source code and one involving a meeting transcript. In May, Bloomberg reported, based on an internal memo, that the company banned generative AI tools at one of its largest divisions and was preparing its own tools.

How do I prepare the team to use AI without exposing company data?

Buy the corporate tool first for the tasks the team already does in hiding, write a short rule on what data can go into each tool and train each area with examples from its own routine. Measure the shift of usage to corporate accounts over the following months and treat declared use as a skill, so nobody needs to hide it.