Two in three business leaders now believe their company has already suffered a data breach caused by staff using AI tools without permission, according to new research from Writer, the enterprise AI agent platform. The finding matters to any small business owner who has let staff loose on ChatGPT or similar tools without a policy in place.
Two-thirds of leaders (67%) believe their organisation has already suffered a data leak or security breach caused by an employee using an unapproved AI tool, Writer’s survey found. The research, published this week, puts a number on a problem that has been building quietly inside UK workplaces: staff pasting confidential information into public AI chatbots to get work done faster.
The scale of the behaviour is stark. More than a third of employees (35%) admit to having fed proprietary, confidential or sensitive company information into a public AI tool. A further 16% say they use AI tools that are explicitly banned by their employer, most commonly because they feel it is “whatever it takes to get their work done” (40%) or because the tools their employer has approved are simply not good enough (32%).
Writer’s researchers estimate that as many as 5,557 large UK enterprises may already have exposed themselves to this kind of risk. For smaller firms without a dedicated IT or security team, the exposure is likely to be just as real, since there is often no equivalent of a company-approved tool for staff to use instead.
The governance gap extends beyond simple data leaks. More than a quarter of staff (28%) say they have seen an AI tool produce a result that was dangerously wrong, unethical or biased, yet three in ten feel unsafe reporting it for fear of being seen as blocking the business’s AI rollout. Executives, meanwhile, are strikingly out of touch with that fear: 90% believe staff feel safe to speak up.
Confidence in controlling AI agents once they are switched on is also shaky. More than a third of leaders (35%) say they are not confident they could “pull the plug” on an AI agent that started causing financial or reputational harm, despite six in ten believing such an error would cost a senior executive their job. This sits alongside the kind of autonomous-agent oversight problems recently flagged in tools like NinjaOne’s Katana AI safety controls, and echoes warnings from Anthropic about AI worms spreading across devices.
For a small business, the lesson is not to ban AI outright. It is to give staff an approved tool that is good enough that they have no reason to go elsewhere, and a short, written policy on what can and cannot be typed into it. That pairs naturally with wider defences: the same discipline covered in recent reporting on rising cyberattacks and in guidance on spotting data theft tactics used by ransomware gangs.
What to do on Monday: ask every manager which AI tools their team actually uses, not which ones are officially sanctioned. Write a one-page rule on what information can never go into a public AI tool, and name an approved alternative so staff are not left guessing.






