New behavioural data from Redflags has revealed a striking concentration of AI tool usage within UK organisations: just 3% of employees account for 18% of all AI-related activity on work devices, averaging 235 AI events each, compared with 35 for a typical colleague.
The findings come from the Redflags Behavioural Impact Report 2026, which draws on real, on-device telemetry rather than self-reported surveys, a distinction that matters. The report analyses over 29 million behavioural nudges delivered across 44 organisations throughout 2025, covering financial services, engineering and manufacturing, government, and other sectors.
Overall, employee visits to AI websites surged 43% year-on-year in 2025. But the headline number tells only part of the story. The 91% increase in the number of companies actively monitoring AI usage in the same period suggests that security teams are beginning to grasp the scale of the problem, even as governance struggles to keep pace.
Shadow AI and the data egress problem
The report identifies several AI-related behaviours that are keeping security teams up at night: employees uploading files to AI sites, using AI tools without corporate account logins, and accessing unapproved applications. These create data egress points that are difficult to detect without continuous behavioural visibility.
OpenAI accounts for 93% of all AI site visits observed in the data, with Gemini at 5% and Copilot, Perplexity, Claude and DeepSeek each registering under 1%. The dominance of a single tool doesn’t necessarily reduce risk; it depends entirely on whether that tool is being used under a corporate account with appropriate data governance in place.
Tim Ward, CEO of Redflags, said: “The speed at which AI usage is growing inside organisations is remarkable, but what’s equally striking is how many companies are only now starting to understand what’s actually happening on their employees’ devices. Governance is racing to catch up with behaviour, and the gap between the two is where risk lives.”
Ward added, “The human brain is wired to seek novelty. New AI tools trigger dopamine responses associated with excitement and reward. This makes them inherently compelling to employees, regardless of whether they’re approved by the business. Understanding that this is a human behaviour challenge, not just a technology policy one, is critical to building an effective response.”
Phishing: the perennial threat isn’t going anywhere
Alongside the AI findings, the report delivers a timely reminder that foundational threats remain stubbornly persistent. Clicking on links in external emails from unknown senders was the most commonly tracked risky behaviour, flagged by 93% of the organisations in the study.
However, the data also demonstrates that behavioural interventions work. Redflags’ nudge-based approach, delivering just-in-time prompts on employees’ devices at the precise moment of risk, produced an average 35% reduction in dangerous link clicks across the dataset, with peak reductions of 83% in the best-performing organisations.
The mechanism is grounded in cognitive science. Link-clicking from unknown senders is typically a fast, instinctive System 1 decision. Nudges interrupt that automatic response and prompt a more considered System 2 evaluation, whether that’s hovering over a link to verify a URL, or pausing to scrutinise the sender. Over six months, the report recorded a 28% increase in the hover-to-click ratio, indicating that employees are building more cautious habits over time.
The credential loss data reinforces this. A 22% average reduction in passwords being entered on sites reached via unknown-sender email links suggests the nudge effect compounds across the full phishing chain, not just at the point of click.
What this means for security teams
For CISOs and security awareness leads, the report offers a useful benchmark and a methodological argument. The Redflags data is unusual in that it is measured before and after intervention, on actual devices, in real working conditions, not modelled, simulated, or self-reported. That makes it one of the few datasets available that can demonstrate genuine behaviour change rather than claimed behaviour change.
The 3% finding in particular has practical implications. Identifying and monitoring the small cohort of power users who account for a disproportionate share of AI activity, whether enthusiastic early adopters or individuals bypassing policy, may be a more targeted and efficient use of security resources than blanket controls that affect all employees equally.
