The problem HR is facing
HR and IT news sites keep returning to the same problem in 2026: shadow AI, meaning AI tools employees use without their employer's approval. In a survey of 2,000 employees in the UK and US, run by Sapio Research for the security company BlackFog in November 2025:
- 49% used AI tools at work that their employer had not approved.
- 60% agreed that using unapproved AI is worth the security risk if it helps them work faster or meet a deadline.
- Leaders were the most relaxed about it: 69% of presidents and C level respondents said speed beats privacy or security, against 37% of people in administrative roles.
- 27% had shared employee data in unapproved tools, and 33% research or data sets.
For HR the last number is the uncomfortable one. Employee data is HR data.
How much of your team's data is at risk?
Move the sliders to match your organisation. The defaults are the survey numbers.
A way forward
The reflex is a ban. Research on how people react to rules suggests why that backfires: when a rule blocks a goal people care about, here getting the work done, many keep the goal and hide the behaviour. Psychologists call this reactance. A ban turns 49% of your people into rule breakers, and you lose sight of what they paste where.
What works better is to make the safe way the easy way:
- Give people an approved tool that is as fast as the one they use now. People in the survey chose speed. Compete on speed.
- Use a simple traffic light for data. Green: public information, fine anywhere. Amber: internal, only in the approved tool. Red: personal and employee data, never in any AI tool without a check.
- Ask first, then decide. A short, anonymous amnesty survey ("which AI tools do you use, for what?") tells you what people need. Punishing the first honest answers ends the honesty.
- Leaders go first. The survey found leaders were the most willing to trade security for speed. Rules that leaders visibly break will not hold.
What this means for HR
- You own the most sensitive data. Performance reviews, sick notes, salaries and complaints should be red in the traffic light, and HR should say so first.
- Training is your lever. Most people do not know which data is risky. One 20 minute session with real examples from your own organisation beats a policy document.
- Watch the culture signal. If people hide their AI use from you, they will also hide mistakes. That is a psychological safety problem, not only an IT one.
Honest limits
One vendor commissioned online survey in two countries, with self reported answers. Other 2026 surveys found similar shares of unapproved use, roughly between half and two thirds, depending on how the question was asked. The simulator is an illustration based on these shares, not a risk assessment.
Sources
- Sapio Research for BlackFog (2026). Shadow AI survey of 2,000 UK and US employees, fielded November 2025. source
- Brehm, J. W. (1966). A theory of psychological reactance. Academic Press.