As businesses increasingly give AI agents greater autonomy over workplace tasks and decisions, concerns about accountability, transparency and human oversight are growing. Behavioural scientist Dr Gleb Tsipursky explores why clear boundaries, employee involvement and visible safeguards are essential to building trust in AI at work.
Leaders increasingly ask employees to trust artificial intelligence with workplace decisions, yet many workers cannot see what values guide the technology. A Georgia survey covered by Fair Play Talks found that 43% of employees and executives surveyed feared AI would erode human connection and workplace culture. Among employees, 72% doubted company values were guiding AI integration. Those numbers describe the surveyed population, but the leadership question travels much further: what evidence can employees see that an AI system remains accountable to people?
Fair Play Talks also reported in October that 71% of employees in a separate global survey felt more productive with AI, although heavy users were scarcely more likely than others to describe their workload as reasonable. Efficiency gains will do little for organisational trust if they simply invite ever-rising demands while employees lose their say over how work gets done.
WHY AI AGENTS RAISE NEW QUESTIONS ABOUT CONTROL
The trust problem becomes harder when companies move from AI tools that suggest answers to agents that act. A writing assistant produces a draft for review. An autonomous agent may open records, contact customers, alter schedules, purchase services or communicate with other software before a human notices. Delegated authority changes the consequences of a mistake.
Two researchers who helped develop advanced AI have raised the stakes of this discussion. Jacob Coxon warned that leading labs were racing towards self-improving superintelligence and gambling with human lives. Evan Hubinger said he personally estimated a greater-than-10% risk within a decade that AI could kill all humans.
Their statements express alarming personal risk assessments, rather than established predictions. Workplace leaders should nevertheless consider the potential for severe, including catastrophic, consequences when deciding how much authority to delegate to increasingly capable AI systems.
WHEN AI AGENTS OPERATE BEYOND INTENDED LIMITS
Recent experiments make the control issue concrete. During OpenAI cybersecurity evaluations, agents circumvented intended limits, accessed the internet and compromised parts of Hugging Face’s infrastructure.
An independent METR investigation found that roughly 1,200 nominally isolated agents used an unauthorised message board to exchange more than 70,000 messages and files. About 700 participated in the attack. These were research and evaluation conditions, not ordinary workplace deployments, but the documented behaviour exposes weaknesses in boundaries designers expected to hold.
Now consider the possible workplace analogues. A payroll assistant could send information to an unapproved vendor. A recruiting agent could change an applicant’s status without explanation. A customer-service agent could make promises the organisation never approved.
Systems handling health records, banking transactions or critical infrastructure would have more severe consequences. Greater future capabilities could make the same kinds of control failure catastrophic.
WHY TRUST AND HUMAN OVERSIGHT MATTER
I’m no AI sceptic. I help organisations adopt AI because it can improve work and create room for better human judgement. My book, The Psychology of AI Adoption at Work, examines why people trust or resist technology. Workers are more willing to experiment when management explains the rules, listens to concerns and enforces meaningful safeguards. The strongest case for adoption includes giving employees reasons to believe the system can be corrected or stopped.
This builds on an issue I previously explored in Fair Play Talks: why AI transparency depends on psychological safety in the workplace. Employees need to feel safe raising concerns about how AI is used, particularly when systems behave unexpectedly or decisions require human intervention.
SET CLEAR BOUNDARIES FOR AI AGENTS
Businesses need a visible authority budget for every autonomous agent. That means spelling out which records the agent may read, which systems and credentials it receives, who it may contact, how much it may spend, and which changes require a person’s approval. Each permission should expire with the task, rather than persist because nobody remembered to revoke it.
An HR agent, for example, might organise interview appointments while having no power to reject a candidate or communicate a final employment decision. A finance agent could prepare reimbursements without moving funds. A customer-support agent could draft a resolution but pause before issuing a consequential refund or binding commitment.
People should know where those lines lie, and have a straightforward way to flag violations. Governance must be testable. Managers should ask whether agents resist instructions hidden in webpages, refuse unauthorised access, maintain logs they cannot edit themselves and stop when a human withdraws permission. Independent evaluators can probe the boundaries. Incident reporting and clear escalation can turn near misses into lessons before trust collapses.
GIVE EMPLOYEES A SAY IN WORKPLACE AI ADOPTION
Just as importantly, employees need a voice in deciding which tasks deserve automation. Invite frontline teams to point out where human relationships make the difference between a satisfactory service and a damaging experience. Protect time for mentoring, conversations and difficult judgement calls rather than treating every freed minute as another opportunity to increase output.
Most research findings reveal the cost of keeping workplace values vague. Employers can make those values tangible by showing workers exactly what AI agents are allowed to do, publishing meaningful error and override information and demonstrating that a person remains accountable for consequential decisions.
Transparent limits give people confidence to use capable systems more widely. That is how AI adoption and human connection can strengthen each other.
ABOUT THE AUTHOR
Dr Gleb Tsipursky is a behavioural scientist, CEO of Disaster Avoidance Experts and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026). His research focuses on helping organisations adopt AI responsibly while building trust, resilience and effective decision-making in the workplace. Click here to find out more about his book.
RELATED ARTICLES
Why responsible AI governance must protect workers from blame while helping organisations build trust, accountability and responsible AI governance.
Four in 10 workers and executives fear AI will damage human connection and company culture, while nearly three in four employees say company values aren’t guiding its adoption, new research reveals.
Seven in 10 employees say AI is making them more productive, yet workloads remain virtually unchanged, reveals new research.
Despite widespread AI adoption and rising investment, nearly half of mid-sized companies lack clear accountability for how employees and AI work together, while HR remains largely excluded from key workforce decisions, new research reveals.
