A still poorly documented occupational-health effect
A particularly interesting Wall Street Journal article uses the personal stories of several startup founders to illustrate a still poorly documented effect of working with AI agents. The more autonomous and faster machines become, the more they can accelerate the pace at which humans must decide, supervise and respond.
The situations described are extreme, but they reveal several very concrete occupational-health issues: work intensification, hyperconnectivity, reduced recovery time, blurred boundaries between work and personal life, and potentially workaholism.
Agents that never sleep
Aditya Sharma, co-founder of Keel, explains that he may stay awake very late to prevent his agents from becoming blocked overnight. Even when he is no longer at his computer, he finds it difficult to detach psychologically and continues wondering whether his agents need him.
The issue is therefore no longer only working time. It is also the inability to disengage psychologically. When activity continues in the background and may call on the user at any time, periods normally reserved for recovery risk becoming mere waiting periods between interventions.
Peter Pezaris, founder of Proxon, regularly works until two in the morning and describes using agents as ‘addictive’. He estimates that his business, with six human developers, operates roughly thirty times faster because of them.
This acceleration has a paradoxical effect: the agents’ work creates further human work. Development can move faster, more clients can be onboarded and more projects launched, but this then means handling more requests, making more decisions and supervising a greater volume of activity.
This is a classic work-intensification mechanism. A productivity gain at task level does not necessarily release time. It may be immediately reinvested in new objectives.
“Every minute that I’m not working, I’m missing out on not doing a week’s worth of work.”
When an hour spent with agents can theoretically accomplish what previously required several days, rest may gradually be perceived as a lost productive opportunity.
Abby Grills, co-founder of Riveter, describes another phenomenon. Her company has only two people and regularly postpones recruitment because new tasks can be assigned to agents. Yet she acknowledges being constantly tired.
AI can therefore allow very small organisations to handle a volume of activity once reserved for much larger teams. But if staffing does not keep pace with increased productive capacity, responsibility, decisions and mental workload may become concentrated on fewer people.
Ajay Kalia, founder of Alt, works approximately from 5:30 a.m. to 10 p.m. He began controlling some agents from his Apple Watch, which can vibrate every ten minutes to request authorisation, including while he is exercising.
The example is particularly telling. Work no longer stops when a person leaves their desk. It follows them into the moments that should enable recovery. Attentional fragmentation and hyperconnectivity then become direct features of work organisation.
Can agents favour workaholism?
These accounts also evoke workaholism, which does not simply mean working a lot. It describes a pattern of work that is both excessive and compulsive, characterised in particular by difficulty stopping and mentally detaching from work.
AI agents may create an environment especially conducive to this pattern. They can continue producing while the human sleeps, eats or exercises. Every completed task, notification or temporarily blocked agent creates another opportunity to intervene.
The permanent availability of productive capacity may gradually change how rest is perceived. Not working can begin to feel like leaving a resource unused.
This is precisely why these examples matter for occupational health. A technology can dramatically improve productivity at task level while worsening working conditions in the system in which that task sits.
What should occupational-health professionals look at?
The arrival of AI agents in an organisation should not be assessed only through the time they can theoretically save. We must also examine what the organisation does with that productivity gain.
- Identify real work intensification.
Determine whether AI actually reduces workload or simply increases the number of cases, clients, projects or targets handled in the same time. - Assess expected availability to agents.
Determine whether a worker must respond when an agent is blocked, whether agents continue operating at night or over weekends, and whether notifications can interrupt rest. - Protect genuine periods of disconnection.
The organisation should define times when no human supervision is expected, even if an agent is temporarily blocked. The cost of an agent waiting a few hours should not justify routinely interrupting human rest. - Monitor staffing levels.
Greater productive capacity should not automatically delay recruitment or indefinitely expand existing roles. - Examine the invisible work agents create.
Checking outputs, correcting errors, restarting agents, arbitrating, handling exceptions and coordinating are work in their own right. They must be identified and included in workload assessment. - Look for difficulty switching off.
Compulsively checking agent status, intervening at night, feeling guilty when not working or treating every rest period as lost productivity may be warning signs. - Do not measure productivity alone.
A relevant deployment assessment should also track hours, interruptions, perceived workload, autonomy, recovery time and collective work.
The question when deploying agents may therefore not simply be: ‘How much time does this technology save us?’
We should also ask: ‘What happens to that time once it has been saved?’
This is probably where much of the real effect of agentic AI on work—and ultimately on health—will be determined.