This piece is adapted from a Linkedin article by Garth Watson, Co-Founder at ERM Libryo: "Just Five More Minutes"

A safety moment

Last month I noticed I'd started going to bed a full hour later than I used to, and it took me a while to work out why. I hadn't taken on more meetings. My calendar looked the same as it always had. What had changed was what happened after the calendar emptied out, when I'd sit down "just to check on" an AI agent that was still working through something, and end up watching it for another forty minutes before I noticed the time.

My sleep tracker doesn't argue with me about it either. I've always been fairly disciplined about protecting my evenings, exercise, a proper stretch of thinking time, enough sleep to actually function the next day, these were never negotiable for me. And yet here I was, an hour short most nights, telling myself it was just this once.

I want to treat this as a safety moment in its own right; the same way we'd flag a loose stair rail or a near miss on a site visit. Except this one crept up on me slowly enough that I didn't clock it as a hazard until the pattern had already set in, and I suspect most safety frameworks haven't caught up to it either.

The hazard we haven't named

AI doesn't just add to your workload. It changes your relationship with time itself.

Before agentic tools, the size of a task puts a rough ceiling on how long you'd work at it. You finished, or you ran out of things to usefully do, and either way there was a natural stopping point. An AI agent removes that ceiling entirely. It can keep going as long as you let it, and every extra few minutes produces something: another draft, another pass, another improvement you didn't ask for but now can't unsee. A year ago, having a fleet of agents to work through the night on your behalf was still science fiction. Now it's just Tuesday. That is genuinely intoxicating, and I don't mean that as a figure of speech.

Why it's so hard to stop

The pull isn't a character flaw. It's mechanism, and biology.

Researchers who study reward and attention have shown that the brain dopamine response peaks while you're waiting to see what happens, well before the outcome itself arrives, and that's precisely the mechanism a slot machine runs on. It's the not-knowing that hooks you, far more than whatever the machine eventually pays out. An AI agent that's still "thinking" while you wait produces almost exactly the same effect, except now it's wrapped around your actual job, which makes the compulsion feel like diligence instead of what it is. Early research on this is still young, mostly done on students and companion chatbots rather than power users doing real work, so I'd hold it loosely. But some of the first studies on AI dependency, modelled on the same criteria used for substance-use patterns, already describe roughly a third of heavy users showing compulsive checking and failed attempts to cut back. That lines up uncomfortably well with how it actually feels.

It's the not-knowing that hooks you, far more than whatever the machine eventually pays out.

The true cost

The traditional argument for giving people flexible, powerful tools was that flexibility gives time back. What actually happens is closer to the opposite: tools capable of working at any hour tend to colonize every hour. This is sometimes called the autonomy paradox, the freedom to work anytime quietly resets everyone's expectation that you will. Workplace research has tracked a steady climb in after-hours digital activity for a few years now, with knowledge workers interrupted roughly every two minutes during the day. That constant fragmentation pushes the real, focused work into the only quiet window left, which for a lot of people is now somewhere between eleven at night and one in the morning.

This shows up especially clearly for anyone whose work has no natural finish line, and compliance and EHS professionals know that feeling well. Regulations change constantly, across more jurisdictions than one person can realistically track by hand, and a tool that promises to keep watching for you is very hard to switch off. At Libryo, we build regulatory tracking tools precisely to solve that problem, so nobody misses a change because they couldn't watch everything themselves. But it's worth being honest that the same capability can just as easily convince someone there's always one more update worth checking tonight.

Solving the missed-update problem doesn't automatically solve the never-off-duty problem that can quietly replace it.

What it costs isn't visible on a timesheet, which is part of why it's easy to miss. What it actually costs is sleep, and the movement and thinking time a lot of us used to ring-fence deliberately. Sleep and exercise aren't lifestyle extras. They're the maintenance schedule for the prefrontal cortex, where judgment and planning live, and that's exactly the system that degrades first. The safety framing here is stark: cognitive research puts the impairment from roughly seventeen hours without sleep in the same range as a 0.05% blood-alcohol reading, and a full day awake pushes that closer to double, well past most countries' legal drink-drive threshold.

Nobody would accept a colleague operating machinery, or signing off a risk assessment, in that state.

Strip away exercise on top of that for long enough, and you get the WHO's own definition of burnout: exhaustion, cynicism, and falling effectiveness, now formally recognised as an occupational phenomenon in its own right. The irony is hard to miss. AI is sold as a multiplier of output but used without limits it erodes exactly the judgment that made the output worth having.

What the rulebook overlooks

Most of the current regulatory attention, including in the EU AI Act, is aimed at systems that monitor workers: tools that track productivity, flag underperformance, or make automated decisions about people, and rightly demand human oversight where they do. What it barely touches is the mirror image of that problem: the worker who wields the AI and struggles to put it down. That's not something an algorithm is doing to anyone. It's something capable, motivated people do to themselves, because the tool is genuinely that good. No regulator mandated my last few late nights, which is exactly why no regulator is going to catch it either.

Why this one belongs on our risk register

We're not short of a framework for this. We're just not pointing it out at AI yet.

Occupational health and safety already have the language for workload, fatigue, and psychosocial risk, the same categories that would already sit in a mature risk register for shift work or long-haul driving. ISO 45003 specifically, names workload, work intensification, and blurred boundaries as hazards to be managed, not moods to be pushed through. AI-driven overwork fits that definition cleanly. We've mostly missed it only because it arrives dressed as productivity, and because the person most exposed is often the most senior and the most willing.

Five practical changes 

1. Name it. Put AI-enabled overwork on the risk register as its own item, distinct from generic workload or burnout categories, because the anticipation-driven mechanism behind it needs its own controls, not borrowed ones.
2. Treat fatigue as a fitness-for-duty question. Given how closely sleep loss tracks with alcohol-level impairment, a submission finished at 1am deserves the same scrutiny we'd give any other safety-critical decision made under the influence, rather than the quiet admiration it usually gets instead.
3. Lead from the top. Policies about logging off carry very little weight if the person who wrote them is visibly online at midnight. Seen behaviour from senior people travels through a team far faster than anything written in a handbook.
4. Create a switch off. Design the stopping point into the system itself, rather than leaving it to individual willpower. A tool that quietly queues the next ten tasks by default is asking to be run past midnight. One that offers a natural pause instead, a daily digest rather than a live feed, for example, gives people somewhere to stop without having to fight the tool for it.
5. Talk about it openly. Being always-on shouldn't quietly earn admiration it hasn't actually deserved, and people need to hear, out loud, that stopping is allowed. None of this is about wanting people to be less ambitious. It's about wanting that ambition to still be there in a year, rather than burnt through in six months.

How we overcome this new safety hazard

If there's one line worth taking away from this, it's this one. The instinct behind "just five more minutes" isn't laziness, and it isn't poor discipline. It's the same instinct that makes good people good at their jobs: discomfort with leaving something unfinished. I don't think the answer is arguing anyone out of that instinct. It's learning to hear "just five more minutes" for what it actually is, not evidence of dedication, but an early symptom of a safety risk that’s worth paying attention to. We already know how to manage hazards once we've agreed to call them that. This one just needs the same honesty we'd bring to anything else on the risk register, even when it's dressed up as the thing that's working.

"Just five more minutes" isn't evidence of dedication. It's an early symptom worth paying attention to.

In a world where a single missed safety signal can be the difference between a routine shift and an incident report, that awareness isn't optional - it's the job. Curious how a site-specific legal register keeps safety and compliance requirements visible before they become a problem? Get in touch with the Libryo team.

Related reading

• How to Track Regulatory Changes Across Multiple Jurisdictions (Without Losing Your Mind) - the practical flip side of this piece, on staying on top of constant regulatory change without it eating your evenings.
• What is AI Slop and How to Avoid It? - the quality risk that shows up when AI output gets pushed too fast, the twin of the burnout risk covered here.
• How to Stay Ahead of Dynamic EHS Compliance Regulations - for teams managing the same constantly-shifting compliance demands that make it so hard to switch AI tools off at night.

Further reading: Dawson & Reid on fatigue and alcohol-equivalent impairment (Nature, 1997); Mazmanian, Orlikowski & Yates on the autonomy paradox (Organization Science, 2013); Microsoft's Work Trend Index on the "infinite workday" (2025); ISO 45003:2021 on psychological health and safety at work; WHO's ICD-11 classification of burnout (2019); and Article 14 of the EU AI Act on human oversight.