Newsletter #29 · Human Skills in the Age of AI
What the New Job Does to Your Brain — Twenty Conversations, No One Who Waits
In short: If you take one thing from this: AI agents do not get tired, and they do not wait. The person supervising them does both. Ahead: what reading everything on a screen does to understanding; what every switch between conversations costs; the colleague who is always on, and the law on being off; where the ideas come back from, and one rule for whoever signs. Part 1 was last week: Everyone Is a Manager Now.
For a while, each of my AI colleagues sent me a push notification when it finished. I read the first lines and zoomed into some. Then I switched them off. I was on the phone all day, and the information was more than my head could hold. What remains is text. My agents write a lot, I see a lot, and somewhere in me sits the fear of missing something. So I skim. I learned speed-reading about 15 years ago and still highly recommend it. However, it did not make the problem disappear.
Last edition was about the new role: the specialist who manages agents. This one is about what that role does to the person holding it, and what to do. The answer is not less AI. It is a different working day.
If you run a company, this is your problem first: it hits your best people. Boston Consulting Group surveyed 1,488 US workers for Harvard Business Review this year: fourteen percent of AI users reported "brain fry", mental fatigue from using or supervising AI beyond their capacity. Those who reported it made 39 percent more serious errors; perceived productivity rose from one tool to two, flattened at three and fell beyond. In July, four McKinsey authors wrote that the constraint on AI gains is not the technology but the capacity of the people using it. How many tools and agents one person runs, and whether anyone teaches them the self-leadership and self-care to decide for themselves when their day ends, is decided at the top, or by nobody.
(Second of two editions on working with AI agents: the first was the role, this one the brain.)
Reading on a screen, all day
The new job asks for reading, more of it and faster. The cost is older than AI. Pablo Delgado and colleagues at the University of Valencia combined 54 studies with 171,000 readers in 2018: people understand informational texts slightly less well on screens than on paper, and the gap grows under time pressure. Virginia Clinton at the University of North Dakota found a second effect across 33 studies: screen readers were no slower, but worse at judging what they had understood. Screens make us confident, not better.
With AI output the confidence has a second source. Researchers at Microsoft and Carnegie Mellon found among 319 knowledge workers that the more people trusted the AI, the less critical thinking they reported putting in. So the text that most needs careful reading, the agent's confident report, is the one we skim. Speed-reading helps, with one caveat: Keith Rayner and colleagues reviewed the evidence in 2016: there is no way round the trade-off between speed and understanding. Skimming works when you already know the territory. It decides what to read slowly; it does not replace reading.
What I read slowly is this newsletter, because it leaves the house. For that I block the time, close every other tool and go paragraph by paragraph. Even then, a catch. I ask for a correction, wait while the agent rewrites, and in that moment my head wants to use the pause for something else. After a coffee it is worse: I want everything at once, get impatient, skip a paragraph. How calm I am decides how well I read. That is not a technique. It is a state, and a day with twenty conversations works against it.
Every switch has a price
Stephen Monsell, now at the University of Exeter, showed with Robert Rogers in 1995 that changing from one task to another costs time and accuracy even when the change is predictable. The cost per switch is small. The bill per day is not.
Gloria Mark at the University of California, Irvine, has measured this for twenty years. In 2004, people in her studies stayed on one screen for about two and a half minutes. From 2016 onward it was 47 seconds. And in a lab study with 48 people, those who were interrupted every two minutes finished their work faster than the uninterrupted, at the price of more stress, frustration and effort. That is the agent day in miniature: faster, and worse.
Agents make the switch structural. An agent finishes and sits there, three others finish too, and you, last edition's manager, are the bottleneck. My days are not the same twice, with workshops and travel, so I have no two fixed slots. What I have is a morning rule: the results first, all of them, before anything new starts. And I built myself dashboards that show what ran, what is waiting for me and where I need to step in. A management dashboard for AI colleagues. The agents do not need it. I do.
If you sign budgets and your best people are the ones supervising the most tools, one thing is enough. Forward this edition to whoever runs people development for you, with one line: "I want us to work on this over months, not in an event. Find me something like it." What to look for: a programme that starts with the leadership team, runs over months rather than days, and treats human skills and AI as one subject. If you run people development and just received that line, or still have budget for this year: that is the work I do, since 2010 with organisations from start-ups to adidas and Fraunhofer (references). Thirty minutes, and you'll leave with a clear next step whether or not you work with me: book a call.
The colleague who is always on
A human employee works their contracted hours. An agent works whenever you type, at eleven at night too, and nobody stops you. The limit has moved from the colleague to you.
The research on being reachable is clear enough. Larissa Barber, then at Northern Illinois University, and Alecia Santuzzi named the urge to answer work messages immediately "telepressure" and tied it to worse sleep, burnout and more sick days; their data point to a workplace problem rather than a worker problem. The same applies to an agent that will always answer. The hopeful note is a working paper: Microsoft and Harvard researchers gave AI tools to 7,000 knowledge workers for six months; users saved about two hours a week on email and worked less outside working hours, not more. The tools do not make the day longer. People do.
The European Parliament asked for a right to disconnect in 2021; there is still no directive. Portugal, where I live, went ahead in 2022: employers have a legal duty not to contact staff during rest periods, except in cases of force majeure. Neither mentions agents. They do not need to. My own rule came from being tired: I do start agents in the evening, and I read nothing of theirs before the next morning. The work runs at night. The reading does not.
Where the ideas come back from
The part that is easy to forget when the agents are fast: thinking has a pace, and it is slower than theirs. Sabine Sonnentag at the University of Mannheim found with Charlotte Fritz that the heavier the workload, the less people manage to detach from it after hours, and the less they detach, the worse the strain. A meta-analysis of 26,600 people confirms that detachment, actually stopping, is more strongly linked to less fatigue than relaxing or controlling your evening. And the ideas need the gap too. Ut Na Sio and Thomas Ormerod at Lancaster University combined 117 studies on incubation, putting a problem aside: it helps, and for verbal puzzles most when the break held something light rather than rest; a demanding task weakened it. Marc Berman at the University of Michigan showed the walk matters: after fifty minutes in a park, working memory improved by about a fifth; after the same walk through city traffic, barely at all.
So I do what the tools do not invite. I walk, I go to the sea, I make music, I go to the gym even when my head insists I should be working, and there the best ideas come. And every day after lunch, twenty minutes of nap or meditation. That is where the new use cases arrive: a process to rebuild, a more flexible website, landing pages that join campaigns by themselves. None of it comes at the screen.
Ideas from outside come too, now and then on LinkedIn, but the limit is rarely the idea. It is the toolset: team size, servers, data protection, what a company allows. I run a small company and try everything the law allows, so mostly I am the one showing clients what is possible: in marketing and sales, in organising their day, in using AI as a coaching partner for becoming a better manager of others and of themselves. That is how it spreads inside a company too. When Erik Brynjolfsson at Stanford and colleagues gave 5,000 customer-support agents an AI assistant, productivity rose 15 percent on average and about 30 percent for the least experienced; the authors argue the tool captured what the best agents did and spread it. What spreads is not the licence. It is what the person next door does with it.
The McKinsey authors' five principles for "brain-powered organisations" start with calibrating the load; the second asks leaders to "subtract before they add". That is the line I would underline for whoever signs. Remove the low-value work before you add the agents, or the agents become its new owner, with a human reading everything they produce.
Your turn
If you sign budgets. Decide two things nobody else will: how many agents and tools one person runs before someone asks whether that is still work, and that reading an agent's result after hours is a choice, not an expectation. Then send this edition to whoever runs people development, with the line from the middle of this page.
If you lead a team. A dashboard that shows what the agents ran, what is waiting for a human and where to step in; for the team, or one you build for yourself. Where the work allows it, try collecting results in fixed windows rather than as they arrive, and notifications off by default.
If you develop people. Put recovery into the AI programme, on paper: a protected hour without AI, a walk that counts as work. And teach reading: skim to decide, read slowly what leaves the house. If you need the mandate, forward this upward; the opening is written for the person who signs.
Your 1% this week, whoever you are. After your next long AI session, twenty minutes without a screen. A walk if you can, a nap if you dare. Write down what came back.
Until then: how many conversations do you have open right now, and which of them is waiting for you? Tell me in the comments. I read every reply.
If you want to bring these themes into your organisation, book a 30-minute call. You can also explore the Leadership Essentials Workshop Series for teams.