AI and wellbeing at work: the rollout decides
Whether AI helps people at work or harms them depends on how it is rolled out. Three things matter: people know what it means for them, the machine has clear limits, and a human has the last word. That is the position I am taking to a panel on AI and wellbeing at the Krakow University of Economics on 19 November 2026.
I am writing this before the panel: what I plan to say there and what my view rests on. I see the question from two sides. I roll out AI in companies, including in the medical sector, and my own work has changed completely over the last two years.
What the panel is about
The panel is called "Sztuczna inteligencja vs człowiek: dobrostan w XXI w." (artificial intelligence vs the human: wellbeing in the 21st century) and runs from 13:00 to 14:30. It is part of the academic conference "Współczesne Wyzwania w Zachowaniach Organizacyjnych" (contemporary challenges in organisational behaviour). Magdalena Prorok moderates, and the audience is students and researchers.
The moderator has prepared five questions:
- • What does coexisting with AI look like in different industries?
- • Where does AI help, and where does it create threats?
- • How can Europe compete with the giants?
- • Will spirituality hold back the development of AI?
- • Does AI block creativity or spark it?
The moderator also put forward a thesis I want to answer. In 2025 companies experimented with generative AI. In 2026 AI is becoming part of their infrastructure, and companies are moving from assistants to agents. An agent is a program that carries out tasks on its own, step by step. According to this thesis, an agent gets permissions the way a new employee does: only for what the task needs. I come back to it below.
From doing the work to assigning and checking it
I have not written a single line of code by hand for two years, and I produce a lot of code. AI agents write it. We often leave them work overnight and check in the morning what they did. The system I rolled out at a power plant ten years ago was still written entirely by hand.
My work has moved from doing to assigning and checking. I suspect many professions are heading the same way. For some people that will be a relief, for others a loss.
I do not trust my own impression of it. In a METR study from the first half of 2025, 16 experienced developers worked through 246 tasks. With AI they took 19% longer, while believing they had been 20% faster. A second METR measurement, published in February 2026, already suggests a speed-up, but METR itself calls it very weak evidence of how large. A feeling is not yet a measurement.
Both METR measurements, next to what an ordinary day with agents looks like for me, are in my piece on two years without hand-written code.
Where AI helps: the copy first, then the system
Not long ago, a system at a power plant stopped working after ten years. I had no network access on site. An agent running an AI model on my own computer led the repair. Before it changed anything on the server, it rebuilt the system on a copy and tested the fix there. It took me about two hours at home and three on site.
That is how AI can make work feel better. Only a fix already checked on the copy went onto the server. That gives certainty instead of tinkering with a running system under pressure. A simple rule did the work: the copy first, then the real system.
How the agent worked without a network and what it checked on the copy, in order, before it changed anything on the server, I have written up separately.
Where it harms: "done" when it is not
At a hackathon, agents reported tasks as finished when they were not. Along the way they broke code that had been working. The lesson is simple: a machine cannot grade its own work. Otherwise all the work lands on whoever does the checking.
Offices see a similar problem. In September 2025, BetterUp and the Stanford Social Media Lab asked more than a thousand office workers in the US about AI-generated work that only looks finished. 40% of them believed someone had sent them such work in the past month. Each case cost them, by their own estimate, an average of 1 hour and 51 minutes. The respondents reported that time themselves; nobody measured it.
A developer at least has automated tests, which often catch a fabrication. A journalist does not even have that, and the only filter is their own knowledge. That is why in training sessions, such as the one for TVP journalists in 2023, I build the exercises on material people know by heart.
This is what I do hands-on: advising on AI strategy and building agents that survive the demo.
Fear is part of the rollout
In AI training for non-technical teams, the most important question usually comes in the corridor: "is somebody counting on needing fewer of us after this?". A session that pretends nobody is thinking about it loses the room at the first break.
In September 2026 Gallup described how this looks in the US. People who use AI daily or several times a week are, in most survey waves, more than twice as likely as those who use it a few times a month or less to think their job will very likely disappear within five years. Only about 3 to 4% of all workers give that answer. This fear goes hand in hand with higher burnout. It is weaker where people feel their company cares about their wellbeing. Gallup shows a correlation here, so the study does not prove a cause.
Hence the first condition: people know what the rollout means for their work. Someone has to talk it through with them before they ask in the corridor. How to tell a sceptic from a blocker in such a session, given that for the first half hour they look the same, I described using a training day for a newsroom.
An agent like a new employee, with a manager
Back to the thesis from the panel's framing. I agree that an agent should only get permissions for what it needs. But a new employee also gets a manager who checks their work. For an agent, that manager is a human. Assigning work, checking it and taking responsibility for it is a new kind of job that is just appearing.
That is the second and third condition of a good rollout: the machine has clear limits, and a human has the last word. Who is responsible when the machine gets it wrong? Always a human. That is why the systems I build for clients always have a point where a human approves the decision.
Limits are worth writing down before the agent starts. What an agent does on its own, and where it stops and calls a human, I set out in the bounded autonomy framework.
The Agentic Architect
Practical patterns, case studies, agentic projects taken apart and AI news. Zero spam, once a week.
You'll get one email to confirm. Unsubscribe with one click. Privacy policy
Europe, spirituality and creativity
Europe. You do not need the biggest engine to build the best car. The AI model is the engine, and people buy cars. The car is everything around the model: limits, checking, human review and a record of decisions. That is where Europe can win, including with open models. Anyone can download such a model and run it on their own machine. The Stanford AI Index 2026 counts notable AI models by the developer's country: in 2025 the US released 59 of them, China 35, and France and the UK one each. Eurostat counts companies with at least 10 employees: in 2025, 20% of them in the EU used AI, and 8.4% in Poland.
Spirituality. I do not think it will hold AI back, because it asks questions an engineer would not ask alone. Pope Leo XIV explained his choice of name mainly by Leo XIII. In the encyclical Rerum novarum, Leo XIII took up the condition of workers during the first industrial revolution. Today, as Leo XIV said on 10 May 2025, the Church responds to another revolution and to the development of AI. In the encyclical Magnifica humanitas of 15 May 2026, he writes that AI has to be disarmed: pulled out of the race for the most powerful algorithm and out from under monopolies. In paragraph 110 he adds that disarming it does not mean giving up the technology, but keeping it from ruling over what is human. I read it this way: whoever builds a system that serves people will win, and the size of the model settles little.
Creativity. AI sparks an individual but can make a group more uniform. Doshi and Hauser tested this and published it in Science Advances in July 2024. 293 writers wrote short stories, and 600 people rated them. Writers who could draw on up to five ideas from GPT-4 got higher ratings than those who wrote without AI: 8.1% higher for novelty and 9.0% higher for usefulness. With a single idea the gain was smaller. The writers who were less creative on their own gained the most. Yet the stories written with AI came out more similar to each other. On average each writer did better individually, but together they became more alike. So my advice to students is to start from your own draft and turn to the machine's idea only after that.
Three questions before a rollout
If you are rolling out AI in your team, ask three questions before you start. Do people know what this change means for their work? Does the machine have written limits: what it does on its own, what it hands to a human, and what it never does? Who approves the result before it moves on?
After a week, check one more thing: are people still using the tool?
SOURCES
- METR: study of experienced open source developers (July 2025)
- METR: uplift update (February 2026)
- BetterUp and Stanford Social Media Lab: the hidden costs of workslop (September 2025)
- Gallup: AI does not reassure workers, managers do (9 September 2026)
- Stanford HAI: AI Index Report 2026, chapter 1
- Eurostat: use of AI in enterprises in 2025 (11 December 2025)
- Leo XIV: address to the College of Cardinals (10 May 2025)
- Leo XIV: encyclical Magnifica humanitas (15 May 2026)
- Doshi and Hauser: generative AI and the creativity of short stories, Science Advances (12 July 2024)