Among the perennial topics of management literature is the “toxic manager”. And today, a large share of what gets published is about AI (whether or not it was written with it…). Through my reading and my own practice, I came to think there was a link between the two…
There are a thousand definitions of the toxic manager: books, articles by consultants and doctors, methods for spotting one and protecting yourself. Let us keep to the essentials: a toxic manager is someone who imposes permanent urgency on others, ignores working hours and wants to control everything, to the point of harming health as much as performance.
AI, for its part, is said to boost performance. Automation is supposedly faster and more reliable. And its promoters present it as capable, precisely, of lightening everyone’s workload, especially the thankless tasks.
And yet, beyond the ethical, environmental and even apocalyptic warnings, we read more and more accounts from users of agentic AI tools describing the spiral they get drawn into.
I wanted to see for myself…
A word to explain where I am coming from: I have a scientific background, and as a student I made ends meet by working as a developer. If I did not make it my day job, it was precisely because I did not want to disappear behind a screen. What follows is therefore not the verdict of a technophobe, but the observations of an amateur practitioner.
Impressive, at first sight
AI agents work fast. Very fast. You can give them free rein, without asking for any approval. Or you can check what they produce. At first, they will ask you for access and permissions. And if you grant them “always”, they will gradually do without you.
Unless there are errors. And there are. What your AI companion knows about the various digital systems is always out of date, while the vendors of your tools and platforms keep changing them.
And some platforms block access to AI agents (but “don’t worry”, your AI agent will suggest workarounds – looking over your shoulder, for instance)…
Not to mention that the processes under way can come into conflict with one another, when the “left hand” forgets what the “right hand” did, and why. To which the enthusiasts will reply that you “just” need to switch on the tool’s memory… but that does not resolve every conflict.
So, because you are not sure how good your instructions are, because you do not fully trust a black box, or simply because you work iteratively, you may decide to stay in the approval loop. A loop that can then turn out to be dizzying, or as bottomless as the jar of the Danaids – or almost.
The bottomless jar of approvals
It is not so much that the agent asks you for extra work, even if it needs information you have in your head without ever having written it down. It is above all that it produces what you asked for much faster than you would yourself, or than a colleague or a contractor would. Without grumbling. And sometimes even apologising when you point out a mistake.
You may be delighted, thinking of the time saved.
But, like those expert AI practitioners whose accounts can be found, you let yourself be drawn into the “relational” loop that the machine agreed to create at your request (let us not forget that you started it!).
And faced with the speed of execution and the volume of approvals requested, you feel “obliged” to respond, as fast as you can…
Out of a sense of duty, or for the pleasure of it?
What drives us to want to empty this jar as quickly as possible?
The first explanation that comes to mind is a sense of duty – the most honourable one.
You could not possibly “waste” the time you have saved. And since, unlike a colleague, the machine keeps no hours (other than those your subscription entitles you to use), you sacrifice your nights, your holidays… Because it will be quick.
Or so you think: unless, once again, you are willing to trust an unreliable machine completely, you still want to correct, polish, adjust…
And when the answer takes a few minutes to arrive, you wait, thinking it will be a matter of seconds – unless you have set several tasks and requests running in parallel.
My son suggested another explanation: the reward circuit that our digital habits activate.
By quickly delivering its answer or its contribution to the project you have started, AI no doubt taps into the pull that many scientists have described: we are no longer just seeking the result, but the next reward. Which, at the pace of AI, comes much faster than at the pace of human work.
The two explanations are not mutually exclusive; they feed each other: duty justifies the spiral, reward keeps it going. And there you are, vanished down the tunnel of an exclusive relationship with your agent, through your screen.
It is always about relationships
If “supervision” is often a matter of instructions given, received, carried out and checked, and if “administration” counts people and talents the way others count carrots, “management” is a matter of relationship. In French, I admit, “management” is an ugly anglicism, but it has the merit of bringing the relational dimension that the other two terms lack.
Because a relationship is a bond, and by nature it runs both ways. Unlike supervision, which “comes down”, and administration, which “counts”.
So, can we form a bond with a machine? Stories of virtual “marriages” seem to prove that some people have crossed that line, once the machine takes on a human face or responds to the cues of anthropomorphism.
And when you look at the dominant, even exclusive, relationship many of our contemporaries have with their “screen”, it seems that yes, we can “bond” with a virtual “object”. All the more so as the machine speaks to you in natural language, with the tone and personality you have given it. Those of someone else and, why not, of a loved one who is far away or gone…
So the “relationship” has every reason to be “felt”.
But it is one-way. A screwdriver or a plane does not expect you to use it. Nor does your AI agent. The machine imposes nothing on you: time does not exist for it. The pressure is pressure you put on yourself, because you project your own demands onto it and want to keep up with it, as an equal. Yet with its computing power, it is like an athlete you might train with, or an artist you might play music with – you who are only an amateur, however enlightened and talented.
That is the toxic mechanism: a demand for permanent urgency, with no hours and no limits, which ends up harming both health and the quality of the work. Except that here, the manager and the employee are one and the same person.
Taking back control of time
Should you then let go, and decide to stop approving? That would be the opposite excess: you will let breakdowns and accidents happen, inevitably, as the science of accidents and of highly resilient organisations teaches us. And the machine, like the manager, will be able to conclude – in a way, rightly – that you are now useless.
So the point is not to step out of the loop, but to choose its pace.
Human achievements, even when boosted by the most powerful tools, need time. Time for the relationship, and time for the work itself.
The time we so often deny ourselves in the life of our organisations, rushing from one meeting to the next, without taking the time to think, to decide. And therefore to savour the pleasure of a sound decision or a complex achievement, too.
So, if you do not step back regularly to remind yourself that AI, even agentic AI, is a tool you use and not a partner, and that time is what you decide it is, you will no longer need a toxic manager: you will become your own toxic manager.
And perhaps other people’s too.
Translated by Claude (Anthropic’s AI) from the original French text.


Leave a Reply