Is AI really degrading work? Reading Carbonell's « Un taylorisme augmenté »
I recently came across a review by François Jarrige, published in La Vie des idées, of Juan Sebastián Carbonell’s book Un taylorisme augmenté. Critique de l’intelligence artificielle (Éditions Amsterdam, 2025). Direct title, uncomfortable thesis — especially when you’re, like me, a developer who spends his days writing code with a generative AI in the same window.
I’ll be honest: this book made me think, and in part proved me right against my own will. Here’s why.
The thesis: AI doesn’t replace work, it fragments it
Carbonell dismantles a myth we’ve been hearing since the 90s — the idea that automation is bringing about the “end of work.” His thesis is more precise, and darker:
“AI is neither a tool for requalification nor an instrument of polarisation, but a tool for the degradation of work in the hands of companies, in the form of an augmented taylorism.”
— Un taylorisme augmenté, p. 72 (my translation)
In plain terms: we don’t replace people, we reduce their tasks to verification, cleanup, correcting machine errors. The translator becomes a proofreader of machine output. The journalist rewrites generated drafts. The oncologist validates an algorithm’s suggestion. Autonomy over the creative act, Carbonell writes, is lost.
He adds a line I find exactly right:
“The functioning of generative AI depends very heavily on human labour.”
— Un taylorisme augmenté, p. 111 (my translation)
The whole system runs on human work — annotations, corrections, validations — that has simply been made invisible.
Where Carbonell is right
I see this pattern with some prospects. Someone reaches out saying: “We want to put AI into our process, to move faster.” When I dig in, often the real ask is: “We want to produce more, with fewer people, with more control over the output.” That’s not an augmentation project. That’s a control project.
I think of a mid-sized company that contacted me a few months ago to “add AI” to their customer support. The initial ask seemed reasonable: an assistant to handle FAQs. But as we dug in, the real equation appeared — the two-person support team was deemed too expensive, and leadership wanted to keep only one, who would supervise the AI’s replies. Not an automation project. A role-elimination project dressed up as tech modernisation. And the person who would have remained would have spent their day rereading tone-deaf replies to frustrated customers, instead of actually talking to those customers. Degrading, yes. And over time, humanly expensive for everyone — including the company that ends up paying the bill in turnover and reputation.
I turned it down. Carbonell is right: that’s the natural slope the moment you introduce the tool without questioning the intent behind it.
Where I disagree
What bothers me in the book, as Jarrige summarises it, is the idea that the tool itself carries that logic. As if generative AI were by nature a tool of degradation.
My day-to-day experience says otherwise. I code most of my projects with Claude Code — an AI agent that reads, writes, and tests my code alongside me. I don’t feel deskilled. I feel more powerful: I can take on more ambitious solutions, explore paths I’d have set aside, and push a problem all the way through instead of stopping at the first acceptable answer — because the cost of exploration has dropped.
What changes isn’t that the machine does the work for me. It’s that I have to step up a level. I no longer type out loops by hand — I design the architecture. I’m not rereading syntax errors — I’m arbitrating design decisions. My craft has shifted. It hasn’t degraded.
Jarrige himself, in his review, criticises Carbonell for failing to investigate the concrete resistances, the alternative uses. That’s the missing piece. The tool isn’t neutral — but it isn’t one-directional either.
The real pivot: who deploys it, and for what
I think Carbonell is right on the facts and wrong on the cause.
What degrades work isn’t AI. It’s the project entrusted to it. A company looking to industrialise, cut costs, and meter every keystroke — it was already using spreadsheets, KPIs, and tracking tools long before AI. AI just hands it a new productivity gain to spend in the same direction.
Conversely, a freelancer or a small team trying to do better work will use the same tool to loosen their constraints, not to tighten them.
Taylorism isn’t Claude Code or ChatGPT. It’s a way of organising work, and it long predates AI. AI is just the latest note played in that score.
What that changes for me — and for Avalonis
I come out of this book with a conviction I already held, now a little sharper.
When I build an n8n workflow for a client, or automate part of their back-office, I ask myself the same question every time: is what I’m building going to make a painful task disappear, or is it going to turn a person into a machine-watcher? They aren’t the same. One liberates, the other embitters.
That’s also why Avalonis takes on few clients. Few — but each supported with the rigour of a solution designed for their craft, and an obsession with their satisfaction. Not a template rushed through the door. Industrialising bespoke work is a contradiction in terms. Forging is something else entirely: choosing the material, knowing the person who will use it, and leaving a trace of your hand in the finished object.
AI hasn’t turned me into a taylorist. It has made me more demanding about what I refuse to build.
Have an automation project and want it to augment your team rather than replace it? Let’s talk.