Nobody voted on this: how to build a team that can still change its own mind

A team can get faster, flatter, and happier while quietly losing the ability to decide how it works. Here are some notes on how to tell, and what to do about it.

Watch a team adopt AI tooling and you will see something that looks like a win. Cycle time drops. The hierarchy flattens, because the junior person and the senior person now ship work that is harder to tell apart. Autonomy scores go up. Nobody is unhappy.

Then ask one question: when did you last change the tool, rather than change yourselves to fit the tool?

Most teams cannot answer. They have spent a year reorganizing around what the tooling makes easy, and the reorganization was so comfortable that nobody logged it as a decision. The norms shifted. The definition of a good pull request shifted. The cadence of review shifted. None of it was voted on, because none of it arrived as a proposal. It arrived as a default.

I want to argue that this is the central problem in organizational design right now, that neither of the two dominant framings can see it, and that the useful vocabulary for it was worked out sixty years ago in a psychiatric clinic in the Loir-et-Cher.

01. Two readings, both wrong

The vertical reading asks who approves the output. Human in the loop, sign-off authority, escalation paths, an accountable name at the bottom of the document. This is the governance conversation, and it is a real conversation. It is also entirely about positions in a hierarchy, which means it takes the hierarchy as the thing to be arranged.

The horizontal reading says the hierarchy is over. Everyone has a copilot. The leverage per person went up across the board, the org chart flattens, the distinction between senior and junior work compresses. This is the AI-native org post that gets written every week.

The two look opposed. They are not. Both take the arrangement as given and argue about where people sit inside it. Neither asks where the arrangement came from, or whether the team can reach it.

That question has a name.

02. A psychiatric clinic in the Loir-et-Cher

La Borde opened in 1953 at Cour-Cheverny, founded by the psychiatrist Jean Oury. Félix Guattari arrived in 1955 and stayed for the rest of his life. Most people know Guattari, if at all, as the co-author of A Thousand Plateaus, which is a difficult book. The clinic work is not difficult, and it is where his one genuinely portable idea comes from.

Oury and Guattari were treating an institution as much as they were treating patients. Their diagnosis had two failure modes.

The vertical failure is the pyramid: information up, orders down, everyone certain of their level and blind to every other one. The horizontal failure is the ward where everyone is nominally equal and nobody talks. Flatness without contact. Guattari's claim is that these are not opposites. They are the same failure wearing different clothes, because in both cases the group receives its structure rather than making it.

Transversality is his name for a third axis that cuts across both. It is not a value between vertical and horizontal. It is a measure of how much a group can encounter itself and rearrange its own arrangement.

At La Borde this was material. The clinic ran la grille, a rotating rota of tasks: doctors did laundry, nurses ran sessions, patients worked the kitchen. This was not therapeutic play-acting or a flat-hierarchy gesture. The point was mechanical. Once the role rotates, the role stops being identical to the person occupying it, and the group becomes able to see the structure it is standing in. A structure you can see is a structure you can argue with.

03. The coefficient

Guattari's image for this is odd and good. Picture horses in a field wearing adjustable blinkers. Blinkers shut tight and they collide blindly, each one working from a view that excludes everyone else. Open them and the field becomes navigable.

The coefficient of transversality is how far open the blinkers are: how much of the whole any position in the group can see.

Two things matter about this. First, it is a dial rather than a switch. No group is transparent to itself, and full transparency is not the goal. Second, and this is the whole idea, the dial is adjustable by the group. A group with a high coefficient is one that can reach its own settings.

04. Subject-groups and subjugated groups

The distinction Guattari draws from this is the one you can actually use on a Monday.

A subject-group produces its own enunciation. It can state its purpose in its own words, and it can modify its own structure.

A subjugated group receives its law from elsewhere. It hears its own purpose described by someone who is not in the room. It can be busy, effective, and pleasant to work in. It cannot change its own terms, and usually it has stopped noticing that this is a thing groups do.

Same people. Same org chart. Same headcount. Different coefficient.

This is why the two readings in section 01 miss. A subjugated group can be perfectly flat. It can be high-autonomy and high-morale and shipping fast, because subjugation is not about who reports to whom. A team whose norms, cadence, and working definition of good work are set by what the tooling makes easy is receiving its law from a vendor's defaults. Flat, autonomous-feeling, and legislated from outside.

The test: name three things you changed in the tool this quarter. Name three things you changed in yourselves to fit the tool. If the second list is longer, you know what you are.

That is not a vibes question. It has an answer, and the answer is in your commit history.

05. Where the defaults come from

It is worth being precise about the mechanism, because "the vendor sets your defaults" sounds like a complaint about procurement, and it is not.

Baudrillard, writing about opinion polling in 1978, made a structural observation that has aged strangely well. A poll does not measure what you think. It hands you a menu and records which item you point at. The menu was drawn before you arrived, and whatever you pick confirms that the menu was the right menu. He called this the referendum form: reality reorganized into A/B, and therefore computable, and therefore already latent in the question.

Look at what preference training literally is. Two completions, A and B, both sampled from the model. You pick one. The gradient flows back into the thing that generated both options. The preference is real. The space the preference ranges over was drawn upstream of the preferring.

I want to hold the line on how far this goes, because the strong version is wrong and will make your argument worse. Baudrillard wanted to say the loop is sealed and no information passes. It obviously does; models change because of that signal. What survives is the structural point, which is sharper anyway: the range of what can be preferred is set upstream of the preferring. Take the structure and leave the fatalism.

Now scale that up one level, from the model to the team. Your team is also being handed a menu. The skills file someone wrote in month one, the eval set nobody has looked at since, the default review cadence that came with the tool: these are the options the team will express preferences over, and the team will get very good at expressing preferences over them. The blinkers are not malicious. They are just narrow, and nobody remembers installing them.

06. The authorship worry is standing in the wrong place

There is a standing objection here that needs clearing, because it eats every conversation about AI and work.

The worry: if a model writes it, who is the author? Something has been lost. The work is no longer yours.

The objection to the objection is old, and it predates Guattari by decades. Bakhtin's line is that words arrive already inhabited by other people's voices. Every utterance is a reply to prior utterances and an anticipation of the next. You inherited the genre, the cadence, the moves, the vocabulary. You are answering arguments you half-remember reading. Deleuze and Guattari radicalize this into the collective assemblage of enunciation: individual enunciation never existed in the first place, and all discourse is indirect discourse.

So the anxiety quietly assumes you were the sole origin before the model showed up. You were not. Otherness in your work is not new. What changes is the distribution: which parts of the assemblage do which work, and where the credit lands.

That is not the same as saying there is nothing to worry about, and this is where most versions of this argument stop one move too early. Guattari's entire point is that assemblages differ, and that the differences are the object of study. Three things genuinely change:

• Speed. The other voices arrive faster than you can metabolize them. At some point a difference in rate is a difference in kind.
• Particularity. Bakhtin's borrowed voices belong to someone. A teacher you owe, a writer you are arguing with. A model's voices are a statistical aggregate, and you cannot answer back to a distribution.
• Erasure. The corpus had authors. They have been laundered out of the assemblage.

The worry does not evaporate. It relocates, from authenticity to composition and power, which is a much more interesting place for it to live and a much more tractable one. Authenticity is not an engineering problem. Composition is.

07. What to actually do

The obvious move here is to steal the grille: rotate people through positions and let the arrangement become visible. I think that is wrong, and the reason it is wrong is the useful part.

The grille worked at La Borde because the clinic had a problem you do not have. It was a permanent institution with fused role identity. A doctor had been a doctor for twenty years and would be one for twenty more, and the role had grown into the person. Rotation broke the fusion. A team of fifteen people building something does not have that problem. The roles are already fluid, and rotation for its own sake mostly costs you context.

So rotation was never the mechanism. The mechanism is the encounter: an assumption becomes visible when someone who did not make it reads it, and it becomes changeable when that person has standing to change it. Rotation is one way to force an encounter, and it was the way available in a psychiatric clinic in 1955. It is probably not yours.

Ownership: individual, with obligatory contestation

There is a real tension here that rotation dodges rather than resolves, and it is the vertical and horizontal failure replayed one level down.

Collective ownership of the configuration. Everyone can edit the skills files. In practice nobody does, because nobody is answerable for them. The config drifts, accretes, and becomes load-bearing by neglect. Within a year it is a set of decisions nobody remembers making. That is the horizontal failure exactly: flat, equal, no contact, and receiving its law from a document nobody is reading.

Individual ownership. One person owns the config, keeps it coherent, knows why every line is there. This works, and it produces the vertical failure. The team receives its arrangement from a colleague instead of from a vendor. Marginally better, because the colleague is in the room and can be argued with. Still legislation.

The way out is not to split the difference and not to pass the parcel. Keep ownership individual and make contestation obligatory. Someone owns the config, because coherence needs an owner and drift is the bigger enemy of the two. Someone who is not the owner has a standing job to argue with it, and that person changes.

The difference from rotation matters. The contestant does not take over. They read it with intent to disagree and produce an actual disagreement. That is a two-hour job rather than a quarter-long handoff, it does what rotation was supposed to do, and it costs no context. The owner keeps the coherence. The contestant keeps the coefficient.

The signal that it is working: the config changed, and not because the owner changed their mind.

The model boundary: you are permanently the reviewer

The version of this point that gets made usually says reviewers who never draft become rubber stamps, so rotate them. That describes a division of labor no engineering team has. Developers draft and review constantly, sometimes inside the same hour. Nobody spends three months only reviewing.

The asymmetry is somewhere else. Relative to your colleagues you do both. Relative to the model, you only ever judge. The model generates, you accept or reject. That position does not rotate, and it cannot, because there is nothing on the other side to swap with.

This collides with a distinction I have used before and did not push hard enough. Intelligence is rules-based work; judgement requires experience and taste built over years. The question that framing raises and does not answer is where the taste comes from once you stop doing the work. Judgement was never free-standing. It is a residue of having produced the thing badly a few hundred times.

The easy version of the worry is that judgement atrophies from disuse. That is the optimistic version. Judgement does not just drain. It gets refilled.

Section 05 said the range of what can be preferred is set upstream of the preferring. That was a claim about the option space. At the model boundary it goes one level deeper, into the criterion. After five hundred model outputs, your sense of what a good function looks like has model-shaped priors in it. You are still judging, and you are judging with a yardstick that the thing being judged helped manufacture. Meanwhile the loop closes at the level of the team rather than the weights: what you accept becomes the local definition of good work, and what you were offered to accept came from the model. Nobody has to be fine-tuning anything for this to be a closed circuit.

The terminal version is already shipping. The model drafts, a second model reviews the draft, and you approve the review. At that point you are not a reviewer. You are a notary.

The obvious objection is that this proves too much. Every tool shapes the taste of the person using it. Developers raised on one framework have framework-shaped priors, writers who read one author write like him, and nobody calls that a catastrophe. Fair, and the answer is the three differences from section 06. Speed: the priors arrive faster than you can examine them. Particularity: a framework has authors you can argue with, and a distribution does not. Erasure: you cannot trace which piece of your judgement you borrowed from where. The mechanism is old. The rate and the anonymity are not.

So the practice is not a rotation, it is a calibration. Some of the work gets done unassisted, on a cadence, not for the output and not out of principle. To keep the judging attached to something that is not itself model output.

This doubles as your control. You cannot say what the tooling changed if there is no unexposed baseline to compare against, and the argument about whether to bring the baseline back inside the tooling is itself the transversal move. It forces the team to say out loud what the tool is for.

Notice the adaptations

When the team changes a norm to fit the tooling, make a note of it somewhere you will come across it again. Nothing formal, no ceremony. These changes are invisible now not because anyone hid them but because each one is individually too small to register, and the aggregate is the entire arrangement.

Then occasionally set that list beside the list of changes you made to the tooling, and see which one is longer.

There is a version of this I am already committed to, and it is worth naming the bias. I have argued that the product design job in vertical AI is to encode domain expertise in flat, readable, auditable files rather than codebases: skills files, not code. That argument was about cost and speed. The org argument is better. A skills file is the artifact that makes the arrangement modifiable by the people standing in it. A codebase legislates. A markdown file can be argued with by the accountant, and an accountant who can argue with the configuration is in a subject-group. The reason to build Rig around a portable manifest is not developer ergonomics. It is the coefficient.

The test, again

The vertical question is who signs. The horizontal question is who is augmented. Both are answerable and both leave the important thing untouched.

The transversal question: can this group see and modify its own arrangement, or is the arrangement arriving from outside it?

You do not answer that with a survey, because a survey is a menu. You answer it by looking at what your team changed this quarter, and what changed your team.

---

Notes and reading

• Guattari, "La transversalité" (1964), collected in Psychanalyse et transversalité. The subject-group distinction runs through the same volume.
• Nicolas Philibert's documentary La moindre des choses (1996), the best way into the clinic itself.
• Bakhtin on inhabited speech, in The Dialogic Imagination.
• Baudrillard on the referendum form, in À l'ombre des majorités silencieuses (1978), which is short and worth the afternoon.