Governance Is Not a Dashboard
What AI makes cheap is coordination. What it makes tempting is control.
The room had seven chairs. White fibreglass, swivel-mounted, arranged in a hexagon, each with a panel of buttons set into the armrest instead of a keyboard, because Stafford Beer had decided that executives would not type. Screens covered the walls. Photographs of it still circulate today, usually captioned as something between a Bond villain's lair and the bridge of the Enterprise.
It was the operations room of Project Cybersyn, finished in Santiago in early 1973, and built to help Salvador Allende's government run a nationalised economy in something close to real time. In every photograph it looks exactly like a command centre: one room, total visibility, a handful of people at the top seeing everything at once.
That reading is wrong, and the room is partly to blame for it. The crisis Cybersyn is best remembered for surviving, the lorry owners' strike of October 1972 that was meant to strangle the country's supply lines, happened before the room existed. The work was done over a telex network strung between factories and depots, shifting information sideways so the people nearest each blockage could route around it. No screens. Nothing worth photographing.
That is the difficulty with Beer's work, then and now. What survives is the picture of the control room. What mattered was the part that had no picture.
Only variety can absorb variety
The theory underneath it was the Viable System Model: a design for an organisation that could respond to reality without collapsing into either chaos or dictatorship. The model held up. Almost nobody could afford to build it properly.
Beer borrowed a line from Ross Ashby: only variety can absorb variety. A system facing a hundred different situations needs a hundred different responses on hand, or it filters the world down until fewer responses will do. Ashby's own formulation was that the amount of variety in a system must match the variety in its environment to keep control, and organisations that pursue predictability instead tend to filter out the mavericks and the disruptive ideas along with the noise, which leaves them with nothing to draw on when the environment actually shifts.
Most organisations do the second thing. Not because anyone decides to. Because holding real variety takes people: people sensing the edges, people coordinating between departments, people carrying the whole picture without flattening it into a dashboard. That has always been expensive, so the flattening usually wins by default.
There's a second reason it wins, and it's less often said out loud: the model itself is hard to see. The diagram most people meet first is a maze of nested ellipses, algedonic channels and looping feedback lines.
It's accurate. It's also close to unreadable on first contact, the kind of picture that explains everything to someone who already understands it and nothing to anyone else. Show that to a room full of people trying to run a charity or a co-operative and watch how fast the conversation drifts towards "can we just put the important bits on one screen."
That drift is not a coincidence. A model nobody can hold in their head gets replaced by whatever fits in a dashboard, and a dashboard is, structurally, a command-and-control instrument. People don't set out to strip Beer's work down to its worst version. They reach for the simplified one because the real one was never made legible, and the simplified version happens to concentrate power at the top.
That's the part that's changing, on both counts.
Five systems, one constraint
VSM describes any organisation as five interacting parts.
- System 1 does the actual work: the caseworkers, the classroom, the shop floor.
- System 2 stops those units colliding with each other.
- System 3 holds operations coherent, here and now.
- System 4 looks outward, watching for what's about to matter.
- System 5 holds the question of what the organisation is actually for.
Each part has always run into the same wall. System 1 can only absorb as much complexity as its people and processes allow, so organisations write procedures and shrink what counts as "normal" until the work fits. System 3 has never had the visibility to coordinate more than a handful of units without hiring expensive middle managers to do it by hand, and when it takes on too much of that job directly, the model has a name for the result: hypertrophy of System 3, where the coordinating function swells to cover work that belongs with the units themselves.
System 4 gets cut first when budgets tighten, because scanning the horizon produces nothing you can point to next quarter, and a System 4 that goes missing entirely gets called, in the same taxonomy, a headless chicken: a system with no way of sensing what's coming. System 5 ends up as a handful of people making calls on everyone's behalf, not because the model asks for that, but because it was the only way most organisations found to keep decisions moving at all.
Command-and-control wasn't a choice most organisations made. It was what happened when absorbing variety at the edges cost more than destroying it at the top, and when the one model built to prevent that was too dense to use as a working tool.
If a picture is needed at all, a simpler one does more work: plot an organisation on two axes, power concentrated against power distributed, aligned against misaligned with its own purpose.
Where an organisation sits on that grid tells you more about whether it will survive contact with reality than the full recursive model ever will, not because the full model is wrong, but because a model nobody can act on doesn't count as held knowledge.
What actually changes with AI
Everything above describes a constraint that has held since the 1970s: real variety needs people to carry it, and the model for organising that was too complex for most organisations to use. AI doesn't remove the constraint. It changes who can afford to meet it, and it does so at every level Beer described.
A caseworker whose AI tool can draft, summarise, and cross-reference in seconds can handle more of their own casework's variety without escalating every unusual case upward. That's more capacity sitting at System 1, not less. A coordination function that used to need a room full of people tracking dependencies can now do it with two people and a model, in something closer to real time: System 2's job, done properly, in organisations that could never previously afford to do it at all. Environmental scanning, always the first line item cut, gets cheap enough to sustain without a research department.
None of that requires centralising anything. Worth saying outright, because it's the part almost every commercial AI rollout gets backwards.
We're not the only ones arguing this. One recent piece on management cybernetics makes the case that AI turns every user into a manager of a digital actor, and that the only workable response is to push control down to wherever the work and the information already sit, not up towards a central authority. Another, aimed at agentic AI specifically, splits the design question into levers of freedom, sitting with Systems 1 and 4 where agents explore and sense, and levers of constraint, sitting with Systems 2 and 3 where guardrails and telemetry keep that exploration honest. Different vocabulary, same shape as the argument here.
The command-and-control trap
The easier move, and the one organisations keep reaching for, is the same one the diagram already nudges them towards: point the new capability upward instead of outward. Feed the model's output into one dashboard, hand System 5 the summary, and turn System 3, without anyone announcing it, from "coordinating autonomous units" into "watching them".
We've seen this pattern often enough with clients to name it without flinching: it rarely starts as anything sinister. Someone asks whether the AI tool can flag when people drift from the guidelines. A couple of meetings later, flag has become mandatory, and mandatory has become a report on who didn't comply.
That's variety being destroyed, not absorbed. Fewer exceptions reach a human, which looks efficient, right up until the organisation notices it has become more brittle rather than more capable. The complexity in its environment hasn't gone anywhere. It's just been hidden from the people who were actually positioned to respond to it.
This is the same pattern we described in Decisions Have No Autopilot: AI doesn't replace an organisation's decision-making culture, it turns the volume up on whichever one is already there. Point it at control, and control is what gets amplified.
It's not a hypothetical. A recent talk to systems practitioners on AI in healthcare organisations recommends exactly that direction: a coherent, organisation-wide AI strategy driven top-down through the model, aligning deployment with the institution's strengths and weaknesses from above.
The speaker knows VSM well. That's rather the point. Even people who understand the model in detail reach for the top-down reading first, because the diagram nudges you there and the org chart already agrees with it.
Coordination without surveillance
RACK, in active development, exists because of exactly this pattern, and because a shared practice, like a shared model, only holds up if the people using it can actually see what it's doing. The idea is to let a person or a team write down how they work, once, and turn that into instructions their AI tools actually follow.
What matters more than the mechanism is what it refuses to do, and why: a required instruction has to come with a written reason and can't be overridden downstream, because System 2 coordination only works if everyone can see where a rule came from. A default is a shared convention, adaptable by whoever is nearest the work, no permission needed, because the person doing the work is still better positioned to judge it than anyone watching from above. A personal instruction never leaves that person's machine.
If a shared rule overrides someone's own instruction, RACK says so immediately rather than letting them discover it later in the output. If someone adapts a shared default for themselves, that adaptation is never transmitted, logged, or reported upward. There's no dashboard showing a manager who changed their own defaults, and building one would defeat the entire point of the tool.
That's the difference between a System 2 that helps units coordinate and a System 2 that watches them.
Who this is for
The organisations most likely to benefit aren't the ones with the biggest AI budgets. They're the charities, co-operatives, and small non-profits that have always had to do more with less, and could never stretch to a real System 4 function or a proper System 2 layer.
Many of them also have values that rule out the command-and-control route on principle: they exist to distribute power, not concentrate it. Pointing new capability upward would betray the reason the organisation exists.
There's a sovereignty question sitting underneath this too. Renting organisational judgement from a handful of large AI vendors, most of them sitting in jurisdictions with their own regulatory agendas, just recreates the top-down problem at a different layer of the stack. An organisation can decentralise its own decision-making and still hand the underlying capability to somebody else entirely.
We made this case at length in The Sovereignty Stack Has No Bottom: sovereignty isn't a destination, it's a ratio, and every layer where it gets declared solved conceals another layer where it wasn't. The same ratio applies here.
RACK runs entirely on the machine it's installed on, no account, no model to connect to. Running a model locally, or adopting a tool built this way, doesn't make an organisation fully autonomous. It shifts the ratio. That shift is worth making anyway.
It's the same thinking behind our first experiment, testing whether a small group can run AI for each other on hardware they already own, with a public record of who did what for whom, instead of routing everything through a vendor's servers. Distributed capacity, held in common, with no single point that has to be trusted.
What this doesn't solve
AI doesn't fix a poor decision-making culture. It amplifies whatever's already there. An organisation with no shared way of deciding gets louder, not clearer, once AI enters the picture.
Everything argued here holds only for organisations already trying to spread judgement outward rather than pull it inward. For everyone else, more capable tools just produce faster, better-dressed versions of the same old top-down calls.
What's easier now is the cost of building requisite variety into Systems 1 through 4: sensing, coordinating, and scanning, all cheaper than they've ever been. What's also easier, for the first time, is making the model itself legible enough that people can tell the difference between using it and gutting it.
What isn't easier, and never will be, is the governance choice sitting on top of both of those: whether to use the capacity to spread power out, or to pull it in.
If your organisation is working through this, get in touch, or try RACK for yourself, still in active development but built to be used now.