A continuation of “The Five Dimensions of Human Agency: An Introduction for the Rest of Us.”
The previous piece ended on a question rather than an answer. Once you see agency as five dimensions — Capacity, Room, Will, Direction, and Identity — the interesting question about AI stops being “how much agency do machines take from us?” and becomes:
Who gains agency through AI, who loses it, and where does it become concentrated?
That question is not new. Someone asked a version of it fifty years ago, long before the technology existed to make it urgent. His name was Stafford Beer, and the book is Designing Freedom.
The book where Beer said the quiet part
In 1973, Beer — a management cybernetician, the man behind Chile’s Cybersyn, an attempt to steer a national economy through a network of telex machines — gave five radio lectures for the CBC. No charts, no jargon, no worksheets. Just a sustained public argument about why our institutions fail, and what freedom would have to mean if we wanted to design for it.
His claim was uncomfortable then and is more so now. Freedom, he said, is not a property of technology and not a vague ideal. It is a property of organizational design. The same tools — in 1973 he meant the computer — can be assembled into what he called a liberty machine: a system that pushes decisions down to where the knowledge actually is, and uses its informational power to enlarge the room people have to act. Or the very same tools can be assembled into an apparatus of centralized control — in his words, the most efficient instrument of oppression ever devised.
The technology does not choose between these. Someone chooses. Freedom has to be designed — deliberately, against the institutional default, by whoever is holding the pen.
Why AI sits exactly on Beer’s fault line
Everything that made Beer’s warning theoretical in 1973 is now literal.
AI is, at its best, a coordination technology. It can keep the parts of an organization aware of one another without a manager in the middle relaying messages up and down a hierarchy. That is genuinely liberating: it lets a company widen autonomy, flatten its structure, and let the people closest to a problem actually decide.
But the identical capability, pointed the other way, becomes something else. The same system that can coordinate can also watch — every task, every decision, every deviation — and quietly pull the real choices back to the center. A network can concentrate control as thoroughly as a pyramid, and do it behind a friendlier interface.
This is the trap worth naming plainly: flatness is not freedom. An org chart with fewer layers tells you nothing about who actually holds the pen. You need a criterion that structure alone cannot give you.
The criterion: the five dimensions, applied
That criterion is agency, and this is where the five dimensions earn their keep. They turn a vague word into something you can inspect. Before and after any deployment, you can ask which dimensions actually moved, and for whom.
The most common pattern in real workplaces is easy to miss and easy to sell. AI raises Capacity — people can do far more than before — while cutting Room: they are watched more closely, with less ability to negotiate the target, the pace, or the rating. More gets done, and less can be questioned. That is not the spread of agency. It is its concentration, behind a more capable tool.
Two sentences are worth keeping on a card:
An explainable system without an effective right to depart from its recommendation remains a system of control.
Empowering people without giving them resources or authority merely turns autonomy into responsibility without power.
Both describe the same failure: a design that looks like it distributes agency while actually gathering it in.
And two of the five dimensions should almost never be handed to a machine at all. Direction — toward what are we acting? — and Identity — who are we, and what would we never trade away? These are the judgments that decide what an organization is for. Automating the drafting beneath them is fine. Handing over the judgments themselves is not efficiency; it is abdication.
From a good question to a real check
A criterion you cannot apply is only a sermon. So the last move is from diagnosis to design.
The practical form is an Agency Impact Assessment — a companion to the risk and compliance checks that any serious AI adoption already runs. It is deliberately modest. Before you deploy, and repeatedly after, ask:
- Whose Capacity increases, and whose decreases?
- Whose Room to act expands, and whose contracts?
- Does this strengthen initiative and ownership, or reward passivity? (Will)
- Who sets the Direction — and would anyone notice if the system’s goals quietly became ours?
- What does the change do to people’s Identity and sense of authorship?
- How are these effects distributed — and can we reverse the ones we get wrong?
None of this produces an automatic score, and that is the point. Agency is not a number that settles the argument; it makes the argument visible. It shows you which human possibilities a design is creating, constraining, transferring, or quietly abandoning — while you can still change your mind.
The invitation is the same
Beer ended Designing Freedom with a warning and an invitation. The warning: that we were sleepwalking toward institutions no one would have chosen, produced reliably by structures we never examined. The invitation: that the alternative was available to anyone willing to design for it.
Half a century later, the machine on the fault line is a far more powerful one, and the invitation has not changed. AI will help you build organizations that are faster, flatter, and more capable. Whether what you build is a liberty machine or its opposite is not a question the technology answers. It is a question the designer answers — and the honest first move, now as in 1973, is to admit that the question is being answered at all.
Keep the pen in a human hand. Not out of sentiment, but because handing it over is handing over your capacity to remain yourself.
This continues The Five Dimensions of Human Agency: An Introduction for the Rest of Us. The five-dimensional model and the Agency Impact Assessment are set out more fully in the working paper Human Agency in the Age of Artificial Intelligence (2026) (attached there, look at the end) ; the organizational argument draws on Stafford Beer’s Designing Freedom (CBC Massey Lectures, 1973).