IV. The Machines Make It Possible, and Necessary

For a century, the strongest objection to any scheme of crediting value was practical: you cannot compute it. The information was not there, and even if it were, nobody could process it. That objection is dissolving in front of us, faster than almost anyone outside the technology industry realises.

The trace

Start with the simplest fact. Knowledge work now leaves a record.

Code lives in version control, where every change has an author, a timestamp, a description and a review. Decisions live in documents with complete edit histories. Tasks live in ticket systems that record who picked them up, who unblocked them, and when they closed. Conversations live in searchable threads. Customer outcomes live in analytics. None of this was designed for crediting value. All of it is the raw material for doing so.

Ten years ago, turning this exhaust into a credible account of who moved which goal would have required an army of analysts. Today a capable model can read an entire project’s history in minutes and produce a structured, cited account of contributions: what was done, by whom, and how it connects to the stated objectives. It will be imperfect. So is every performance review ever written. The difference is that the model’s account can be inspected, challenged, re-run and audited, which cannot be said of the manager’s recollection in a December calibration meeting.

Evaluation is easier than production

AI alignment researchers have converged on an insight that turns out to be the formal backbone of Axiocracy. Jan Leike and his colleagues built their approach to supervising powerful systems on a simple assumption: evaluating an outcome is easier than producing it. You may not be able to write a great paper, but you can tell a great paper from a poor one. You may not be able to design a bridge, but you can check whether it stands.

That asymmetry is what makes decomposition tractable. You do not need to know in advance how every piece of work should be done. You need to agree what success looks like, and then evaluate each contribution against it, recursively, through a hierarchy of evaluators, human and machine. The same idea that allows humans to supervise systems smarter than they are allows an organisation to credit contributions it could never have planned.

Constitutions for machines

The second convergence is even more direct. The leading approach to shaping AI behaviour, Constitutional AI, governs an automated evaluator with a short, written list of principles set in advance. The model’s outputs are judged against the constitution, not against the ad-hoc preferences of whoever happens to be reviewing them.

Then, in 2023, Anthropic and the Collective Intelligence Project went a step further. They asked a representative sample of about 1,000 Americans to write the constitution themselves, through an online deliberation platform. Participants cast more than 38,000 votes on over a thousand proposed principles. The resulting public constitution was used to train a model, which performed as well as the one trained on the company’s own constitution while showing less bias on several measures.

Read that again with Axiocracy in mind. A thousand people agreed, in advance and in public, on the principles an automated evaluator would apply. The evaluator applied them at scale. That is the axiocratic loop, already working, in the most advanced technology on earth.

The agentic firm

Now add the development that makes all of this urgent rather than merely possible.

Within a few years, a large share of the work in serious organisations will be done by software agents. Not chatbots answering questions, but agents executing multi-step tasks: writing and shipping code, drafting and filing documents, researching, negotiating, reconciling. They will work alongside people, hand work back and forth with them, and increasingly with each other.

An organisation like that has a problem that no organisation has faced before. It must continuously answer: which of these thousands of contributions, by people and by machines, actually moved our goals? Without an answer it cannot allocate budget between agents. It cannot tell which human judgement is worth keeping in the loop. It cannot decide which workflows to scale and which to shut down. It cannot pay anyone fairly.

Credit assignment stops being a question of fairness and becomes the operating system of the firm. Every company deploying agents at scale will build some version of it. The only question is what kind.

The two roads, again

This is why Axiocracy is not an academic exercise. The machinery of measuring value will be built regardless. It is already being built, inside the metrics dashboards of every platform and every agentic workflow product.

The default version will look like the systems we already know: objectives chosen by the owner, metrics nobody else can read, scores that follow you around, ratings by strangers with no appeal. Trebor Scholz, writing about the platform economy, put it bluntly: every Uber has its Unter. A badly designed credit-assignment engine is the Uber rating generalised to all of working life.

The axiocratic version uses the same machinery with different rules. Goals are agreed before the work. Measures are public and versioned, like law. Every credit decision has an audit trail, an explanation and an appeal. The constitution that governs the evaluators is written by those it governs.

There is a warning here from the alignment research itself. Pan and colleagues showed that more capable optimisers exploit badly specified objectives harder, and that the exploitation can arrive suddenly, in “phase transitions” that give little warning. Agents will game measures faster and more thoroughly than people ever did. That is not a reason to avoid measurement. It is a reason to design the measures in the open, with the people who will be measured, and to revise them constantly. Opaque metrics, gamed by capable agents and defended by nobody, are the worst of all worlds.

The machines have made rule by value created possible. They are about to make some form of rule by measured value inevitable. The only open question is whether it will be agreed or imposed.