Most organisations can count the decisions with no human attached. The exposure is in the ones with a human who cannot change the outcome, and almost nobody counts those.
Ask a board what share of the company's consequential decisions are now made by machines and you will get an anecdote. Someone will name the systems. Someone else will say it depends what you mean by a decision. Nobody produces a number, and the reason is not that the number is hard to compute. It is that almost no organisation has agreed what would count, and the ones that have tend to count the wrong thing.
Every organisation deploying AI at scale can tell you how many systems it runs, how many calls those systems serve, and what they cost. Almost none can tell you what proportion of the decisions that actually affect a customer, a claimant, an applicant or an employee are now determined by one of them. That is a strange gap. It is the single most important fact about how the organisation now works, and it is missing from every dashboard.
The gap persists because the obvious way to measure it gives a comfortably low answer. Count the decisions with no human anywhere in the path, divide by the total, and most firms land somewhere small and reassuring. The systems that worry lawyers all have a person attached. There is a queue, there is a reviewer, there is a sign-off. On paper the humans are still deciding.
The useful frame here is older than any of this. In 1992 Michael Jensen and William Meckling described organisational design as an allocation problem: a decision right should sit where the specific knowledge needed to exercise it sits, and it has to be paired with a system that rewards and holds to account whoever holds it. Push the right away from the knowledge and you get bad decisions. Separate the right from the accountability and you get something worse, because now nobody is answerable for the outcome in a way that changes behaviour.
What a capable model does to an organisation is relocate specific knowledge. The pattern that used to live in an underwriter's experience or a caseworker's instinct now lives in a system that has seen a hundred thousand more cases than either of them. Decision rights follow knowledge, as they always have. The migration is not a governance failure. It is the predictable response to knowledge moving.
The rights migrate to the machine because the knowledge did. The accountability stays with the person whose name is on the form, because the form was designed before the migration.
That mismatch is the thing I have written about as accountability inversion, and I will not restate the argument here. What matters for this essay is a narrower point: if rights are moving, you should be able to say how far they have moved. Most organisations cannot, and the reason is that they are counting seats rather than outcomes.
Sort the consequential decisions in a workflow into three groups. The first is straightforward: fully automated, no person in the path, nobody claims otherwise. The third is also straightforward: a person considers the case, their judgment changes outcomes at a rate you can measure, and the machine is an input rather than an answer.
The second group is the one that matters and the one nobody reports. A person is in the path. They are named, they are employed, they click. And their presence does not change outcomes, because the queue is too fast, the inputs are not shown, refusing is career-expensive, or the model is good enough that reading stopped paying a long time ago. On the org chart this looks like tier three. In effect it is tier one.
A firm that reports only tier one is not lying. It is answering a question about staffing and presenting it as an answer about authority. The two were the same thing for most of industrial history, which is exactly why nobody noticed they had come apart.
This is not a philosophical distinction that regulators will get to eventually. They have started.
In December 2023 the Court of Justice of the European Union decided the SCHUFA case. A credit reference agency produced a score. A different company, a bank, made the actual lending decision with a human involved. SCHUFA argued it was performing a preparatory act, not making a decision. The court disagreed, holding that where the score is so heavily relied upon that a low one results in refusal in nearly all instances, producing the score is itself an automated individual decision within the meaning of Article 22 of the GDPR.
Read what that does. The court located the decision at the point where the outcome was actually determined, not at the point where a person signed. It did so on the basis of an empirical observation about how often the downstream human departed from the recommendation, which is to say: on the basis of something very close to an override rate. The legal test and the measurement question have converged, and most firms are not measuring the thing the test turns on.
So here is the number, defined so that the middle tier cannot hide in it.
Of the consequential decisions produced by a named workflow in a stated period, the share whose outcome was determined by an automated system. A decision counts as machine-determined when altering the system's output would have altered the outcome, and altering the reviewer's attention would not. Fully automated decisions and effectively unreviewed ones both count. Reported with the denominator, the period and the workflow.
The definition is counterfactual on purpose. It does not ask whether a human was present, which is the question that produces flattering answers. It asks whether the human was load-bearing. A reviewer who would have produced the same outcome asleep is not holding the decision right, whatever their job title says, and the organisation should know that before a court tells it.
The first term is easy and every firm already has it. The second is the work, and it needs two instruments that exist.
The Meaningful Override Rate gives you the rate at which reviewers changed outcomes they were positioned to change. A rate at or near zero across a large sample tells you the reviewers are not moving outcomes. It does not yet tell you why, because a very good model produces the same reading as an unread queue.
That is what the seeded error probe resolves. Plant decisions whose correct answer you already know, indistinguishably from real work, and count how many come back. A high catch rate beside a near-zero override rate means the model is strong and the reviewer is awake, and those decisions belong in tier three. A catch rate near chance means the queue is unread, and every decision in it belongs in the numerator whether or not somebody clicked.
Those two numbers together turn the middle tier from a matter of opinion into a matter of measurement. Without them, any claim about how much authority an organisation still holds is an assertion about its org chart.
One property of this measure matters more than its absolute value, and it is the reason to compute it repeatedly rather than once.
Machine decision share drifts upward without any decision to increase it. A model gets better, so refusing starts to feel like friction. A queue gets faster, so there is less time per item. A dashboard is redesigned and the inputs a reviewer would need to disagree move one click further away. None of that appears in a change log as a transfer of authority, and all of it transfers authority. An organisation can find, entirely without intending to, that a class of decisions it believes it still makes has moved, and the only trace is a number nobody computed.
Not to be low. A high machine decision share is not automatically a failure, and driving it down by inserting reviewers who change nothing is the exact behaviour this measure exists to detect. Plenty of decisions should be automated, and pretending otherwise by staffing a formality is worse than automating honestly.
The number is for knowing. It tells a board which decisions it has actually delegated, so that oversight effort goes where authority went rather than where the org chart says it should be. It tells a regulator something checkable instead of a policy document. And it tells the person whose name is on the form whether they are exercising judgment or absorbing consequences for a judgment made elsewhere.
Chenna, M. (2026). What share of your decisions have you already given away? manjchenna.com/essays/digital-decision-rights
Counting the middle tier needs two things that exist: The Meaningful Override Rate and the seeded error probe. The argument about who ends up answerable is Accountability Inversion.