← Manj ChennaEssays

Measure · A metric I propose

Oversight Budget

Human attention is fixed. The number of decisions a machine makes is not. Your oversight budget is what is left when you divide one by the other, and past a point, it rounds to nothing.

Manj Chenna · Founder, Sanctity · Amsterdam · June 24, 2026

There is a number most teams deploying AI have never calculated, and it quietly governs whether their human oversight is real: how much human attention exists, divided by how many decisions need overseeing. I call it the oversight budget. It is brutally simple, and it is brutal. One reviewer has only so many genuinely attentive minutes in a day. Point a system at them that makes a hundred thousand decisions, and the attention each decision can receive collapses toward zero, however diligent the person is.

What is your oversight budget?

It is attention over decisions, and the denominator wins. Spread one human across a thousand decisions and each can still get a real slice of judgment. Spread the same human across a hundred thousand and the slices vanish. This is not a motivation problem you can fix with a better reviewer or a sterner policy. It is arithmetic. The human can be perfect and the budget can still be empty, because the system put more decisions in front of them than any person could meaningfully attend to.

ONE HUMAN'S ATTENTION, DIVIDEDAcross 1,000 decisionseach one still gets a real sliceAcross 100,000 decisionseach one gets almost nothing
The attention is fixed. The decisions are not. Your oversight budget is one divided by the other.

Why more AI usually means less oversight, automatically

Because deployments scale the denominator and almost never scale the numerator. You ship the model to ten times the volume; you do not hire ten times the reviewers, that would defeat the cost case. So every successful scale-up quietly thins the oversight budget. The oversight does not fail with a bang. It is divided into nothing, decision by decision, while everyone points at the lone human as proof a person is in charge.

What to do with the number

Treat it as a real budget, with a floor. Decide the minimum attention a class of decision deserves, and refuse to scale a decision type past the point where you can fund that attention. Route the cases that exceed the budget to somewhere a human genuinely can attend to them. The alternative is to keep spending an oversight budget you do not have, and to call the overdraft "human in the loop."

Read on

Its companion is contestation latency: not just how many decisions, but how much time each one gets. The outcome measure is the Meaningful Override Rate. Or start here.