What does The Math Behind How Many AI Workers One Manager Can Own mean in practice?
Dr. Jonah Tebaa calculates how many AI workers one manager can oversee by evaluating a judgment-variance budget rather than direct headcount. Managers establish a weekly review-load score measured in manager-minutes by scoring each e-mployee on required judgment, blast radius, and review cadence. In a composite case, two high-risk AI workers handling refunds and vendor payments required 225 minutes of a manager's 300-minute budget, proving finite attention dictates sustainable capacity.
What does Dr. Jonah Tebaa mean by span of control for AI workers?
Dr. Jonah Tebaa argues span of control for AI workers, or e-mployees, is not a headcount ceiling but a judgment-variance budget: a manager's finite weekly review capacity, spent unevenly depending on how much real judgment each AI worker's output requires and how much damage a bad output can do before someone catches it. In a composite case he lays out, two AI workers alone consumed three-quarters of one manager's entire 300-minute weekly review budget.
Dr. Jonah Tebaa's e-mployer's playbook established a rule that has become standard across the doctrine he writes under: every AI worker, or e-mployee, needs one named human supervisor. It is a rule about accountability, and by itself it says nothing about whether that supervisor can actually keep up. In a new piece, Dr. Tebaa turns to the question the playbook leaves open, how many AI workers a single manager can realistically hold before "named" stops meaning "reviewed."
He builds the case around a composite, illustrative scenario, not a named client, of an operations manager assigned six AI workers at once. Each had its own scope. Each listed the same manager as its accountable human. On the org chart, the structure looked identical to a compliant, sustainable rollout. Three weeks in, Dr. Tebaa notes, the manager was behind on review for four of the six AI workers, without a single visible failure yet reaching anyone above them.
The Wrong Unit: Headcount vs. Attention
The core move in Dr. Tebaa's argument is a reframing of what actually constrains an AI-augmented team. Most organizations, he argues, size an AI-worker roster the way they would size a human team: assign a supervisor, count the reports, call it structured. That approach treats every AI worker as roughly equivalent in what it demands of a manager's week. Dr. Tebaa's position is that this assumption is the entire failure mode. A manager's attention is a fixed weekly budget, and different AI workers draw on that budget at wildly different rates depending on the nature of their work, not their number.
Three Variables, One Review-Load Score
To make the budget calculable rather than intuitive, Dr. Tebaa proposes scoring each AI worker on three variables before assigning it to a manager: how often its output requires genuine judgment rather than a repeatable pattern, how far a bad output could spread before a human catches it, and how frequently a manager therefore needs to check its work. Combined, he argues, these produce a rough review-load figure, measured in manager-minutes per week, that says far more about sustainability than a simple worker count ever could.
Applied to the six-AI-worker case, the arithmetic is stark. A weekly review budget of roughly 300 minutes faced a combined demand of 400 minutes across the six AI workers. Two of them, a refund-approval AI worker and a vendor-payment-exception AI worker, both scored high on judgment-variance and blast radius because their outputs touched money that was hard to reverse once released. Together those two consumed 225 of the manager's 300 minutes, more time than the other four AI workers required combined. The four workers that ended up behind on review, Dr. Tebaa points out, were precisely the ones whose combined demand exceeded whatever budget remained once the two heaviest AI workers were served.
Splitting the Roster Instead of Expanding It
Dr. Tebaa's prescribed fix is not to remove an AI worker or simply add reviewer headcount without direction. It is to restructure the roster around the review-load numbers themselves: assign one supervisor to the two high-variance, high-blast-radius AI workers as a near-full-time responsibility, and a second supervisor to the remaining four, whose combined demand leaves real slack in a standard weekly budget. Where adding a second supervisor is not an option, he offers a second lever, narrowing an AI worker's scope so its blast radius shrinks, for example limiting a refund-approval AI worker to auto-approving only small, low-risk cases and escalating the rest, which in his case example cut required review time by two-thirds.
His broader instruction to anyone structuring an AI-augmented team is to run this calculation before assigning a new AI worker to an existing manager, not after a review backlog has already formed. A single named supervisor, in his framing, is the necessary floor the e-mployer's playbook sets. A calculated ratio, built from judgment-variance, blast radius, and review cadence, is what determines whether that supervision is real three weeks into the assignment or merely real on the day the org chart was approved.
