Agentic AI
The Ratio Problem: How Many Agents Can One Person Run
Every company will have access to the same agents. The same models, the same vendors, the same integrations. Within a year, the agent your competitor runs on their pipeline will be functionally identical to yours.
Access stopped being the edge. What separates the teams pulling real output from agents from the teams drowning in half-checked drafts is a number almost nobody has calculated: how many agents one person is able to run before quality breaks.
I call it Agent Span of Control. It borrows from a management idea older than the CRM. Span of control asked how many direct reports one manager handles well. Seven was the old rule of thumb. Nobody assumed one manager scales to forty people without something giving.
Agents get treated differently. A team buys twelve of them, assigns them to five people, and nobody asks whether those five people have the hours to review what twelve agents produce. Output volume goes up. Review quality goes down. Six weeks later a bad forecast number reaches the board and everyone blames the model.
The ratio failed before the model ever ran.
The formula
Agent Span of Control = base task capacity ÷ (risk weight × complexity factor)
Base task capacity is how many agent outputs one person reviews in a day without quality slipping. It's role-specific. An SDR reviewing lead-research summaries might handle fifty. A marketing lead judging brand-voice drafts might handle twenty-five.
Risk weight scales for exposure. Low-risk output, easy to catch and fix, weights one. Customer-facing work weights two. Anything financial, legal, or board-facing weights four.
Complexity factor scales for judgment. Routine tasks factor one. Moderate judgment factors two. Expertise-heavy work factors three.
A worked example
Take an account executive. Base capacity of forty reviewed outputs a day. The agents do deal research and outreach drafting, low risk and routine, so risk weight one, complexity factor one.
ASC = 40 ÷ (1 × 1) = 40.
That rep is able to run up to forty low-risk agents before the math flags a problem. If they run five today, they have real headroom.
Now take a customer success manager reviewing renewal-risk scores. Same base capacity, but the output touches churn decisions and needs judgment. Risk weight two, complexity factor two.
ASC = 40 ÷ (2 × 2) = 10.
Same person, same day, a quarter of the capacity. Give that CS manager fifteen agents and you've built a churn problem wearing an AI problem's clothes.
The ratio isn't one number for the company. It's a per-role calculation, and it changes every time an agent gets promoted to a higher risk tier.
Run it yourself
I built a calculator for this. Five questions, ninety seconds, your number with the math shown. Run the Agent Span of Control Calculator
It uses defaults for base capacity by role. The real version runs against your actual team, your actual risk exposure, and your actual pipeline, which is what the Diagnostic Sprint does. But the free version is enough to tell you whether you're running a safe ratio or an accident with a timeline.
Where governance comes in
Most AI governance stops at policy. A document says who's allowed to approve what. Nobody connects it to how many agents each approver is realistically reviewing.
The ratio is the bridge. When you know one CS manager's ceiling is ten agents at medium risk, governance stops being abstract. You know exactly when the next agent added to that role needs a second reviewer, a lower risk tier, or a pause. A RACI for agent decisions only works when the people marked accountable have the capacity to be accountable.
The ratio tells you whether they do.
FAQ
What is Agent Span of Control?
Agent Span of Control (ASC) is the number of AI agents one person is able to review and manage well, calculated as base task capacity divided by risk weight times complexity factor.
How many AI agents can one employee manage?
It depends on the role, risk, and complexity of the task. A low-risk, routine task might support forty agents per person. A high-risk, judgment-heavy task might support two or three. There is no single company-wide number.
What happens when a team exceeds its agent ratio?
Review quality drops before anyone notices. The person approving agent output starts approving faster and catching less. The fix is tighter risk tiering, added oversight, or fewer agents on that role.