Agentic unit economics
Agent governance is often framed as a cost. It is the opposite. The CAO function is a margin protection mechanism — and the only way to prove it is to put agents through the same unit economics discipline as any other line of business.
Most organizations measure what their agents cost per run. Almost none measure what they cost per decision — and those are very different numbers.
A token dashboard tells you a fraction of the story. The true cost of an agentic interaction includes the model, the tools it calls, the infrastructure it runs on, and — critically — the human supervision it triggers. An agent that looks cheap per run can be expensive per outcome if it escalates often or retries frequently.
The number nobody tracks is cost per decision — not per run. An agent that runs forty times to settle one pricing decision has a unit cost that is invisible on a token dashboard. The CAO makes that number visible.
The controls that contain an agent are not free, and they are not constant. Cost per decision is a function of the rung the agent operates at.
Cost per decision divides total interaction cost by completed decisions, not by runs — which is why the Registry's run record carries a decision_completed field. Without it, the denominator does not exist. The Autonomy Ladder adds a second refinement: two of the four terms move with the rung, and they move in opposite directions.
Gate overhead rises. At Rung 5 the jurisdiction gate generates per-action adherence evidence and the security gate requires a full match before the agent may act — real compute, storage, and audit cost on every action. At Rung 2 the overhead is loaded human-approval labor on every proposed action.
Escalation cost falls. A Rung 2 agent escalates every action by definition. A Rung 4 agent escalates only on irreversible actions or policy conflict.
A portfolio that reports one blended cost per decision is averaging across agents whose cost structures differ by construction. The number is not wrong so much as uninformative — it cannot tell you which agent is mispriced, and it cannot support a promotion decision.
When a breach is detected, containment is immediate and automatic: the agent starts no new actions while committed operations close cleanly to its safe-hold point. That is the right trade — a false hold costs a delay, a missed containment costs a compounding breach — but it is a real cost, and a framework that priced the risk it prevents without pricing the control that prevents it would be selling something.
E[false_hold] is where over-tightening shows up as a number. A containment threshold set too aggressively does not announce itself as a governance failure — it appears as throughput that keeps stopping for no reason. Pricing it makes the threshold a decision with two visible sides.
A governed ecosystem is not one that never breaches. It is one that contains breaches faster than they compound.
Damage from a breaching agent accumulates for exactly as long as the agent keeps acting. That makes containment time — not remediation time — the term that decides what a breach costs. Remediation can take days; containment must take seconds. Separating the two is what lets the CAO report that risk was bounded immediately even while the fix was still in progress, and put a number on it.
The v1 business case carried an unbounded P(error) × error_impact term. It was honest about the risk existing and silent about what limits it. With an enforced safe-hold, the impact is bounded by compounding_rate × MTTC — which means governance does not merely acknowledge the risk term, it shrinks it, measurably.
Promotion is an investment decision, not a reward for good behavior. An agent moves up a rung when the trade clears — and not before.
The Autonomy Ladder is usually read as a safety mechanism: it contains agents and revokes autonomy on breach. That is true and it is half the story. Because each rung has a different cost structure, the ladder is also the instrument that tells a CFO where each agent in the portfolio should sit.
Moving an agent up a rung buys lower escalation cost, and pays for it in higher gate overhead and a changed risk term. Promote when the trade clears:
Volume matters. Escalation savings are per-decision, so a low-volume agent may never repay the gate overhead of a higher rung, however well it performs. Decision value matters. The same agent is worth promoting for high-value decisions and not for low-value ones — which is why a rung is earned for a specific region of the reversibility–sensitivity matrix rather than in general. The risk term is bounded, not unbounded, because containment caps how far a breach compounds. That is what makes the trade computable at all.
An organization that has pushed every agent to the highest rung its gates permit has not optimized — it has bought gate overhead it does not need and taken risk it is not being paid for. An organization that leaves every agent at Rung 2 is paying human-approval labor on every action to avoid a cost it never measured.
A governed portfolio has agents distributed across rungs on purpose. The ladder prices both errors.
A CFO already knows the cost of acquiring a customer and the lifetime value of keeping one. An agent that detects at-risk customers and prevents attrition is not a cost line — it is a multiple of avoided acquisition cost. This is the ROI conversation that lands fastest.
The CFO does not need to be convinced that retention has value — they already price it. The argument is simply: this agent converts a known cost (CAC) into a measurable saving. Frame the agent as a retention instrument, not a technology project.
Most agent projects are justified on revenue upside and implementation cost alone. That model is incomplete. A defensible NPV for an agent initiative must include the risk-adjusted cost of error — the term that separates an honest projection from an optimistic one — and, in v2, the cost of the controls that bound it.
The last two lines are what most business cases leave out, and v2 changes both. The risk term is no longer an unbounded P(error) × error_impact — it is an expected loss bounded by containment, because a governed agent stops rather than compounding indefinitely. And E[false_hold] joins recurring opex, because the mechanism that bounds the risk has a running cost of its own. Quantifying both is precisely the work of the CAO function.
When an agent makes a wrong decision, the cost is rarely just the decision itself. It compounds — into remediation, reputation, regulatory exposure, and lost trust. Set that full cost against the cost of mitigating it, and the business case for governance makes itself.
For most organizations operating agents at scale, the cost of mitigation is a fraction of the expected cost of error. That asymmetry is the entire business case for the CAO.
| Error dimension | How it surfaces | Why it is underestimated |
|---|---|---|
| Hallucination | Agent produces confident but false output | Often invisible until a customer or auditor catches it |
| Customer impact | Wrong decision reaches a real account or interaction | Downstream churn is attributed to other causes |
| Reputational | Error becomes visible externally | No line item — but real and lasting |
| Regulatory | Automated decision breaches a compliance rule | Exposure accumulates silently until an audit |
Every formula on this page depends on data that most organizations do not currently capture. Cost per decision, error rates, escalation patterns, value retained — these only become measurable when agents operate inside a governed hub with a common registry, shared best practices, and a protection framework around them.
A CFO cannot manage what they cannot see. The governed hub is the instrument that converts agents from a black box into a managed asset — with a cost, a return, and an owner.
The Agent Registry captures the data behind every formula on this page.