The instinct with autonomous AI is to treat autonomy as a single dial — cautious or aggressive, on or off. That framing guarantees a bad outcome. Set the dial low and the system is safe but useless, drowning every trivial action in human review. Set it high and the same policy that lets an agent reschedule a meeting also lets it move money or change a production system. The organizations that get real value without real exposure stop asking how much autonomy to grant in general, and start deciding it action by action, priced against what a mistake would cost. That is a governance decision, and it has an economics.

The bet: how much freedom, and where

The decision being made is not whether to trust AI, but how to distribute autonomy across the many things an agent could do — most of them trivial, a few of them consequential. The stakes are asymmetric. The upside of automating a low-risk action is convenience; the downside of automating a high-risk one is a loss that can dwarf every efficiency gain the system produced. A single autonomy policy applied uniformly optimizes for neither.

What makes this timely is that autonomy compounds. Every action an agent takes without review is a small productivity win, which creates pressure to widen autonomy everywhere — including where it should never have been widened. Governing that pressure deliberately, rather than letting it drift, is the actual task.

Autonomy is not one decision

The useful move is to classify each action along two axes that determine what a mistake costs: how reversible it is, and how large its blast radius. A reversible, low-impact action — drafting a summary, tagging a record — can be granted freely; the worst case is cheap to undo. An irreversible, high-impact action — issuing a payment, deleting data, changing a live system — is the opposite, and deserves friction regardless of how capable the model is.

These two axes produce four very different governance postures. Collapsing them into one policy is what forces the false choice between safe-and-useless and useful-and-dangerous.

Reversibility and blast radius decide how much freedom an action should get. Autonomy by consequence Require approvalirreversible, large Notifyreversible, large Log & proceedirreversible, small Automatereversible, small Blast radius Reversibility
Reversibility and blast radius decide how much freedom an action should get.

Risk tiers, not a risk dial

In practice the four quadrants collapse into a handful of tiers that an organization can actually operate. Routine actions — the vast majority — run autonomously with logging. Material actions run with a human notified and able to intervene. High-risk actions require explicit approval before they execute. The shape is a pyramid: many actions at the base moving freely, a narrow apex that never moves without a person.

The value of tiers is that they are legible. Everyone — the teams building agents, the people accountable for outcomes, and the auditors who show up later — can see which tier an action sits in and why. That legibility is what turns autonomy from an act of faith into a governed decision.

Most actions run freely; a narrow apex never executes without approval. Three tiers of action High-riskapproval required Materialnotify & intervene Routineautonomous, logged few most
Most actions run freely; a narrow apex never executes without approval.

The economics of approval

Human approval is not free, and treating it as if it were is the second common error. Every approval consumes scarce attention, adds latency, and — if demanded too often — trains reviewers to rubber-stamp, which quietly removes the very control it was supposed to provide. Approval is a budget, and like any budget it should be spent where it buys the most risk reduction.

This reframes the design goal. The aim is not to maximize oversight but to concentrate it: near-zero friction on the high-volume, low-consequence base, and real, attentive review reserved for the rare actions that warrant it. Approve everything and the system is slow and the reviews are hollow. Approve nothing and the risk is unbounded. Tiered approval spends the budget where it matters.

Uniform policies are expensive at both extremes; tiered approval concentrates the spend. The cost of approval policy Approve everything friction & rubber-stamping 100 Tiered approval targeted attention 35 Approve nothing unbounded risk 90
Uniform policies are expensive at both extremes; tiered approval concentrates the spend.

Matching rigor to blast radius

Operationally, the policy is a routing decision made on every proposed action: assess its risk, then route it to the level of scrutiny that risk warrants. Low-risk actions execute automatically. Medium-risk actions proceed but notify a responsible human. High-risk actions stop and wait for explicit approval. The routing itself should be automatic and consistent; what varies is where the human sits relative to the action.

The discipline is that the threshold is set by the cost of being wrong, not by the confidence of the model. A highly capable model proposing an irreversible, high-impact action still routes to approval, because the question governance answers is not whether the model is likely right, but whether the organization can afford it being wrong.

Every proposed action is routed to the scrutiny its blast radius warrants. Routing by risk Action risk? High Medium Low Human approval Notify & proceed Auto-execute
Every proposed action is routed to the scrutiny its blast radius warrants.

Where governance drifts

Two failure modes recur. The first is uniform caution: every action, however trivial, requires review. It feels responsible, but it makes the system uneconomic and trains reviewers into inattention, so the one approval that mattered gets waved through with the thousand that didn't. The second is uniform permissiveness: early wins with low-risk automation build confidence that silently expands to high-risk actions, until an irreversible mistake reveals that no tier boundary was ever enforced.

Both drifts come from the same root — treating autonomy as one setting. The corrective is to make the tier of each action explicit and to hold the boundary between tiers as a governed line, not a matter of momentum.

⚠️
The quiet failure. Requiring approval for everything doesn't add safety; it trains reviewers to approve without looking.

What to weigh

The judgment worth making is not how autonomous the system should be, but whether autonomy is being allocated by consequence — and whether the human attention it spends is going where a mistake would actually hurt. A system that reviews everything and a system that reviews nothing are both ungoverned; they have simply failed in opposite directions.

The durable posture is a small set of legible tiers, a routing rule that assigns each action to one, and a boundary between tiers that expansion pressure cannot quietly erode. Autonomy then becomes a portfolio decision — freedom where it is cheap, friction where it is warranted — rather than a single bet placed once and forgotten.

Key takeaways

  • Autonomy is not one setting; grant it action by action, priced against the cost of a mistake.
  • Classify actions by reversibility and blast radius — those two axes set the right posture.
  • Operate a few legible tiers: routine runs free, material notifies, high-risk requires approval.
  • Human approval is a scarce budget; concentrate it where being wrong is expensive.
  • Set the approval threshold by the cost of error, not by the model's confidence.
  • Reviewing everything and reviewing nothing are both ungoverned — hold the tier boundary deliberately.

Glossary

Blast radius
The extent of damage a single action or failure can cause before something stops it.
Approval economics
Treating human approval as a scarce, costly resource to be spent where it reduces the most risk.
Risk tier
A defined level of autonomy and scrutiny that an action is assigned to based on its potential consequences.
Reversibility
How easily and cheaply an action can be undone if it turns out to be wrong.