The Autonomy Budget: Why Liability Should Scale With What Your Agents Can Do

Every organization deploying AI agents is answering a question it never formally asked: How much is this agent allowed to do without a human? In most companies, the answer was not chosen. It emerged from whatever access the integration happened to receive, whatever tools were wired in, and whatever limits someone remembered to set.

That is an odd way to make one of the most consequential decisions in enterprise technology. We would never hand a new hire unlimited system access, spending authority, and the power to act irreversibly, then wait to see what happened. Yet that is how many agents go live.

Autonomy Is Already Being Spent

Autonomy is not abstract. It is granted in concrete units: which data an agent can read, which systems it can write to, how much money it can commit, how quickly it can act, and whether its actions can be undone. Each of these is a withdrawal from an account nobody is tracking.

I call the alternative an autonomy budget: an explicit, deliberate allowance for what an agent may do unsupervised, sized to the evidence the organization can show that the agent behaves as intended. A low-risk, reversible task earns a generous budget quickly. An action touching sensitive data or producing irreversible outcomes gets a small one, and a human stays in the loop.

Liability Should Follow the Budget

Models will always make mistakes. That is not news, and it is the wrong question for accountability. The question a regulator, a court, or a board will ask is different: Did the organization grant autonomy it could not justify?

This is how accountability works everywhere else. The more authority we delegate to a person or a process, the more oversight, evidence, and control we expect to see. AI should be no different: the more autonomy an organization grants an agent, the more accountability it must be able to demonstrate. Responsibility scales with the budget, not with the novelty of the technology.

A Budget in a PDF Is a Wish

Here is where most governance programs fail. They state the budget in a policy document and then rely on prompts and guardrails to honor it. Those are advisory. A clever prompt, a model update, or an edge case can route around them.

The data layer is different. Entitlements decide what an agent can touch, and if the platform says no, the agent cannot say yes. Data products establish who owns and stands behind the inputs. Semantic layers keep agents working from the same definitions as the business. Lineage proves, after the fact, what data drove an action. A budget enforced in access controls and recorded in lineage is not a wish. It is a fact you can show an auditor.

How to Start

1. Classify actions by consequence. Separate what is reversible from what is not, and what touches sensitive data from what does not.
2. Set budget tiers and enforce them technically. Express each tier as entitlements and approval requirements in the platform, not as guidance.
3. Make increases earn their place. Raise an agent's budget only when monitoring, testing, and lineage show it has earned it, and review the decision like any other material change.

The organizations that win with agents will not be the ones that grant the most autonomy. They will be the ones that can afford to grant the most because they can prove it was earned.