Autonomy is a dial, not a switch

Key takeaways

• Treat autonomy as five levels of authority, not a yes or no decision.
• Choose a level per function and per action type, never one level for the whole company.
• Authority should be earned through evidence from replayed and supervised work.
• Lowering authority must be as easy as raising it.

When organizations talk about automation, the question is usually binary: should this process be automated or not? That framing is a large part of why AI initiatives either stall in a pilot or fail loudly in production. It treats autonomy as a switch. In practice it behaves like a dial.

A switch has two states and one decision. A dial has many positions, can be turned in both directions and can be set differently for each job. Once leaders start thinking this way, the conversation changes from "can the AI do this?" to the far more useful "how much should it be allowed to do, and what would make us comfortable turning it up?"

The five levels of authority

We describe the authority of any autonomous system on a five-step scale. The scale is simple on purpose, because a leader should be able to say in one sentence what a system is permitted to do.

1. L1, Assist. The system summarizes, drafts and recommends. People carry out the work.
2. L2, Execute on approval. The system prepares the complete action. Nothing happens until a person approves it.
3. L3, Operate within limits. The system acts independently inside thresholds you define, such as amounts, counterparties or categories, and escalates anything beyond them.
4. L4, Run the function. The system owns the day-to-day outcomes of a function. People set direction and audit results.
5. L5, Coordinate the organization. Several functions plan and act together, sharing context. This is where research is still active.

Choose a level per function, not per company

The most common mistake is to pick one level for the whole organization. Risk is not uniform, so authority should not be either. Consider three examples from a typical company.

• Routine customer replies are low risk, high volume and easy to reverse. They can reach L3 quickly, with a human reviewing exceptions and a sample of resolved cases.
• Payments and ledger postings carry financial and audit consequences. They often sit at L2 for a long time, and move to L3 only for narrow, low-value categories.
• Hiring decisions involve people's livelihoods and legal exposure. The system can assess and recommend, but the decision stays with a person. That is a deliberate permanent ceiling, not a failure of the technology.

Notice that these are decisions about action types, not about departments as a whole. A finance system might run reconciliation at L4 while keeping payment release at L2. Granularity is what makes autonomy safe enough to be useful.

How authority is earned

Authority should never be assumed. It should be earned in steps, with evidence at every step. A sound path looks like this:

1. Replay. Run the system against historical cases and compare its decisions with what actually happened. Study every disagreement.
2. Shadow. Let it work alongside the live process without acting. Measure how often it would have been right.
3. Supervised. Let it act, with every action approved by a person. Track the approval rate and the reasons for overrides.
4. Independent within limits. Let it act alone inside tight thresholds. Review exceptions and a regular sample.

At each step, the question to answer is the same: what is the measured rate of error, what does an error cost, and can we detect and reverse it quickly? When the answers are good enough, the dial moves up. When they are not, it stays put or moves down.

What leaders should insist on

• A written statement of authority for every system: what it owns, what it may decide and what always needs a person.
• Hard limits that are enforced by the system itself, not only described in a policy.
• Approval queues that give approvers the context they need to decide in seconds.
• A complete record of every input, decision and action, searchable by anyone who audits it.
• A visible, one-step way to lower authority.

Where this leaves us

Most of our deployments begin at L2, where everything is prepared and nothing happens without approval, and advance function by function as the evidence accumulates. Some now run entire functions with only occasional human intervention. None of them got there by flipping a switch.

If you would like to feel the difference between the levels, the dial on our homepage lets you set the authority yourself and watch every step move between the system, an approval and a person.