How Much AI Agent Autonomy Is Right? Three Levels for Every Step of a Workflow

In short: You can’t settle AI agent autonomy for the agent as a whole. The call has to be made for each individual step of a workflow. In the AI-first operating model, I work with three levels for this, which I call Suggest, Execute with Review, and Act under Oversight. The right level depends above all on how easily a mistake can be undone and how quickly it surfaces. Assigning levels is a leadership decision, and it needs a clear rule for when a step moves up and when it moves back down.

Why is “How autonomous is our agent?” the wrong question?

The question sounds reasonable, yet it leads you astray. Take an agent that prepares quotes at a machinery manufacturer. It pulls customer data from the CRM, prices the job based on past orders, drafts the cover letter, and sends the quote. If something goes wrong in the early steps, little harm is done. A quote that goes out with the wrong price, however, commits the company to that customer.

If you assign this agent a single autonomy level, you are left with two bad options. Either you slow every step down to the pace of the riskiest one, and the agent barely saves any time. Or you give every step the freedom of the most harmless one and let unchecked prices reach customers.

A better approach is illustrated in “The AI-First Operating System,” the report the World Economic Forum published with Kearney in June 2026. In a supplier sourcing process, the team sets the objective without AI. The AI handles the financial screening of potential suppliers on its own and delivers a report. In the technical and legal assessment, it flags key issues, while approval stays with a person. Ongoing supplier monitoring goes back to the AI. A single workflow, in other words, contains every level.

Where do autonomy levels for AI agents come from?

The idea predates today’s agents. In automated driving, the SAE J3016 standard has distinguished levels 0 through 5 for years. For AI in the enterprise, there are now two models I draw on.

The World Economic Forum uses a five-level scale based on the Human Agency Scale developed by Shao and colleagues (2025). It runs from an agent that completes a task with no human involvement to a task that remains entirely with a person. Researchers at the University of Washington led by Kevin Feng also describe five levels, but they organize them around the role of the human. At the lowest level, the user operates the agent like a tool. At the highest, the user only observes.

Five levels make sense for research. For work with leadership teams, I reduced the scale to three because they are easier to apply day to day. With five adjacent levels, the lines blur quickly and decisions get stuck in debates about definitions. Each of my three levels stands for a different control mechanism, and that is what matters for leadership.

What are the three levels of AI agent autonomy in the AI-first operating model?

LevelWhat the agent doesWhat the human doesBest suited forExample
Suggestprepares and recommendsdecides and actssteps that call for judgment or have consequences that are hard to reverseThe agent finds comparable complaints and drafts a recommendation. The case handler decides and writes to the customer.
Execute with Reviewhandles routine cases on its ownreviews exceptions and sampleshigh volumes of similar cases where errors are visible and fixableThe agent processes clear-cut warranty claims and routes everything else to a person.
Act under Oversightacts within fixed limitsmonitors metrics and warning signswell-defined, recurring steps with low individual riskThe agent reorders spare parts once stock falls below a minimum. A person reviews the weekly report.

The difference lies in when control happens. With Suggest, a person checks before every action. With Execute with Review, a person checks individual cases after the fact. With Act under Oversight, a person looks at the overall result. Choosing a level also means deciding where in the workflow human time is needed.

How do you assign a process step to a level?

Four questions help.

  1. Can the consequences be undone? A misrouted ticket can be fixed in minutes. A payment that has gone out or a price committed to a customer cannot.
  2. How uniform are the cases? A thousand similar invoices lend themselves to a higher level than twenty special cases a year.
  3. How much judgment does the step require? Where experienced employees would reach different conclusions, the decision belongs with a person.
  4. How quickly does an error surface? A mistake caught at the next customer touchpoint costs less than one discovered in the year-end audit.

Deloitte reaches a similar conclusion. In its view, AI is particularly well suited to work that is repetitive and data-heavy, and to tasks that follow rules and have clear success criteria. Areas involving complex judgment or significant risk still require close human involvement. When two answers point in different directions, I recommend the more cautious level until data justifies a higher one.

When can an agent move up a level, and when does it have to move back?

The line between what an agent can take on and what stays with people shifts over time, a point the WEF makes explicitly. An assignment that is right today should therefore be reviewed again within six months at the latest.

For moving up, define in advance what error rate the samples must show and over what period. There should also have been no serious incident, and the person accountable for the business process has to agree. For moving down, one relevant incident is enough. The step drops one level immediately until the cause is understood. The same applies when the underlying model or key data sources change.

Without such a rule, something happens that I call autonomy creep. Nobody decides to give an agent more freedom. After a few months without problems, though, nobody looks closely anymore, and Execute with Review has quietly turned into acting without oversight. That is why each agent’s autonomy level, along with the date of its last review, belongs in the agent register.

What does the autonomy level have to do with liability and the EU AI Act?

For high-risk systems, Article 14 of the EU AI Act requires effective human oversight. The people exercising it must be able to interpret the system’s output correctly and stop it when necessary. The regulation explicitly names the risk of relying too heavily on a system’s output. For high-risk systems under Annex III, these obligations apply from December 2, 2027. Most agents in a company will not fall into that category. As a benchmark for good oversight, the provision is still useful.

Then there is management liability. Putting a step at the highest level is a business decision. It should be documented so that it is clear later on what information it was based on.

Which mistakes should you avoid when setting autonomy levels?

The most common mistake is granting autonomy based on technical maturity. Knowing that a model can handle a task tells you little about what happens when it gets it wrong. The level should follow the cost of an error.

The second mistake is a review that turns into a formality. An agent that is wrong rarely sounds unsure. Anyone who approves a hundred suggestions a day will eventually just click through. A better approach is a sample review in which the reviewer records their own assessment before seeing the agent’s suggestion. That keeps judgment sharp.

The third mistake is having no rule for moving back down. Without one, autonomy only moves in one direction, which leads straight to autonomy creep.

What should leadership teams review now?

  • One question: For which of our agents is the autonomy level documented step by step, and who made that call?
  • One metric: the share of reviews in which a person overrides the agent’s suggestion. If it sits at zero for months, chances are nobody is really reviewing anymore.
  • One conversation: with the business owners about a rule for when a step moves up and when it moves back down.

Frequently asked questions

What autonomy levels exist for AI agents?

Five-level scales are common, for example from the World Economic Forum or from Feng and colleagues. In the AI-first operating model, I use three levels per process step: Suggest, Execute with Review, and Act under Oversight.

What is the difference between human in the loop and human on the loop?

With a human in the loop, a person approves a step before it takes effect. That corresponds to Suggest. With a human on the loop, the agent acts on its own while a person monitors the outcome and can step in. That corresponds to Act under Oversight.

Who decides an agent’s autonomy level?

The person accountable for the workflow proposes a level for each step. Senior management sets the boundaries, for instance that steps involving binding commitments to customers are never executed autonomously. Both belong in the workflow’s decision matrix.

Further reading: The AI-First Operating Model · The organization after AI agents · AI agent risk and the board · The EU AI Act for boards

Sources: World Economic Forum in collaboration with Kearney (June 2026): The AI-First Operating System. A Blueprint for Operating and Business Model Innovation, pp. 26–27 · Shao et al. (2025): Future of Work with AI Agents (Human Agency Scale), Stanford University · Feng, McDonald & Zhang (2025): Levels of Autonomy for AI Agents · Deloitte (2025): AI-first companies · Regulation (EU) 2024/1689 (AI Act), Art. 14, EUR-Lex · Nicole Lontzek (2026): Die neue Betriebslogik, Chapter 3

Similar Posts