How much an AI agent should decide on its own, and how autonomy levels help you choose

AI agent autonomy levels describe who makes the call at each step of a process. Up to level 3, the system prepares the work and a person approves the important steps. From level 4 the system decides on its own, and at level 5 the process runs without any approval. You set the level per process, not once for the whole agent.

Many managers assume an agent only pays off once nobody touches the process anymore. We think that is the wrong yardstick. In most office work the decision is quick. The time goes into everything before it: finding the data, retyping it, checking it, writing the reply. An agent takes that over at level 3 already.

What do autonomy levels mean for an AI agent?

An AI agent is software that works through a task in several steps by itself. It reads incoming messages, pulls data from your systems, applies your rules and produces a result. The autonomy level answers one question about it: where does it stop and wait for a person?

There are two broad zones. Up to level 3 the agent does the groundwork and a person decides. It drafts the result, checks it against your stored data and puts it in front of someone for approval. Nothing is sent or booked until that person clicks.

From level 4 the system makes its own decisions and needs human approval less and less. At level 5 the process runs fully on its own.

Note the words "important steps". Even at level 3 nobody signs off every small step. People decide where money, a customer relationship or a deadline is at stake.

What does this look like in a real process?

Sales quotes show both zones side by side.

A customer emails a request for an individual quote. The agent reads it, finds the items, fills in prices and margins and has the draft ready in under a minute. Sales checks prices and margins and approves it by hand. The quote then goes out automatically through the usual channel. That is level 3: the system does the work, a person decides on the price.

Recurring quotes with fixed prices and discounts stored in the CRM run without approval. Think of a regular customer who orders the same items every quarter on agreed terms. Sales has nothing to check there, so that path runs at level 5. We explain how quotes with sales approval are built for wholesalers in a separate post.

Same agent, same kind of request, two different levels. That is why the level belongs to the process.

Is an agent worth it if a person still approves at the end?

Yes. Ob Plus, a care and social services provider, is a good case. An AI agent writes their monthly invoices in Lexware Office. At the end of the month the staff dictate all visits in one go, and the agent creates and files the invoices. One employee gives the final approval, then the invoices go out automatically.

So a person still signs off, which puts the process in the range up to level 3. Even so, the work on roughly 300 invoices a month fell from about 25 hours, spread across at least three people, to about one hour. It took one day to build.

Approval costs almost nothing once the agent has done the work leading up to it. We wrote separately about why a human-in-the-loop step often beats full automation.

Which level fits which process?

Three questions are usually enough.

1. Does the decision depend on knowledge that sits in no system? A special discount for a customer who just threatened to leave is one example. Keep the decision with a person, level 3. 2. Is there a fixed rule that everyone on the team already applies the same way? Agreed prices, stored terms, clear deadlines. Then the system can decide, level 4 or 5. 3. What happens if a decision is wrong? If the mistake is easy to fix, a higher level makes sense. If something irreversible goes to a customer or an authority, keep an approval step.

For processes that customers feel directly, we recommend starting at level 3. You get the time back and responsibility stays where it is today.

How does a process move up a level?

A process is ready for the next level when approval has become a formality. The person only ever clicks yes because the draft is always right. At that point the click is just waiting time for your customer.

Moving up is a change to the agent like any other. Every automation we build comes with a maintenance agent that tests changes against past cases. Before the agent starts deciding alone, it runs against the cases a person approved so far, and you see in advance whether it would have decided the same way.

You can also move back down. If the rules change, say after new terms were negotiated, the process can run with approval again until the new rules have settled.

Our recommendation

Don't start by asking whether the agent can do everything alone. Ask how much work sits in front of the decision. The agent takes that over at level 3, and approval stays with the people who know the customer. Pick a first process with clear rules, and the road to level 5 is short.

At Sophera Consulting we build a workflow, an RPA solution or an AI agent, depending on the process, and we agree with you which steps a person decides. An agent is up and running in one to two days, at a fixed price and with no subscription. Which of your processes fits, and at what level it should start, is what we work out in the free automation check.