The four autonomy levels: how a process in your business runs on its own
If you search for levels of autonomy for AI agents, you mostly find frameworks from research papers and industry groups. They describe how independently an AI system acts and how much the user stays involved. For a business with a team, that doesn't help much. What you want to know is whether filing, the inbox or quote preparation keeps running in your business while you're busy with something else. That's why at Gain Autonomy we measure autonomy per process. There are four autonomy levels, and you can check each one with three questions.
Three questions show where a process stands
Whether a process is at Level 1 or Level 3 isn't a matter of gut feeling. Three questions tell you:
- Who starts the run? Someone on your team, or a fixed schedule or an event, such as an incoming email?
- Who checks? Your team on every run, the AI employee itself with an escalation path, or an AI manager who reviews all the reports?
- What do you still see? Every single run, a report after every run, only the exceptions, or only the goals?
The three answers give you the level. That's all it takes. So the person on your team who looks after the AI employees can propose the level themselves. It gets entered when we look at all processes once a month.
| Autonomy Level | Who starts the run? | Who checks? | What do you still see? |
|---|---|---|---|
| 1, Assisted | someone on your team | your team, on every run | every run |
| 2, Autopilot | schedule or trigger | the AI employee hands anything uncertain to a person | the final report |
| 3, Audited | schedule or trigger | an AI manager | only the exceptions |
| 4, Autonomous | schedule or trigger | the AI manager; at the approval points, you or your deputy, as you decide | goals and exceptions, even when you're away for a week |
The four autonomy levels in detail
Level 1, Assisted: your team starts and checks
The AI employee does its task when someone starts it, and your team checks every run. Every process starts here. This is the level where training happens: first, we watch your team at work with you, to see how it does the job today. Then every variant is shown once, with someone sitting next to the AI employee. Whatever the AI employee doesn't know, it asks about, and in a feedback session it gets saved as a rule.
At Level 1, the three questions come out like this: a person starts, a person checks, and you see every run, either yourself or through your team. That's why we don't count the hours of a process at Level 1 as relief yet. Your team still starts and checks every run. What this level looks like day to day, and when we always keep a human involved, is covered in Autonomy Level 1, Assisted: when your team checks every run.
Level 2, Autopilot: the process starts itself and reports back
The AI employee starts on its own on a schedule or trigger, runs through to the final report and hands anything uncertain to a person. A process reaches this level by crossing a fixed threshold: four clean runs, then a period in which your team watches alongside, and only then does it run on its own.
The three questions: the schedule or an event starts the run. The AI employee checks itself against its rules, and whatever it can't match with certainty goes to a person via the escalation path. After every run, you get a report: what's done, what's still open, where there was an exception.
Level 2 is the level our guarantee covers for processes no. 1 and no. 2. What "Autopilot" honestly means, and what it doesn't, is covered in Autonomy Level 2, Autopilot: the process starts on its own and reports back at the end.
Level 3, Audited: an AI manager checks
An AI manager reviews the AI employee's work and bundles the reports, so you only see the exceptions. Instead of one completion email per process, you get a morning report. You only have to decide at the approval points: the places where a step can't be undone, such as a payment or a submission to a public authority. There, a person explicitly says yes, every single time, before anything moves on.
The three questions: a schedule or trigger starts the run. The manager checks whether every expected run happened and whether the reports fit together. You see the exceptions and the questions only a person can answer. How a reviewing role like this is set up, and what it may not do, is covered in Autonomy Level 3, Audited: who checks the AI when you no longer do?
Level 4, Autonomous: the week-away test is passed
The process has passed the week-away test and keeps running even when you're away for a week. This is how the test is set up in our program: for five working days, you don't step into any of the processes being tested. A person on your team stands in for you at the approval points, and every time a person steps in, it gets logged. The test is passed if every tested process ran every day, nothing was left unfinished and you weren't needed.
The three questions: a schedule or trigger starts the run. The manager checks. At the approval points, you or your deputy decides, and you decide who that is. You set the goals and handle exceptions. How the test is prepared, and what stays with you afterward, is covered in Autonomy Level 4, Autonomous: the week-away test.
The level applies per process, not per business
An autonomy level belongs to one single process. In the same business, filing can be at Level 2 while quotes are still at Level 1 and no AI employee has been trained for bookkeeping yet. That's normal, and it's a good thing: every process has its own variants, its own risks and its own pace.
Three things follow from this:
- No process skips a level. A manager can only check reports that exist, so Level 3 comes after Level 2. And a week-away test needs a manager and a deputy, so Level 4 comes after Level 3.
- Every process gets its own target level. Which level a process should reach next is decided when we look at all processes once a month.
- The levels apply to AI employees that take over processes. AI employees you develop ideas and concepts with keep working with you in conversation. They aren't on any level, because they're not meant to run on their own.
How a process moves up a level
It all starts with a map of every recurring process in your business: who does what, how often, how long it takes, where the work comes in and in which system it gets done. For every process, one question is added: does someone need to think here, or does it simply have to get done? Only the second kind climbs the levels. How to set up this map and pick your first process is covered in Which processes can you automate? The process map.
Once a month, we look at this map together. For every process, we enter the level it's at now and decide what comes next. Two rules apply. Only a level that has actually been reached gets entered, never one that just feels like it. And if a process slips back, for example because runs fail or a person has to step in again, the lower level gets entered the following month.
The thresholds between the levels are fixed:
| From | To | The threshold |
|---|---|---|
| Level 1 | Level 2 | four clean runs, then watching alongside, then on its own with a final report |
| Level 2 | Level 3 | an AI manager reviews the process, the approval points are defined, each with a deputy |
| Level 3 | Level 4 | the week-away test is passed |
That turns "we use AI" into a number you see every month: how many processes are at which level, and how many hours a month no longer sit with your team. Those hours only count from Level 2 onward, and even then only after subtracting what stays with the team: follow-up questions, spot checks, approvals.
Two level models, one stance
If you read Kevin's blog on kevinwelter.com, you'll come across a second level model, and a sentence that might surprise you at first on a brand called Gain Autonomy. In his article "Agentic AI Explained: What It Is and Isn't", Kevin writes: "When people talk about agentic AI, the word autonomy comes up quickly: systems that act on their own. I think that is the wrong axis. Autonomy, in my setup, is not a property a system has or lacks, but a dial I set per role and write down."
That's exactly how the four autonomy levels are built. Here too, autonomy isn't a feature of the AI. It's a dial we set per process, measure against fixed thresholds and write down on the map. The same AI tool can carry one process at Level 1 in your business and another at Level 3. For us, too, the wrong axis would be the question of how independent an AI system is out of the box. That's why we record the level with the process and nowhere else.
The two models still measure different things. On kevinwelter.com, it's about what a role may decide on a single task: what it may establish alone, what it handles under a procedure written down beforehand, and what stays with the human, such as anything that spends money or goes outside the business. There, the three levels are called "1. Alone", "2. By procedure" and "3. Human only", and they count the other way around: Level 1 is the most freedom there, while with us it's Level 4. It's all laid out in What Can an AI Lead Role Decide? The four autonomy levels measure how far an entire process runs on its own.
The two fit together. Even a process at Level 4 can contain a step that only a person signs off, such as a payment. Then a person signs off there, you or your deputy, and the process waits until that has happened. The autonomy level tells you how much of a process still crosses your desk. The decision level tells you which step in it never happens without a person. Put the two models side by side and you see the same stance: autonomy gets set, measured and written down.
What this looks like for us
At a tax advisory firm, Kevin and the team brought two processes up to Level 2, Autopilot. The first process is filing digital tax assessment notices, up to 30 to 40 a day. Before, the office team needed around 2 minutes per notice. An AI employee for filing now does this work. The second process covers notices that arrive on paper. For it, a second AI employee checks the inbox on a fixed schedule, every 30 minutes, spots the emails with paper notices and hands them over to the AI employee for filing.
Apply the three questions and the level follows on its own. Runs start without a nudge from the team, and for the paper notices, through the fixed schedule in the inbox. Checking happens via the escalation path: anything the filing employee can't match with certainty goes to the office team by email, and the AI employee is trained on it afterwards. And at the end, filing reports back with a completion email. Filing was signed off after four clean runs. After that, the team watched alongside it for a while. Both processes are at Autonomy Level 2, Autopilot (as of September 2026).
In Kevin's own business, 14 AI employees are at work. An AI manager reviews their work; that role is still on trial. Kevin tries new levels in his own business first, before they come to yours.
You get the map itself from us as a template. Our autonomy map is a list of all the processes that simply have to get done, with the level for each month and a count of how many processes are at which level. You fill it in with your team in the first month, and from then on it's updated every month.
Frequently asked questions
Can my whole business be at Level 4?
The level belongs to the individual process. A business where many core processes are at Level 3 or 4 is the goal at the end of a long road. You travel it one process at a time.
How long does it take to get from Level 1 to Level 4?
We don't make a promise on that. Our goal for the first month is that process no. 1 starts on its own and reports back with a completion email. In the third month, the goal is an AI manager who bundles the reports into a morning report, and after that comes the test: a week without you. The only commitment is what the guarantee says: if your processes no. 1 and no. 2 aren't running on Autopilot after 90 days, Kevin keeps working with you at no extra cost, as long as you keep the agreement.
Do the levels also apply to everyday ChatGPT use?
Only if it turns into a process that starts on its own. When you develop ideas or rework texts with an AI, you're sitting at the table, and that's how it should be. AI employees like that keep working with you in conversation and aren't on any level.
What happens if a process at Level 2 makes mistakes?
Anything the AI employee isn't sure about, it sends to a person on your team via the escalation path. In a feedback session, the AI employee is trained on the new variant, and it's saved as a rule. If a person has to step in again because results are wrong, the lower level gets entered the following month, until the process clears the threshold again.
Which process would be no. 1 for you?
If a process came to mind while you were reading, one that crosses your desk every day, that's a good place to start. On the call, we look at which level it would be at today and which process is worth starting with.