Which processes can you automate? The process map

Business process automation with AI agents, or AI employees, as we call them, starts with a simple question: which processes can you automate? The short answer: the ones that run the same way again and again, whose rules someone on your team can demonstrate, and where all you want to know at the end is that they're done. For a business with a team, the honest answer is longer. In a business with a team, several processes usually qualify. The real question is which one goes first. For that, you need an overview of every recurring process in your business. We call it the process map. In this article, we show how it comes together, how you sort each process and how process no. 1 comes out of it.

Why the question sounds different in a business with a team

If you work alone, you know your processes. After all, you do them yourself. In a business with a team, that's different. Most processes aren't known best by the owner but by the people who handle them every day. They know which cases come up, where things get stuck and what gets left undone. The owner often knows the version from a few years ago, or the version of how it's supposed to work.

That leads to a typical mistake. Without an overview, you start with the process that's annoying you most right now. That feels right. But the most annoying process is rarely the one that saves the most time or money. And sometimes it's also the hardest, because it's full of exceptions.

That's why we turn the order around: first collect every process, then sort them, then score them, then choose. That's how the first month of Owner Autonomy starts: with a map of every recurring process in your business, put together with your team.

The process map: everything that keeps happening, on one list

The map is a table. Each row is a process, each column a question. In our template, the worksheet is called "Process Inventory". It has these columns:

• Process: a short name with an action word, for example "Match delivery notes".
• Area: marketing, sales, fulfillment, customer service, bookkeeping and finance, HR, admin and back office, leadership and management.
• Who does it today? A role or a name. "Me" is an answer too, and often the most revealing one.
• Trigger: What starts the process? A fixed time, an event such as an incoming email, or a nudge, where somebody says "go".
• Start, work, end: What's on the table, what gets done, and how can you tell it's finished?
• Handoff and report: Who gets the result, and who finds out that it's done?
• How often and how long: How many times a month and how many minutes per run, honestly estimated, including rework and follow-up questions. That gives you hours and cost per month.
• Input and systems: Email, mail, portal or phone, and which software the work happens in.
• Thinking or doing? More on this in a moment.
• Irreversible? Is there a step in it that can't be undone?
• Pain today: What's annoying, where do mistakes happen, what gets left undone?

Two of these columns deserve a closer look.

The "nudge" trigger. If a process only starts because somebody says "go", that somebody is needed. Very often, that's you. Nudge processes are a good hint at where your business depends on you. For every nudge, it's worth asking: could this be a fixed time or an event instead?

Start and end. This is the real test of whether something is a process. "Customer service" has no start and no end; it's a topic. "Take a complaint by email and pass it on to the warehouse" has both. If you can't name the start and the end for an entry, split it into two or three processes.

How the map comes together with your team

You don't collect alone. First, you go through area by area and write down whatever comes to mind. Then you talk to every person on your team who does recurring work. A short conversation per person is enough, with six questions:

1. What do you do over and over, every day, every week, every month?
2. What starts it: a time of day, an email, a call, somebody saying "go"?
3. What's in front of you then, and what's finished at the end?
4. Who do you pass it on to, and who finds out it's done?
5. How long does it take once, including follow-up questions?
6. What annoys you about it, what gets left undone, where do mistakes happen?

Be open about what this is for: finding out which work an AI employee can take off people's plates. People who know why they're being asked give more precise answers.

There's one trap when estimating. People like to list only what takes "a lot" of time. The small tasks slip through. Yet they add up: 5 minutes, 20 times a day, comes to 35 hours a month. Without a map, those are exactly the processes you overlook.

Thinking or doing: the most important column

Not all work should run on its own. Some work you don't want to give away at all, because it's about ideas, judgment and decisions. For that, there are AI employees you work with in conversation, thinking partners. Other work simply has to get done. At the end, you just want to know it's done, and a report on it. For that, there are AI employees that take over processes.

On our home page, it reads like this: the levels apply to AI employees that take over processes. AI employees you develop ideas and concepts with keep working with you in conversation. So thinking processes stay in conversation. Doing processes go onto the four autonomy levels, which we describe in our main article The four autonomy levels: how a process in your business runs on its own.

Here's how to sort:

Doing, if most of these statements are true:

• The steps are similar every time, even if there are variants.
• Someone on your team could write down or demonstrate the rules.
• At the end, you can check whether it's right.
• Afterwards, you want a report, not a conversation.

Thinking, if these are closer to the truth:

• The result depends on your judgment or your taste.
• There's no "right", only "better".
• You'd rewrite the result yourself every time.

Unclear usually means: there are two processes in one. An example from the sample sheet in our template, a sample tax advisory firm: "Prepare a client meeting" is thinking. "Gather and file the client's documents" before it can be doing. Split such entries.

Don't be afraid of many variants. A process with lots of case distinctions stays doing, as long as every distinction can be written down as a rule and demonstrated. And don't be tempted to mark everything as doing because it feels like bigger savings. A thinking topic built as a process will be back on your desk within two weeks.

Choosing process no. 1: time, money, risk, feasibility

Only the processes marked "doing" go into scoring. Each one gets 1 to 5 points on several criteria. Our template has six; four of them are the focus here.

Time. How many hours a month go away? The basis is the hours from the map. A process with only a few minutes per run can still come out on top if it happens often.

Money. What does the process cost today per month, hours times hourly rate? And what do mistakes in it cost: rework, follow-up questions, late fees? The other direction counts too. Some processes save little time but bring money in, such as checking open invoices. That's why our template scores the revenue lever as a criterion of its own.

Feasibility. Can an AI employee learn this today? Three questions decide:

• Digital or paper? Whatever arrives on paper has to become digital first. That works, but it's one more step.
• Can it get to the data? And is read-only access enough for that?
• Are there clear rules? "It depends" is fine, as long as someone can say on what. If nobody can, it's closer to thinking.

Note: feasibility is measured by the rules more than by the technology. Whether an AI employee can get into a system can usually be sorted out. Whether your team can explain the rules decides whether the process will be running soon.

Risk. What happens if there's a mistake? What counts most here is whether a step is irreversible: a payment, a submission to a public authority, an email to a customer. Such processes don't go first. And when they do come up, the irreversible step gets an approval point: the AI employee stops until a person explicitly says yes, every single time. That's why a lot of risk earns few points in the scoring.

On top of that, the worksheet scores the revenue lever and the pain for the team. Pain carries the least weight. It's no reason to go first, but it's a good tiebreaker when two processes are level.

From points to a decision

Every criterion has a weight you can adjust. Points and weights produce a ranking. The ranking is a suggestion. You make the decision together, and for process no. 1, we check three questions:

1. Can it run on its own by the end of the first month?
2. Does it have no step that can't be undone?
3. Can someone on your team demonstrate it?

Three yeses, and it's a good candidate. Numbers 2 and 3 get noted alongside it. In Owner Autonomy, within the first two weeks, we agree in writing which processes are no. 1 and no. 2.

One warning to close this step: don't keep adjusting the weights until your favorite process is on top. Then you haven't scored, you've justified. If your gut says something different from the points, that's exactly the point to talk about.

What comes after the choice

Once process no. 1 is chosen, the actual work on it begins. First, it gets written down, in a way an AI employee can take over: with trigger, start, work for each variant, end and report, plus the escalation path. How that works is covered in Documenting processes so an AI employee can take them over.

Then the AI employee is trained, and the process sits on the first level: someone starts it, and your team checks every run. What happens there is described in Autonomy Level 1, Assisted: when your team checks every run.

The map doesn't get put away along the way. Only once a process is running does the next one come up, and again from the map. That turns a one-time overview into a plan for the coming months.

What this looks like for us

At a tax advisory firm, process no. 1 was filing digital tax assessment notices. Put that process on the map today, and you can see the lever: up to 30 to 40 notices arrive there per day, and the office team used to need about 2 minutes per notice. That sounds like little. On a full day, it comes to up to 60 to 80 minutes. How much of that actually goes away today, we haven't measured.

On top of that, the process was very feasible. The notices arrive digitally, and the rules can be demonstrated, even though the assessment notice process as a whole, digital and on paper, distinguishes around 45 of them.

Process no. 2 is the notices that arrive on paper. It's a very similar process with a different start: the paper notices reach the firm's inbox by email. Before, filing them took the office team up to 2 hours, depending on volume.

Because the start is different, it became two processes. For process no. 2, a second AI employee came on board: it checks the inbox every 30 minutes, spots the emails with paper notices and hands them over to the first AI employee, which files them. Both processes are at Autonomy Level 2, Autopilot (as of September 2026).

The map and the scoring described here are our templates from Owner Autonomy. The sample sheet in them shows a sample tax advisory firm with eight processes. Only two of them are based on the real numbers above; the rest are example values.

Frequently asked questions

Should I start with the process that annoys me most?

Only if it also comes out on top in the scoring. Annoyance is an honest signal, but not a good selection criterion. The most annoying process is often the one with the most exceptions, which makes it the hardest start.

What about processes that arrive on paper?

They can be taken over, but they need one more step: the paper has to become digital first, for example as a scan that arrives by email. In the scoring, such a process gets fewer points for feasibility. As process no. 1, one that already arrives digitally is usually the better choice.

Is a process that only takes a few minutes worth it?

If it happens often, yes. What matters is the hours per month, not the minutes per run. That's why the map has both, and the hours follow from them.

Who fills in the map?

You and your team together. You collect first, your team adds what you can't see. Without your team, the processes that never cross your desk are missing. And those are exactly the ones you know least.

Examples from everyday work

If you're looking for ideas on which kinds of tasks are suited to AI at all, Kevin has eight examples of individual tasks, each with effort and limits: Automating Processes With AI: 8 Examples. That article is about individual tasks. This one is about choosing in a business with a team: which of many processes goes first and why.

Your next step

In the end, no list from the internet answers the question of which processes you can automate. The answer is in your own map. If you want to know which process would be no. 1 for you, we'll look at it together on a call.