Blog

AI process automation: what changes with AI compared to fixed rules

AI process automation means that a workflow in your business runs on its own, and at the points where a person used to have to read and sort things, an AI now reads and sorts them. In classic automation with fixed rules, every step follows a rule that someone built beforehand. With AI, what changes is the kind of input a process can handle and what happens when something doesn't fit. But some things don't change, and that's exactly what tends to get skipped when people talk about AI business process automation. This article covers both: what AI does differently, what stays the same and where the limits are.

How process automation with fixed rules works

The familiar form of digital process automation often works with fixed paths. If A, then B. A form is submitted, so an entry is created and a confirmation is sent. That works well as long as the input looks the same every time and there are only a few variants.

It gets difficult at three points. First, with input that has no fixed form: a letter has no fields, an email has no fixed structure. Second, with many variants: every new kind of case needs a new rule, and somebody has to build it. Third, with the unknown: if a case doesn't fit any rule, it sits there, gets processed wrongly or breaks the workflow, depending on how the rules are built.

In many businesses with a team, these three points are exactly where a person sits today. Opening the mail and deciding what it is. Reading an email and knowing which case it belongs to. Spotting a special case and asking someone.

What AI changes

The input can be unstructured

An AI employee reads letters, emails and notices, even if each one is laid out differently. It recognizes what kind of document it is and matches it to the right client, property or case.

That means a process can start where the work actually arrives: in the mailbox, in the email inbox, in a download from a portal. Nobody has to copy the data into a form first.

It learns variants the way a person does

With fixed rules, every variant is a path of its own that someone has to build. An AI employee learns variants the way a new person on the team learns them: someone shows it a case, explains how to recognize it and what happens with it, and it saves that as a rule. The rules are written in plain language, the way your team explains them.

That doesn't mean there are no rules anymore. There are actually a lot of them. The difference lies in who writes them down and how easily a new one can be added.

Anything unknown goes to your team

With us, a fixed rule applies to every AI employee: if it doesn't recognize a case, it doesn't guess. It sends the case to a person on your team and lists it as open in the report. In a feedback session, the AI employee is trained on the new variant, and it's saved as a rule.

That's how the escalation path is built with us: a case the AI employee recognizes as unknown doesn't just sit there. It lands with a person and in the report, and next time it's known. What the escalation path doesn't catch is a case the AI employee believes it knows and still sorts wrongly. That's what the report, the sign-off and the review steps are for, as described below under the limits.

A report after every run

With fixed automation, there's often mainly a technical log, which only gets read when something goes wrong. With us, an AI employee sends a report after every run that a person can understand: what's done, what's still open, where there was an exception and what needs a decision. If there's nothing to decide, it says "none". Then you know it thought about it.

The same letter, two ways

An example makes the difference tangible. At a trades business, a letter from a customer arrives, scanned as a PDF. It says that an appointment needs to be moved, and, in passing, that the last invoice still leaves a question open.

A fixed rule looks for features that someone defined beforehand: a sender, a subject line, a field. A letter with two concerns and no fixed structure slips through that rule. It ends up in a catch-all folder, and someone has to read it after all.

A trained AI employee reads the letter, recognizes both concerns and matches it to the customer. It passes the appointment request to the person who plans appointments. The question about the invoice goes to bookkeeping, because it's about money and a person decides there. If it can't match one of the two parts with certainty, it sends the letter to a person through the escalation path. At the end, the report shows what it matched and what's still open.

A real example from a tax advisory firm follows further down.

What AI doesn't change

If you think of AI process automation as "switch it on and let it run", you'll be disappointed. Three things stay just as they are with any new employee.

Training stays

An AI employee doesn't know your business. It doesn't know which type of notice goes where, which client has two properties or which email is actually a complaint. It learns that from your team, in four steps. More on that shortly.

Sign-off stays

With us, an AI employee is signed off after four clean runs. After that, your team watches alongside for a while, and only then does it run on its own with a completion email.

Before that, the process is at Autonomy Level 1, Assisted: the AI employee does its task when someone starts it, and your team checks every run.

What happens at this level is covered in Autonomy Level 1, Assisted: when your team checks every run.

A person stays involved with money, people and legal consequences

For this, a house rule applies at every level: we keep a human involved when decisions are about people, about money or about anything with legal consequences. The AI employee prepares, and a person decides. Every step that can't be taken back has an approval point in front of it. A payment, a submission to a public authority, an email to a client only happens once a person has explicitly said yes, for every single case.

The four steps of training

With us, training follows the same path you know from a new person on the team.

  1. Watch. Together with you, we watch your team handle the process the way it does today, on a normal day, with real input. Every point where someone pauses to think is a decision, and it later becomes a rule.
  2. Sit alongside. The AI employee gets real cases, one variant at a time. It says what it sees and what it would do. The person on your team confirms or corrects. Every variant is shown once, including the rare ones.
  3. Remember and ask. It saves what it has learned. Anything it doesn't know goes to a person through the escalation path, followed by a feedback session.
  4. Sign off. Four clean runs, then monitoring side by side, then running on its own with a completion email.

For training to work, the process has to be described beforehand: trigger, start, work for each variant, end, reporting, plus the escalation path. We call this description a process profile. How it comes together is covered in How to document business processes: the template for handing a process to an AI employee.

Where the limits are

AI makes processes possible that would be tedious with fixed rules. But it doesn't turn every kind of work into a process.

Tasks that need someone to think along. A quote for a new major client, a reply to a delicate complaint, the question of which job to take on: the result depends on your judgment. With us, work like that stays in conversation, with you or with an AI employee as a thinking partner. The autonomy levels only apply to AI employees that take over processes.

Processes without a clear end. If nobody can say how you can tell the process is finished, an AI employee can't check it either. Then the process has to be described first.

Mistakes still happen. Even a trained AI employee can recognize a variant wrongly. That's why, with us, there's the escalation path, the sign-off after clean runs, the report after every run and the approval points in front of every step with no way back. We build these review steps together with the process, not afterwards. Who checks when many processes are running is covered in AI governance in SMEs: who checks the AI when you no longer do?

ChatGPT on its own isn't a process yet. Maybe you already use ChatGPT. It reads a letter when someone pastes it in. It doesn't start on its own, and someone still has to file things afterwards.

It only becomes AI process automation with a trigger, rules, an escalation path and a report.

What this looks like for us

At a tax advisory firm, up to 30 to 40 digital notices arrive a day. Before, the office team needed about 2 minutes per notice. For the notices that arrived on paper, it took up to 2 hours, depending on the volume.

In our assessment, this would have been difficult input for fixed rules. The notice process distinguishes around 45 rules: many types of assessment, different types of tax and of letter, and letters from the city with different layouts. An AI employee reads the notices, recognizes the type and files them, both digital ones and those that arrive on paper.

The paper notices reach the firm's inbox by email. A second AI employee checks the inbox every 30 minutes, spots the emails with paper notices and hands them over to filing. So here, the trigger is a fixed schedule.

Everything that AI doesn't change applied here too. Kevin first watched the office team. Then the AI employee was trained, one type of assessment at a time, with someone sitting next to it. It was signed off after four clean runs, and after that, the team watched alongside it for a while. Today, filing runs on its own, and Kevin gets completion emails. Anything the AI employee can't match with certainty goes to the office team by email, followed by a feedback session. Both processes are at Autonomy Level 2, Autopilot (as of September 2026).

Frequently asked questions

What's the difference between AI process automation and RPA?

RPA copies a person's clicks in fixed on-screen forms. AI reads and sorts, even when the input looks different every time. The two can be combined. A comparison of all three ways is in Intelligent process automation: workflow tool, RPA or AI employee, and which fits when.

Which AI is behind it?

We decide that with you. In the first sessions, we work out together where your AI employee runs, which models it uses, how far the whole thing can be set up in line with data protection rules and what matters to you. We deliberately don't name models here. They change faster than your processes do.

Do I need to know how to program?

No. Your team shows, explains and signs off. The setup, meaning where the AI employee runs and which access it gets, we sort out with you. What process automation covers in a business without developers of its own is in Process automation in a business: what works without developers and where a person stays involved.

What exactly is an AI employee?

An AI system with a fixed role, rules, an escalation path and a report that takes over a task in your business. Kevin explains it in detail on his blog: What Is an AI Employee?

Your next step

AI process automation pays off where a person reads, sorts and files things today, every day, by rules your team knows. If you want to know which process that is in your business, we'll look at it together on a call.