Process Automation vs. AI Agent: Which Fits When

A familiar picture from this year's vendor meetings: a company wants to solve invoice approval "with AI", a vendor proposes an AI agent, and the project sits in the five-figure range. Then the process walkthrough happens, and it turns out the approval follows four rules. Amount below a threshold, supplier on the master list, purchase order reference present, invoice amount equal to order amount. No one needs an AI agent for a process like that. They need a rule in the workflow system.

This is the question many management teams are answering wrongly this year. Not "AI yes or no", but: is classic process automation enough, or is an AI agent genuinely required? The answer drives project cost, running cost and liability. This article gives you the decision rule, the numbers behind it, and a test you can run yourself in 20 minutes.

What process automation actually means

Process automation means software executes a sequence that a human has described in full beforehand. Three variants matter in practice.

Workflow automation routes tasks and approvals between systems and people. The holiday request goes to the manager, then to HR, then into the time-tracking system.

Robotic Process Automation (RPA) operates user interfaces the way a person would. A software robot opens the ERP, copies values and enters them into a second system. RPA is the answer to missing interfaces, not to missing intelligence.

Rule-based data processing checks values against conditions. Amounts, dates, master-data matches, thresholds.

All three share a property that gets underrated in the AI era: they are deterministic. The same input produces the same output, today and in six months. The behaviour is testable, documentable and, if it comes to it, explainable to an auditor.

What an AI agent does differently

An AI agent is given a goal instead of a sequence. It decides for itself which steps to take and in what order, using tools along the way: databases, email, your CRM. We covered the technical difference between a system that answers and a system that acts in AI chatbot vs. AI agent.

The gain shows up where rules run out: unstructured input, and cases that cannot be listed exhaustively. A complaint email in free text. A supplier quote as a PDF with no fixed layout. A customer request that is really three requests.

The price is predictability. An agent can take two different paths from the same input. That is not a defect in the product, that is how it is built. Anyone who puts this property into a process that used to be deterministic is trading predictability for flexibility. Make that trade deliberately, not by accident.

The decision rule: the five-line test

Take the process you are currently thinking about and try to write it down in five lines as if-then rules. Completely, without "and in special cases Ms. Meier decides".

If it works: you need process automation. An agent would be more expensive here, slower and harder to audit, without delivering a single advantage.

If it fails because there are too many special cases: first check whether those special cases are necessary. Very often the process is not complex, it has simply never been cleaned up. A tidied process is frequently rule-based again.

If it fails because the input is unstructured: that is the one solid reason for an AI agent. Free text, images, speech, changing document formats.

Gartner states the same rule from the vendor side: use AI agents where decisions are required, automation for routine workflows, and assistants for simple retrieval. See the Gartner prediction of 25 June 2025.

The numbers: why this distinction saves money

Gartner expects over 40 percent of agentic AI projects to be cancelled by the end of 2027. The reasons given are escalating costs, unclear business value and inadequate risk controls. The sentence that matters most for buyers appears further down in the same release: many use cases positioned as agentic today do not require agentic implementations at all.

There is a market problem on top of that. Gartner estimates that of the thousands of vendors positioning themselves in this space, only around 130 deliver genuinely agentic systems. The rest practise "agent washing", the rebranding of existing products such as assistants, RPA tools and chatbots. You then pay agent prices for automation you could have bought directly for less.

At the same time the direction is clear. Gartner forecasts that by 2028 at least 15 percent of day-to-day work decisions will be made autonomously by agentic AI, up from 0 percent in 2024. The technology is coming. The only question is whether you put it into the right process.

The real bottleneck is not the technology

Two German data points put this in context.

According to the Bitkom survey of 11 March 2026 of 604 companies with 20 or more employees, 41 percent now use AI, up from 17 percent a year earlier. Another 48 percent are planning or discussing it. In the same dataset, 51 percent say they are struggling to cope with digitalisation.

The second figure explains the first. KfW Research analysed in February 2026 which German SMEs actually use AI. Across the board it is 20 percent. Among companies with a digitalisation strategy it is 35 percent, among companies with no digitalisation activity in the past three years it is 19 percent. At the bottom end, among companies with no innovation activity and no graduate staff, it is 8 percent.

So the difference is not the AI tool, it is the maturity of the process underneath. Automate an undocumented workflow and you automate disorder. We described what a clean entry point looks like in AI automation: how companies get started properly.

The normal case is hybrid

In practice, neither side usually wins outright. The robust design looks like this: the AI agent takes the part that requires understanding. Classic automation takes the part that has to be executed.

An example from invoice processing. The language model reads the PDF and extracts supplier, amount, order number and line items. From that point everything runs deterministically again: master-data match, amount check, approval level, posting. Understanding is agent work, posting is rule work.

This split has a second benefit. It limits what an agent can do in the worst case. An agent that extracts text can misread a figure that a downstream rule then catches. An agent with direct posting rights can post.

The cost side that never appears on the pitch deck

Both approaches are usually compared on project price. What decides the case is operation.

A rule in the workflow system costs almost nothing once it is live. It runs until the process changes. The effort is maintenance, and maintenance is plannable.

An AI agent costs per transaction. Every request consumes compute, and those costs grow with volume, not with the project. On top of that come three items that rarely appear in quotes: ongoing quality checks on the output, rework whenever the vendor changes the underlying model, and the logging you need for traceability anyway.

So evaluate every agent proposal with two numbers: what does the rollout cost, and what does one transaction cost in the second year of operation at realistic volume? If a vendor cannot give you the second number, the proposal is incomplete. With rule-based automation you can usually answer that question yourself, and that is its underrated advantage.

Oversight and liability belong in the decision

As soon as a system decides for itself, control becomes a mandatory question. The EU AI Act, Regulation (EU) 2024/1689, requires in Article 14 that high-risk AI systems be designed so that natural persons can effectively oversee them while they are in use. Not every agent in a company falls into that category, but checking whether it does belongs before procurement, not after.

In practice that means three decisions before an agent goes live: which actions may it perform without approval? Above which amount or criticality does a human take over? How is every decision logged so it stays traceable? For classic process automation these questions are far easier to answer, because the rule itself is the documentation.

The first step

On Monday morning, take the process most often named in your company as an AI candidate. Sit down for 20 minutes with the person who runs it daily and write it out as if-then rules. Then count the lines and the genuine special cases.

If the process fits into five lines, request a quote for workflow automation, not for an agent. If it does not fit because the input is unstructured, you have a properly justified agent case and an argument you can carry into any vendor conversation.

If you want to draw that line across several processes at once, we do it with you in the AI Potential Check. We map your workflows, separate what is rule-based from what is not, and quantify effort and benefit for each case. Get in touch and we will come back to you within one working day.