Automating business processes: where it pays off and where it does not

Business process automation is usually written about as though the answer were always yes. It is not. Automation is an investment with running costs, and a noticeable share of the projects we look at do not pay off under an honest calculation.

This is not a warning against automation. It is a warning against starting without doing the arithmetic. This piece provides the arithmetic, plus the process types where it regularly works out and the ones where it does not.

What automation actually costs

The implementation cost is in the offer. The other three items usually are not.

Operations. Automated processes run on servers, need updates, monitoring, and somebody who reacts when they stop. Budget ten to twenty percent of the implementation cost per year, more for AI-supported processes because model usage is billed on top.

Changes. Processes change. A new supplier sends a different invoice format, a tax authority changes a reporting duty, sales introduces a new category. With an automated process each of these is a small project, whereas a human simply does it.

The exceptions. This is the item most often underestimated. Automation rarely covers a hundred percent of cases. The remaining five to fifteen percent still land on a human, who now handles nothing but the difficult cases without the easy ones as relief in between.

The calculation

The annual benefit is easy to establish:

Time saved per case  ×  Cases per year  ×  Fully loaded cost per hour

Fully loaded cost, not gross salary. For commercial roles in the Mittelstand that typically runs between EUR 45 and 75 per hour depending on region and qualification.

Against that you set: implementation cost once, plus operating cost per year. A project is viable if it pays for itself within two years. Anything longer is a bet, given the current pace of the technology.

Two examples with realistic numbers.

Case A, which pays off. Order intake: 40 orders per week arrive by email and PDF and are typed into the ERP by hand. 15 minutes per order, 48 working weeks.

• Benefit: 0.25 h × 1,920 cases × EUR 60 = EUR 28,800 per year
• Cost: EUR 40,000 implementation, EUR 6,000 per year operations
• Result: pays for itself in roughly 21 months, then just under EUR 23,000 annually

Case B, which does not pay off. Holiday requests: 200 per month, each costing 4 minutes to process.

• Benefit: 0.067 h × 2,400 cases × EUR 60 = EUR 9,600 per year
• Cost: EUR 30,000 implementation, EUR 3,600 per year operations
• Result: pays for itself in roughly five years

Case B is the more interesting one. It has more cases than Case A and still does not pay off, because the saving per case is too small. Frequency alone is not an argument. The product of frequency and time saved is one.

For Case B there is also usually a better answer than AI: a decent form with an approval rule. That costs a fraction and never breaks.

Four process types that almost always pay off

Extracting data from documents that arrive in many formats. Supplier invoices, orders, delivery notes, certificates. High frequency, meaningful time per case, and the variety of layouts is exactly why nobody has automated it so far. This is the classic case where AI beats conventional software. Which steps of such a workflow actually need AI and which are cheaper with rules is covered in intelligent process automation.

Answering recurring inquiries. Customer inquiries, supplier inquiries, internal requests to IT. Not as a fully automatic chatbot, but as a draft that a human reviews and sends. The time saving is in the writing, the quality control stays with the human.

Reconciling data between systems. When employees regularly have two screens side by side comparing values, that is an automation candidate with a very clear calculation. Often it does not even need AI, just an interface.

Checks with clear criteria and unclear inputs. Completeness checks on documents, plausibility checks on submitted figures, pre-sorting applications. The rule is clear, the material is unstructured.

Four that almost never pay off

Processes with fewer than roughly fifty cases a year. The effort for setup, testing, and operations exceeds the benefit. A person doing this twelve times a year is the cheaper solution, even if it feels unfashionable.

Processes that change every few months. Every change is rework. If the process itself is still unstable, you are automating a state of affairs that will not exist next year. Stabilize first, then automate.

Decisions with high error costs and no review step. If a mistake costs a customer, a certification, or a six-figure sum, you need human approval. That takes time, and that time eats the saving. In those cases automate the preparation of the decision, not the decision.

Processes that run on relationships. Price negotiations, complaint calls with long-standing customers, anything with a conflict in it. Technically feasible, commercially damaging.

The most expensive mistake

The biggest single mistake is not picking the wrong tool. It is automating a bad process.

Many workflows in the Mittelstand grew historically. A form exists because somebody had a follow-up question twelve years ago. An approval step exists because something once went wrong. If you automate that workflow, you set it in concrete. It will run faster and remain just as pointless, except now it is expensive to change.

So every automation should be preceded by a short, uncomfortable question about each step: what happens if we remove this step entirely? In practice this exercise regularly eliminates a quarter of the steps. That is a saving which costs nothing and takes effect immediately.

Only then is it worth asking how the remaining workflow gets automated.

Where to start

Take one process, run the numbers as above, and put the calculation on the table before you request an offer. With a figure in mind you can judge any proposal, and you will spot immediately when a vendor wants to automate a process that does not carry itself.

If the numbers work and you want to know what the implementation looks like: thirty minutes with an engineer. What we concretely take over is on our AI automation page. If they do not work, we will tell you that too, and you will have saved yourself a project.