Which processes are worth automating—and where are you simply burning money? A brief guide to strategic selection.

The current discourse surrounding Artificial Intelligence (AI) is marked by a paradox. On the one hand, Large Language Models (LLMs) from OpenAI or Claude promise unprecedented productivity gains; on the other hand, many pilot projects remain stuck at the stage of isolated experimentation. The problem rarely lies in the algorithms' performance, but rather in a lack of pre-selection for specific use cases. Those who attempt to automate highly complex, strategic, one-off decisions will be disappointed. However, those who identify the "Sisyphus tasks" of daily digital life will achieve rapid success.

Until recently, process automation (often summarized under the term Robotic Process Automation, or RPA) was a binary business. A process had to be 100% rule-based to be automated. Any deviation from the standard protocol led to a system crash. The result: only a fraction of corporate workflows were suitable for automation, as most real-world data—emails, quotes, meeting minutes—exists in an unstructured form.

The Innovation: When Software "Understands" Instead of Just Executing

The technological leap today is that, thanks to LLMs, we have broken through the barrier of unstructured data. Modern workflows, such as those we design at mühlemann+popp, use AI as a cognitive link.

Technically speaking, the AI acts as an interpreter: it extracts relevant parameters from a free-text document (e.g., an email or PDF), validates them against existing databases, and prepares the decision for the downstream system (ERP or CRM). We refer to these as "Agentic Workflows"—systems that don't just blindly follow orders but autonomously manage intermediate steps within defined guardrails.

The Reality Check: Separating the Wheat from the Chaff

Not every process that is technically automatable makes economic sense. Our project experience shows that a cold analysis of the following criteria determines the Return on Investment (ROI).

A Classic Winner: Invoice Processing

A prime example of a worthwhile AI application is the processing of accounts payable. This is where high volumes meet semi-structured data. The AI identifies the creditor, extracts the amount, invoice number, and cost center, and suggests the allocation based on historical data. The human shifts from the role of "data entry clerk" to the role of "controller."

A Classic Loser: Strategic Negotiations

In contrast, consider supplier negotiations for major contracts. Here, empathy, market knowledge, and long-term strategy play the primary roles. Since these negotiations happen rarely and every case is unique, the development costs for AI support far outweigh the benefits.

Conclusion & Business Value

The introduction of AI workflows is not an IT project; it is an exercise in business efficiency.

• For SMEs: The focus should be on relieving administrative cross-functional tasks (HR onboarding, billing, customer inquiries). This offers the greatest leverage for scaling without increasing headcount.

• For Enterprises: The value lies in data consistency across departmental boundaries and the acceleration of global processes.

• When should you wait? If your processes are currently not documented at all. AI cannot cure a bad process; it only accelerates it. Order must precede automation.

At mühlemann+popp, we assist companies in drawing this line between technical feasibility and economic sanity. Our goal is not maximum automation, but optimal automation—where technology truly supports the human element.