Fünf KI-Experimente bei WIKA

We spent the summer running five AI experiments at WIKA Group. Three delivered things the team could not do before, and two failed. The failed ones changed what I will build for every client after them.

It started in June with one talk about where AI actually stands. The team leaders and I ran an ideation workshop, and the CEO decided to run five experiments at once, each with its own idea owner.

The two that failed

Both were chatbots for finding information. The users tested the same job with Copilot on SharePoint, and it simply worked. That settled it for me. When a company already runs on Microsoft, I will not build them a search chatbot again.

The three that worked

I can only describe them at a high level, because the details stay with WIKA. One helps the developers write and check code. One works through the technical norms a certification depends on. And one puts 47,000 tools from a stack of supplier catalogs into a single searchable database, so finding the right tool takes minutes instead of half an hour.

The part I did not expect

At the September update, the CEO told me he had been calculating with zero return from the start. He is an engineer, and I suspect that, for him, an experiment fails until it doesn't. Three out of five showing real value was the opposite of what he planned.

The number I keep repeating

One analysis from these experiments costs $320 per run through a general agentic AI setup. The process we built for it does the same job, or a better one, for $4.20 per run. Agentic systems earn their price on complex problems that never repeat. The moment work repeats, it deserves its own process at a hundredth of the cost.

The experiments keep running through September, and the team is measuring what each one actually saves. I am curious how this has gone for others who ran their first AI experiments. Did the failures teach more than the wins?