Flat illustration of a laptop in a video call, next to a three-step process with a person as the review step and a shield with a check mark
Back to blog
Generative AIProcess AutomationHuman in the Loop

AI Expert Panel Recap: Getting Past the Hesitation

Sven HennessenEvents

TL;DR

  • The foundation comes first: AI doesn't fix broken processes, and it's only as good as the data it works with.
  • Once that's in place: get past the hesitation, put compliance first, keep a human in the loop. Secure AI solutions are practical today, on your own hardware or with European providers.
  • What a model costs depends on the case: how often the use case runs, how large the context and request are. Models hosted in the EU are affordable, and local installations can complement them.

In April I was the expert at an online panel hosted by the Mittelstand-Digital Zentrum Saarbrücken, a publicly funded digitalisation hub for SMEs, on process automation with generative AI. A look back at two live demos from real client projects, a question about costs, honest feedback, and why data and processes come before any AI project.

In mid-April I was a guest of the Mittelstand-Digital Zentrum Saarbrücken, if only virtually. The online expert panel was called "Process automation supported by generative AI", and in this case the expert was just me. On the agenda: opportunities, risks, and how to use generative AI safely in business processes.

An audience at every stage

The audience covered pretty much every starting point you find among mid-sized companies: businesses with no concrete AI plans yet, businesses in the middle of planning or first trials, and those already using AI. Add to that a mix of service providers, manufacturers, retailers and trades, many of them small firms.

Presenting remotely has its own rules. The chat stayed quiet, and there's no audience in front of you to read.

Two live demos from real projects

Instead of only showing slides, I demonstrated two solutions live. Both come from practice.

The first demo, the 8D report, came out of a conversation with a client. In manufacturing, a tolerance deviation often triggers an 8D report, a multi-page document with eight mandatory sections. The measurement data is already digital. In the demo, simple threshold logic checks the measurements, and at the end of a batch an LLM drafts the report from the deviations. Quality management reviews and signs it.

The second demo, on inventory planning, is the next evolution of a client project. Behind every order suggestion sits a complex calculation of consumption, replenishment lead time, stock and safety stock. The result is a single number, and there's little room on screen to explain how it came about. An LLM sums that up in a few sentences. The quantity itself is still calculated by the algorithm; the model only explains.

Excerpt from Visposition: a suggested order quantity of 120 with an AI explanation citing consumption, lead time, stock and the safety target as the reason

Strictly speaking, by the way, this wasn't explainable AI, even though that's what my slide called it. What gets explained isn't how an AI model reached its decision, but what a classic algorithm calculated. More "explaining AI" than explainable AI. That explanation is now part of Visposition, the product that grew out of the project. How explanations like these build trust step by step, and later automation, is the subject of Explainable AI for SMEs.

The cost question

At the end came the question that comes up in almost every session: what does a model like this actually cost?

My honest answer: it depends on the specific case. How often does the use case run? How large are the context and the request? The good news for mid-sized businesses: models from European providers can be booked at an affordable price. Local installations can be a complement. When your own hardware pays off is something we worked through in Is a Local AI Setup Really Worth It for a Company?.

What the feedback showed

After the panel, the Mittelstand-Digital Zentrum sent out a survey. The feedback was positive overall. The live demos and the practical examples went down particularly well, including as a good starting point for people who are just getting into the topic.

There was criticism too, and it was fair. The applications in the demos were sometimes hard to make out on the shared screen. I now pay even more attention to size in remote talks. And both examples came from manufacturing, so someone working at a public institution, for example, can't transfer them one to one.

What I'd tell every mid-sized business

Every AI project rests on a foundation without which none of the rest holds up: data and processes.

AI doesn't fix broken processes. Digitise chaos and you get digital chaos, not a return on investment. And an AI decision is only as good as the data behind it: garbage in, garbage out.

Both demos only work because that foundation is in place. For the 8D report, the measurement data is already digital and checking it against the tolerances follows clear rules. In inventory planning, an algorithm calculates on the data from the ERP system, and the LLM only explains the result. Where data is patchy or nobody can describe the workflow cleanly, even the best model won't help. How to find out whether your data is good enough is covered in Is Your Data Good Enough?.

Once the foundation is there, three more ideas come in:

Get past the hesitation. Running LLMs securely is practical today, on your own hardware inside the company network or with European providers that have no US parent company.

Compliance first. Sending process data to US services puts trade secrets at risk and runs into trouble with the GDPR. The CLOUD Act and FISA 702 apply even when the data centre is in Europe. The right choice of provider and operating model avoids this, and it belongs at the start of a project, not the end.

Human in the loop. In both demos, the AI prepares and explains, and a person reviews and decides. That's how trust is built. Only once it's there can low-risk tasks be fully automated, step by step.

Sources

Need support?

Do you have a process where data from several systems comes together and someone spends a lot of time on boilerplate text or explanations? Let's look at the workflow together and build software that uses generative AI where it genuinely helps. Privacy-compliant, with a person making the final call.

Digitalize your process