AI for mid-sized companies: pragmatic instead of pilot projects

01 — Approach

No lighthouse project, just processes that run

The typical AI story in mid-sized companies goes like this: a pilot project is announced, discussed for half a year, a demo gets built, and then everything fizzles out because nobody carries the demo into daily operations. We reverse the order. First we find the routine tasks that eat time in your business every day, then we build the smallest possible solution that runs in production. Not a lighthouse that impresses, but processes that run faster starting next month.

We take this approach out of experience, not theory: we have been working with language models since the first GPT-3 APIs, and before that we spent twenty years delivering digital projects for companies of every size. To see how this fits into the bigger picture of what we do, have a look at our AI agency overview.

02 — Practice

The use case audit

We don't start with a strategy workshop but with a sober audit: together with your people, we look at the actual workflows and find the spots where recurring manual work meets text, documents or data. In almost every business, the same candidates with immediate time savings show up.

• Preparing proposals: pulling together text modules, pricing bases and customer data
• Email triage: classifying incoming mail, routing it to the right place, preparing draft replies
• Summarizing documents: condensing contracts, minutes and reports to the essentials
• Invoice checking: matching receipts against orders and terms, flagging discrepancies
• Screening applications: structuring documents and pre-checking them against the job profile

Want to know what such an audit would look like at your company? Get in touch for a no-strings first call and we will sketch the process for you in 30 minutes.

03 — Practice

Prioritization by effort and impact

The audit usually yields ten to fifteen ideas. We build first where little effort meets a lot of impact: hours saved per week, number of employees affected, error costs that disappear. Whatever is expensive and uncertain goes on the back burner, and explicitly so. Saying no to use case number twelve is part of the consulting.

04 — Practice

Delivery in weeks, not quarters

We take the first use cases into production within two to six weeks, depending on integration depth. That is possible because we don't launch research projects; we combine proven building blocks: ready-made platforms, automation tools like n8n and Make, and direct API integrations where needed. You don't need your own data science department for this, just one contact person in-house who knows the business side. What the technically more demanding cases look like, such as agents and RAG over your own documents, is what we describe under AI implementation.

05 — Practice

Tool selection without vendor lock-in

We don't sell licenses and we don't take commissions. So we choose tools based on your case, not on our portfolio. Ground rule: your data and prompts belong to you, models stay swappable, and every building block must be replaceable without everything else collapsing. Today's best model can be mediocre in six months. A setup that only works with one vendor is a bet you don't have to take.

06 — Current

Where mid-sized companies stand right now

The 2026 numbers show a widening gap. According to Bitkom, 41 percent of German companies now actively use AI, up from 17 percent in 2024. In the classic SMB segment the picture is more sober: current surveys by KfW and DIHK show that only around a fifth of mid-sized companies use AI in day-to-day operations. The biggest hurdles companies name are data protection questions and missing know-how, not the technology itself.

07 — Measurement

Measurable success criteria per use case

Before launch, every use case gets a criterion it has to be measured against: minutes per transaction, cycle time, error rate, or simply the question of whether the team still uses the tool voluntarily after four weeks. After the ramp-up phase we take stock, and whatever misses expectations gets fixed or switched off. We have run the same attitude for years in tracking & analytics: decisions based on real numbers instead of good feelings.