Introducing AI in your company, from the first idea to a running system

Where do we start with AI?

Start with one process, not with a tool, a platform or a strategy paper. Good first candidates have many similar cases, data a system can reach, and a result a person can check much faster than doing the work: incoming invoices, customer enquiries, quotes or questions to internal knowledge. Finding that first case is the hard part for many companies. KfW Research called identifying the first practical application "eine zentrale Hürde", a central hurdle, in its study of AI in the Mittelstand published on 29 July 2026.

Know-how is the other hurdle. In Bitkom's survey published on 14 September 2026, 85% of companies that do not use AI said they lacked the technical know-how, and 51% lacked the people. Both are reasons to start small and get help. Neither is a reason to wait. The guide on choosing the first AI workflow shows how to score candidate processes against each other.

How many companies already use AI?

Depending on the survey, between 20% and 57% of German companies use AI, and the figures differ mainly because each survey counts something different. Four recent surveys side by side:

Bitkom counts companies with 20 or more employees, KfW counts small and medium-sized enterprises, and IAB counts establishments of all sizes. IAB asks only about generative AI, KfW about at least one AI technology. The surveys were also taken at different times, between spring 2025 and summer 2026. Quote any of these numbers only together with its base. In the two surveys that report a previous year, use rose sharply: Bitkom from 36% to 57%, ifo from 40.9% to 54.5%. For your own decision, how many others use AI matters less than whether one of your processes gets measurably faster or cheaper with it.

What stops companies from using AI?

Mostly the same few things, and they differ between companies that already use AI and those that do not. In Bitkom's survey published on 14 September 2026, companies that use AI named data protection requirements (66%), legal uncertainty (56%), hard-to-estimate costs (54%), concern that company data falls into the wrong hands (45%), lack of employee acceptance (42%), poor result quality (26%) and lack of traceability of results (25%). Companies that do not use AI named lack of technical know-how (85%), legal uncertainty (66%), costs (63%) and missing data (40%).

Each of these barriers has a concrete answer, and most of them are covered in a guide of their own:

AI Act checklist for deployersAI business caseChange management for AI

Why do so many AI pilots get stuck?

Most pilots show that something can work, and then nobody takes the steps that make it work every day. Lünendonk's study of 180 CIOs, IT heads and business unit managers in the DACH region, reported by CIO.de on 17 September 2026, found that 58% of upper Mittelstand and large companies are piloting AI agents, 19% use them productively in individual areas and 1% end to end. 66% said that fewer than a quarter of their piloted agents leave the test lab. The study's author summed it up: "Was fehlt, ist die operative Konsequenz", what is missing is operational follow-through.

The same study names legacy IT without integration layers and poor data connectivity as the main barriers to scaling agents. In our reading, a pilot stalls for a few recurring reasons:

• No acceptance criteria: nobody agreed what result would justify the next step, so every result is open to debate.
• No integration: the pilot ran on a copy of the data or in a separate tool, and connecting it to the ERP, CRM or document system was never planned.
• No owner: the pilot team moves on and nobody is accountable for quality, cost and changes.
• No early approvals: IT security, the data protection officer or the works council see the system for the first time just before go-live.
• No baseline: the old process was never measured, so the benefit cannot be shown.

Each of these is cheaper to settle before the pilot than after it. The guide on moving an AI pilot into production covers the handover in detail.

We bought licences and tried ChatGPT. Why has nothing changed?

Because a licence gives people a new tool and leaves the process as it was. Bitkom's president put it this way on 14 September 2026: "Wer nur eine Lizenz kauft, hat KI eingekauft, aber noch nichts verändert", whoever only buys a licence has bought AI but has not yet changed anything. In the same survey, only 3% of companies using AI said they exploit its potential rather strongly.

Still, it is a useful starting point, because usage shows where people already find AI helpful. Three things to do next:

• Find out who uses what, how often and at what cost, including private tools used for work. That tells you where the demand is.
• Pick the task that comes up most often inside a process, such as answering one type of enquiry or checking one type of document, and measure how long it takes today.
• Turn that task into a workflow connected to your systems, with approved sources, permissions and a person who approves the result. That is where the time saving becomes visible.

The free AI usage review covers the first step: one page on AI spend and quality, built from exports you choose to send.

Which steps lead from the first idea to a running system?

Each stage ends with a decision and with evidence for the next one. Skipping a stage usually means doing it later, under more pressure.

1. 1

   Check readiness on your own process

   Take the candidate process and one real sample case, and check whether data, systems, rules and people are in place. What you get is a list of gaps, each with a name next to it.

2. 2

   Design the first workflow

   Describe the work step by step: where cases arrive, what the system reads, what it prepares, where a person approves, and what goes back into which system. Measure today's process at the same time.

3. 3

   Run a pilot with acceptance criteria

   Agree in writing, before the pilot, which failure types are unacceptable, which share of correct cases is good enough, what a handled case may cost and what each result will fund. Then test on your own cases, often first in shadow mode, where the system works alongside people without its results being used.

4. 4

   Go into production on your own systems

   Connect the workflow to the ERP, CRM, document management or ticket system with the users' permissions. Decide where it runs according to the data, for example in your own cloud account or on your own servers, so that data does not leave the company. Pass an acceptance test on your own cases before go-live.

5. 5

   Operate and improve

   Name an owner, monitor quality and cost, run the tests after every change, and write a runbook, a written guide for running the system and handling incidents. Then pick the next process and reuse what you built.

Do we need an AI strategy and a roadmap before we start?

You need a short one, and its first item should be a running system. Many companies already have a strategy paper: in KPMG's survey of 480 decision-makers in Germany, published on 9 June 2026, 98% had an initial AI strategy, but only 39% said top management actively steers it, and 37% had clearly defined responsibilities and governance. For a Mittelstand company, a strategy that fits on a few pages is more useful than a thick deck. It should answer:

• Which business goals AI should serve: fewer routine hours, faster replies, fewer errors, new services.
• Which processes are candidates, scored on the same questions, and which one goes first.
• Which data may go where: what may use a cloud service in an EU region, and what must stay in your own cloud account or on your own servers.
• Which shared foundations you build once, such as an LLM gateway, one controlled access point to all models with keys, budgets and logs.
• Who is responsible for AI overall, who owns each system, and which rules staff follow.

An AI roadmap then puts the processes in order. Plan the first one in detail and keep the rest short, because what you learn in the first one will change the order of the others.

Who must be involved, and when?

Bring in the people who can stop a go-live at the very start. In a German company that usually means:

• Management, which sets the goal and signs the budget, and finance, which asks for the business case.
• The process owner and the people who do the work today, because they know the exceptions and will review the results.
• IT and IT security, for systems, permissions, hosting and operation.
• The data protection officer, for the legal basis, processing contracts and, where needed, a data protection impact assessment.
• The works council, where there is one: an AI system that records who did what can trigger co-determination under section 87(1) no. 6 of the Works Constitution Act (BetrVG).

These are not formalities. Among companies using AI, data protection requirements are the most named barrier (66%), and 42% name a lack of employee acceptance, according to Bitkom's 2026 survey. A speaker quoted in IHK Karlsruhe's magazine in 2026 put it this way: "Wer nur die Technologie implementiert, aber die Belegschaft nicht mitnimmt, wird scheitern", whoever implements only the technology without bringing the workforce along will fail.

We bring the technical side: what is stored, where data is processed and what the works council will be able to see. Your company decides, advised by its data protection officer and lawyers, and nothing here is legal advice. The guide on works councils and AI system logging shows the configuration in detail.

AI strategy and rolloutUse casesFor managing directors