Tried out is not the same as introduced
Most management teams tested AI long ago. The open question is what comes after the test.
A chat licence, a pilot in one department, a workshop with the leadership team. That is trying things out. AI only counts as introduced once a process runs on it regularly, someone is responsible for it and the result is measured.
Nobody knows exactly how many companies have taken this step. You often read that only one in five mid-sized companies uses AI. The figure of around 20% comes from KfW: according to KfW Research (11 February 2026), around 20% of mid-sized companies used AI in the period 2022 to 2024, roughly 780,000 businesses. Every form of use is counted, not only productive use.
The industry association surveys measure differently. Bitkom reported in September 2025 that 36% of companies with 20 or more employees use AI, up from 20% the year before. Another 47% are planning or discussing its use.
None of these sources separates "tested" from "embedded in operations". IW Köln (the German Economic Institute) gives a qualitative pointer (more on that shortly). We see the gap between testing and operations again and again.
That does not make mid-sized companies backward. They are cautious. A company without a data team and without spare IT capacity does not bring every new tool into operations just because it is available. It only becomes a problem when caution turns into standstill because nobody knows what the next step looks like.
What "actively using" AI means in operations
"A company uses AI" says little about how deep that use goes. IW Köln (IW-Report 33/2025, 4 July 2025) describes use as mostly selective, limited to individual parts of the business and largely based on free tools. Overall, the institute rates use as rather superficial.
Our working definition: a company uses AI actively when at least one business process runs on AI regularly, has a named owner and has its result measured.
Three levels make the difference visible:
- Tried out: a few employees have access to a chat tool and occasionally use it to summarise documents. Nobody knows exactly who uses it and how.
- Pilot: a team tests drafting quotes for a few weeks. Afterwards there is a presentation but no decision.
- In operation: incoming enquiries are pre-sorted with AI every Monday. One person in internal sales is responsible and checks the results. What may go out without approval is defined. Lead time and error rate are tracked.
The tool can be the same in all three cases.
Three questions for a self-assessment:
- Is there a process that runs on AI every week without depending on individual enthusiasts?
- Can you name one person who is accountable for the result and for approvals?
- Do you know how long the process took before and how long it takes today?
Three times yes: you are running AI. Otherwise you are in the testing phase. That is no failing. Only the next step is missing.
The three hurdles according to Bitkom and DIHK
Why does AI not get past the test in many companies? Technology is rarely the reason given. It comes down to people, time and rules.
Bitkom (press release of 15 September 2025, 604 companies with 20 or more employees). Three obstacles are almost level (source):
- legal hurdles and uncertainty: 53%
- lack of technical know-how: 53%
- lack of staff resources: 51%
Data protection requirements follow close behind (48%). This is not a study of mid-sized companies alone. For businesses with 50 or 200 employees, the figures will probably still sound familiar.
DIHK digitalisation survey 2026 (4,686 companies, surveyed in November 2025). The DIHK (Association of German Chambers of Industry and Commerce) asks about the challenges of digitalisation as a whole, not only AI. 58% of companies name time, 56% complexity, and just under a third legal uncertainty (source).
Put side by side, three hurdles emerge:
- Competence. Bitkom: "lack of technical know-how". In practice this means nobody in the company can judge with confidence which process is suitable for AI.
- Capacity. Bitkom: "lack of staff resources"; DIHK: time as the biggest challenge. The people who know the processes are tied up in day-to-day business. There is enough time for a pilot, but often not for routine operation.
- Law and rules. Tied for first place in the Bitkom survey; in the DIHK survey, named by just under a third and referring to digitalisation in general.
And the costs? Many people expect them on this list. They are on it, but not at the top: Bitkom puts lack of budget at 36%, the DIHK puts money at 40%, in each case well behind the hurdles named first. Bitkom's spending plans point to restraint rather than rejection: among the companies that use, plan or discuss AI, 29% want to increase their AI spending, 60% keep it stable and 5% cut it. That is why the cost question belongs at the end of the chain.
Why the order matters
In the surveys, the hurdles sit side by side. In operations, they depend on each other. What follows is our own reasoning.
First route: you start with the cost question
Many management teams start with a cost-benefit calculation. That is understandable. The costs are quick to put a figure on: licences, implementation, internal hours. For the benefit there is no basis, because nobody in the company knows which process is suitable and what would actually change.
So the calculation runs on cautious assumptions. Cautious assumptions set against fixed costs usually produce "not worth it" or "not yet".
Something else weighs more heavily: the calculation touches none of the three hurdles. Competence is still missing. Capacity remains unplanned. And the legal question stays abstract, because nobody can say which data should go into which system.
Second route: you start with competence
Two or three people who know their processes well learn to work with AI first. They notice where AI saves time in their work and where it makes mistakes. At the end there is a short list of concrete use cases.
That makes the capacity question something you can plan: fixed hours for a specific process instead of "time for AI". For that process, the benefit can be measured. The legal question becomes concrete as well: specific data, a specific system, one responsible person. And "Is AI worth it?" becomes "Is this use case worth it?". You answer that question with your own figures.
Part of the legal uncertainty comes from missing competence. If you don't know what a tool does with your own data, you cannot assess the legal position either. This gives the chain:
Competence in the team (step 1) → a measured workflow that fits the available time (step 2) → law and approvals as concrete detailed questions (step 3) → cost decision based on your own figures
If you start with the costs, you are assessing something that does not yet exist in your own company.
Step 1: Build competence before buying tools
The obvious first step is a licence agreement. We recommend a different one: choose two or three people who know your processes in detail. Rarely the IT managers. Usually the people who write quotes, handle complaints or sort enquiries today.
Why so small? Because missing know-how and missing time come together. A large training programme is meant to solve the first hurdle and makes the second worse. A small team learning on a real process ties up little capacity. Give it fixed hours in the calendar, otherwise learning stays a side project.
The team learns on your own material, at first only with non-critical content such as public product information or anonymised examples.
The three-folder principle
You don't need a platform to get started. A simple structure is enough, which the team sets up alongside its first attempts:
- Business context. Customers, offering, tone of voice, internal rules. The knowledge a new colleague takes in during her first weeks. Without this context, AI gives generic answers instead of answers for your company.
- Data. The documents a process works with: price lists, product sheets, past quotes, typical enquiries. This is where gaps in data quality quickly show. Eight questions on this are asked in the checklist of eight data readiness questions.
- Work instructions. The steps a person carries out today: what comes in, what gets checked, what goes out, who approves? This description is the template an AI will later work from.
Whoever fills the folders has to describe a process precisely enough for an outsider to understand it. This is exactly how you build the competence that is missing when a tool is simply bought. And the structure is not tied to any tool. If you change systems later, the context, data and instructions stay with you.
Competence and the AI Act
The AI Act (Article 4) addresses the topic too. According to the text of the regulation, the article on AI literacy has applied since 2 February 2025. What it means for your company in concrete terms is for your legal counsel to clarify. For operations, in our view, something else counts: a deliberately built core team is the fastest way to bring competence into the company. How governance and law run alongside is covered in step 3.
Step 2: The first workflow where the benefit can be calculated
The first productive use case decides whether AI is seen in the company as a working tool or as an experiment. Don't choose it by the biggest conceivable impact. Choose it by whether the benefit can be shown and whether your team has the time.
Five criteria for the first workflow
- Frequent. Daily or several times a week. Something that happens once a quarter produces too few runs.
- Rule-based. The rules can be written down. If your most experienced employee decides purely on gut feeling, the process is too open to start with.
- Measurable. You know how long a case takes and how often it needs reworking. If you don't have the figure, collect it before you start.
- One responsible person. Named, and in charge with AI and without it. A committee is too slow.
- Clear approval level. Set in advance: what may the AI do on its own, and what does a person approve? To start with, we recommend the most cautious level. The AI prepares, a person checks.
Plan capacity before the workflow starts
Many pilots deliver a usable result, and afterwards nobody has time for routine operation. Plan in advance who spends how many hours a week on checking, corrections and maintaining the instructions. At first, checking costs time before it saves time. What else has to be settled on the way into routine operation is listed in the handover checklist from pilot to routine operation.
- Start with a single workflow.
- Reallocate time instead of piling it on. Decide which task will be left undone in return.
- If it doesn't fit, choose something smaller. Better one sub-step than a postponed start.
A typical pattern
Incoming enquiries are often a good first workflow. Today, one person reads every enquiry, assigns it, adds information from the CRM and forwards it. With AI, the system proposes a category, the responsible team and a draft reply. The responsible person checks and approves. The assignment rules sit in the work instructions folder.
This fits what companies already use AI for. In the DIHK digitalisation survey 2026, the companies using AI most often name, after generative AI for text, images or code (78%), personalised customer outreach and customer support (43%).
Measure instead of estimating
A return-on-investment calculation before the start rests on assumptions that are hard to justify. A before-and-after comparison on three values is more reliable:
- Time per case: how many minutes of work go into one enquiry?
- Lead time: how long from receipt to reply?
- Errors and rework: how often is something assigned wrongly or corrected?
Collect the values for a few weeks before the start, then over the same period with AI. Count the checking time honestly. What you get is a figure from your own company rather than an extrapolation.
What IW Köln says about productivity
Every management team asks: how much more productive does AI make us? IW's answer is sober. In its expert report "How will AI change productivity in Germany?" (18 March 2025), it projects productivity growth of 0.9% a year for 2025 to 2030 and 1.2% for 2030 to 2040. That is no productivity miracle, and according to IW the gains do not arrive automatically.
There are three limits you should know:
- A projection. The values are a model calculation for the economy as a whole. They say nothing about your company.
- Self-assessment. In the DIHK survey, 41% of companies using AI rate the impact on their productivity as high. That is the companies' own estimate rather than a measurement.
- Correlation is not causation. According to IW-Report 33/2025, 66% of large companies use AI and 36% of small ones. The figure does not explain why. Well-organised companies are probably more likely to introduce AI. No survey can separate whether AI makes them more productive or whether they already were.
The only reliable figure for your company is your own before-and-after value from step 2.
Step 3: Let governance and law run alongside
In the Bitkom survey, legal uncertainty ties for first place among the obstacles. The concern is justified. But waiting until every legal question is fully settled means not starting. Governance does not have to be finished before the first workflow. It grows with it, from the start and on a small scale.
For the first use case, three decisions on one page are enough:
- Who approves? One named person decides. At the start the rule is: a person sees everything before it leaves the company.
- Which data may go into which system? A list is enough: non-critical, personal, confidential, and which tool may process what. This also answers the most common question in the team: "Am I allowed to enter this here?"
- Who keeps the record? Write down which tool runs in which process, who is responsible and which approval rule applies.
That is operational organisation. It also makes conversations with data protection officers or the works council more concrete: you talk about a process with known data and one responsible person, instead of about "AI" in general.
This page does not replace a legal review. Our deployer check shows how we go through a company's use of AI.
Consulting as a shortcut instead of another cost risk
If you start with competence, you soon run into the capacity question. External support can close this gap. It can also widen it: if the end result is a workflow that only the consultants can operate. Then you have shifted costs and built no competence.
The yardstick: support pays off when it saves your team learning time and the knowledge stays with you afterwards. You can recognise this before you hire anyone by three points:
- A clearly limited scope. Start, end, budget, for example one process with approval rules. Open-ended hour allowances without a goal are a warning sign.
- Your team works on it from the start. Context, data sources and work instructions are documented well enough that your team can keep running the workflow on its own.
- An agreed measuring point. A baseline before the start, the same value after a few weeks. If nothing shows, raise it before any more budget is spent.
Not every company needs support. If you have people with time and know-how, a structured start following the steps above is often enough. Your team has to walk the path either way.
What runs differently on Monday morning
Four steps are enough to get started, and they fit into 30 days:
1. Name the owners and let them learn (week 1). Two or three people who know the process. One of them is accountable for the result. Fixed hours in the calendar. From the first week, the team works with an AI tool, at first with non-critical content, and sets up the three folders.
2. Choose one process (weeks 1 to 2). Only one. Frequent, rule-based, measurable, and matched to the time you have freed up. The team describes it in the work instructions folder.
3. Measure the baseline (weeks 2 to 3). Time per run, runs per week, rework. Two weeks of tally marks are enough.
4. Set the approval rules and start the first runs (weeks 3 to 4). Write the page from step 3. After that, the first real cases run with AI, each one checked and approved.
The weeks that follow give you the reliable comparison, measured over the same period as the baseline.
It looks unspectacular: one process has a responsible person, and the result is measured. That is exactly the difference between testing and operations.
If you would like to assess beforehand where your company stands, use our readiness check or book a no-obligation intro call. The four steps also work without us.
Which process in your company could be the first candidate next Monday?
Frequently asked questions
How many mid-sized companies actually use AI productively? Bitkom reports 36% AI use for 2025 among companies with 20 or more employees; KfW reports around 20% among mid-sized companies for any use of AI from 2022 to 2024. None of these figures separates testing from operations.
Why should we start with competence and not with the cost calculation? Without experience in the company, any cost calculation rests on assumptions and solves none of the hurdles named above. With competence, you end up calculating with your own figures.
Do we need a data team or an IT department to get started? No. To get started, two or three people who know the processes and get fixed time for it are enough. Time is the bottleneck: Bitkom reports a lack of staff resources as an obstacle at 51% of companies.
Where can we find reliable information on the AI Act? In the text of the regulation on EUR-Lex. According to the regulation, Article 4 on AI literacy has applied since 2 February 2025. Which rules apply to your use depends on its purpose and is a matter for your legal counsel.
When does external support pay off when getting started with AI? When it saves your team learning time and the knowledge stays in the company afterwards. A clearly limited scope and an agreed measuring point show this before you hire anyone.
This article describes our practical perspective. It does not contain a legal assessment; for that, we refer you to the linked primary sources and your legal counsel.
Sources
- Bitkom: Durchbruch bei Künstlicher Intelligenz (in German), press release, 15 September 2025
- KfW Research: Press release on AI use in mid-sized companies (in German), 11 February 2026
- DIHK: Digitalisation 2026: Businesses stay the course, 28 January 2026
- DIHK: Digitalisierung 2026: Unternehmen halten Kurs (in German), 28 January 2026
- DIHK: DIHK-Digitalisierungsumfrage 2026 (in German), results (PDF), January 2026
- IW Köln: Künstliche Intelligenz als Wettbewerbsfaktor für die deutsche Wirtschaft (in German), IW-Report 33/2025, 4 July 2025
- IW Köln: Wie wird KI die Produktivität in Deutschland verändern? (in German), expert report, 18 March 2025
- Regulation (EU) 2024/1689 (AI Act): EUR-Lex