Strategy & Market · 2026-10-07 · 12 min read
What does AI implementation cost in the mid-market? The bill does not start at the licence
Michael Kaiser
Co-Founder & Head of Systems, Vincency
The short answer first: anyone asking what AI implementation costs is usually budgeting the wrong line item. The representative Bitkom survey of 603 German companies shows where the money actually goes: 51 percent of AI users name infrastructure as a major cost block, 50 percent name data preparation and 41 percent name integration into existing systems. Licences come fourth, at 21 percent. The expensive part is not the model, it is the plumbing.
This piece is written for the managing director who typed "KI-Implementierungskosten" into a search box expecting one number. There is no honest single number, but there is an honest structure: five lines on the bill, and a representative dataset showing which of them bites hardest.
What 603 companies say the money goes to
| Cost block | Share of AI users calling it a major cost | What is behind it |
|---|---|---|
| Infrastructure | 51 percent | Cloud, compute, hosting, the machines the model runs on |
| Data preparation | 50 percent | Cleaning, structuring and permissions on your own data |
| Integration | 41 percent | Interfaces into ERP, CRM, DMS and the grown system landscape |
| Licences and subscriptions | 21 percent | The seat price every vendor prints on the pricing page |
| External consulting | 21 percent | Day rates for assessment, architecture and rollout support |
| Training | 19 percent | Sessions, prompt libraries, documentation under Article 4 |
| Token consumption | 8 percent major, 11 percent minor | Metered usage, the smallest block in the survey |
The order of this table is the answer to the search query. The cost of AI implementation is not driven by the licence line every vendor advertises, but by three items no vendor puts on a price page: the compute behind the model, the state of your own data and the systems it has to reach.
The five lines on the real bill
An implementation that is allowed to touch company data and processes decomposes into five invoices, and each behaves differently. The assessment is the cheapest line and the only one that can be bought at a truly fixed price. Data preparation is the wildcard: it scales with the state of your shares, exports and archives, and every second AI user calls it a major cost for that reason. Integration is day-rate work, roughly 700 to 1,500 euros per day in the current German market, multiplied by however many systems have to answer when the model asks. The pilot is deliberately small. And the rollout carries two lines people forget: the training duty under Article 4 of the AI Act, and the running cost of a system that just became production.
The licence itself, the line everyone googles, sits inside this structure as the most predictable and least important number. A seat runs 10 to 99 euros per user and month across the documented market prices, or a subscription wrapper as covered in the flat-rate breakdown. Neither line decides whether the project lands. The three top rows of the table do.
Why 63 percent cannot calculate it, and how to fix that
The same Bitkom survey explains the query volume behind this article. Among companies not using AI yet, hard-to-calculate costs are the third most common barrier at 63 percent, behind missing technical know-how at 85 percent and legal uncertainty at 66 percent. The anxiety does not dissolve with adoption either: 54 percent of current AI users still name calculability as a problem. The market is not short of motivation, it is short of a trustworthy bill.
The fix is procedural, not financial. Costs become calculable when the first purchased artefact is an assessment with fixed scope: an inventory of systems, data sources and candidate use cases, priced as a defined package. From that document the integration work can be quoted against real interfaces instead of estimated against hope, which is exactly how the day-rate risk in the middle of the bill shrinks. Whoever skips the assessment and asks vendors for a total price gets either a guess or a price padded for the guess. Our comparison of mid-market consulting day rates shows what the individual lines cost; the assessment is what makes them add up predictably.
The context, and the honest remainder
Two reference points size the topic. Bitkom counts 57 percent of German companies already using AI, up from 36 percent a year earlier, while the official Destatis statistics for 2025 put German enterprise AI use at 26 percent overall and 36 percent for the 50-to-249-employee mid-market. Both can be true: Bitkom surveys companies with 20 or more staff and counts usage broadly, Destatis measures a stricter definition. The gap between them is roughly where the mid-market sits: past experimentation, not yet at systematic implementation.
Three qualifications close this out. First, the Bitkom figures are self-reported assessments by companies about their own spending, not invoices, and a named "major cost block" is a judgment, not a euro amount. Second, every euro figure here is a market anchor with a retrieval date, not a quote for your project; the only number that counts is the one after your assessment. Third, a calculable budget is not the same as a cheap project: the survey shows 45 percent of companies raising their AI spend in 2026, and the companies winning are the ones buying the plumbing, not the ones negotiating the licence.
Related service
IT strategy as a guided service
The analysis described here can be done in house. If you lack the time or the distance from your own systems, we take it on: inventory, a prioritised order and, if you wish, the implementation.
IT strategy consultingFrequently asked questions about the cost of AI adoption
What does AI implementation really cost a mid-market company?
The honest answer is a bill of five lines instead of one number: assessment, data preparation, integration into existing systems, pilot and rollout including training. A representative Bitkom survey of 603 companies shows where the money actually goes: 51 percent of AI users name infrastructure as a major cost block, 50 percent data preparation, 41 percent integration, only 21 percent licences. A budget cut to the licence line systematically underestimates the project.
Why are licences not the main cost item?
Because the licence is the tool, not the work. A seat costs between 10 and 99 euros per user and month today, readable on any vendor page. What no vendor puts on the pricing page: the cloud and compute behind it, preparing your own data so the model can use it at all, and the integration into ERP, CRM or DMS. Exactly these three items are named as major cost blocks by 41 to 51 percent of companies.
How can the budget be calculated at all?
Through fixed scope instead of estimation. 63 percent of companies without AI name hard-to-calculate costs as a barrier, and even among users the figure is still 54 percent. The project becomes calculable when the first phase is an assessment with fixed scope and fixed price, producing a defensible roadmap. Only then is integration commissioned, grounded in the assessed inventory rather than hoped-for figures.
Which cost items are forgotten most often?
Three: data preparation, named as a major cost block by every second user in the Bitkom survey and absent from every price list; the documentation and training duty under Article 4 of the AI Act, in force since February 2025; and the running costs after go-live, because a pilot that works turns into a product that wants maintenance.
Is a pilot enough to test the implementation cost?
For testing feasibility, yes; for calculating total cost, no. A pilot proves a use case works, but it usually runs on clean sample data without interfaces. The expensive items, data preparation and integration, only start when the pilot moves into regular operation. Reading the pilot as a cost test means budgeting with the cheapest phase of the project.
Sources, status and note: All figures were read on the primary sources, retrieved on 7 October 2026: the Bitkom press release of 14 September 2026 for the cost blocks (51, 50, 41, 21, 21, 19 and 8 percent), the adoption figures (57 percent users, 45 percent raising budgets) and the barrier figures (63 percent of non-users, 54 percent of users naming hard-to-calculate costs), from a representative survey of 603 companies with 20 or more employees; and the Destatis ICT statistics on enterprise AI use for the official German shares of 26 percent overall and 36 percent for the 50-to-249-employee class, reporting year 2025. This is orientation, not a quote; the only figure that counts is the one after your assessment.
Related insights