Short answer: Budget roughly 15 to 25 euros per user per month for licences. But the licence is rarely the biggest item: rollout, training and ongoing support usually cost more in year one than the software itself. Plan for both together.
Why the licence question is misleading
The cost of AI is almost always framed as a licence question. What does one seat cost, times headcount, done. That calculation is never right, and it is wrong in both directions: it underestimates the rollout effort and overestimates how many full seats a business actually needs.
The current data suggests many companies are stuck exactly here. In its study on AI in Germany, Bitkom finds that 36 percent of companies with 20 or more employees use AI, up from 20 percent the year before; in the follow-up survey of 14 September 2026 (603 companies with 20 or more employees) it is 57 percent. The same report notes that companies use only two AI applications on average. The licences have been bought, but the breadth is missing. And breadth does not come from more licences, it comes from onboarding.
What licences actually cost
| Product | Billed annually | Billed monthly | Prerequisite |
|---|---|---|---|
| ChatGPT Business | USD 20 per user/month | USD 25 per user/month | from 2 users |
| ChatGPT Plus (individual) | no annual plan | USD 20 per month | none |
| Microsoft 365 Copilot Business | EUR 15.60 net (promotional, list EUR 18.20) | EUR 21.84 net | Microsoft 365 Business plan required |
| Claude Team, standard seat | USD 20 per seat/month | USD 25 per seat/month | 2 to 150 users |
| Gemini in Google Workspace | included in all paid tiers, limited in Business Starter | Workspace subscription | |
| ChatGPT Enterprise | no list price, quoted by sales | negotiation | |
Two things stand out. First, full seats from the major vendors sit remarkably close together, mostly around 20 units per user per month. Second, monthly billing is consistently and substantially more expensive. If you are confident in your choice, the annual contract saves roughly a fifth.
The item almost everyone forgets: the rollout
A licence creates no value. Value appears when people know what they are allowed to use the tool for and what not. That is precisely where it stalls. The ifo Institute reports for May 2026 that 51.2 percent of solo entrepreneurs and micro businesses use AI, up from 30.4 percent, and that just under 59 percent pay for AI services. The tools are there. Whether they are used systematically is another question.
A clean rollout realistically involves: selecting and testing tools, a written usage policy, a data protection assessment, basic training for everyone and deeper training for the people who set the tone. Then ongoing support, because models, features and prices change faster than any training deck.
Worked example: 25 employees, 12 licences
A service business with 25 employees equips 12 people with Microsoft 365 Copilot Business at the promotional price of EUR 15.60 net. Internal costing rate: EUR 65 per hour.
| Item | Year 1 | Year 2 |
|---|---|---|
| Licences (12 x EUR 15.60 x 12 months) | EUR 2,246 | EUR 2,246 |
| Rollout project (20 hours) | EUR 1,300 | – |
| Basic training (25 people x 2 hours) | EUR 3,250 | – |
| Ongoing support (3 hours per month) | EUR 2,340 | EUR 2,340 |
| Total | EUR 9,136 | EUR 4,586 |
Licences account for only about 25 percent in the first year. That is the real message of this table. Budget for licences alone and you have failed to budget for three quarters of the cost, then wonder later why the rollout quietly died in day-to-day business.
The counter-calculation is encouraging, though. Recovering EUR 9,136 at EUR 65 per hour requires saving about 141 hours a year. Spread across 12 licence holders and 220 working days, that is roughly 3.2 minutes per person per day. In practice that threshold is almost always cleared as soon as anyone uses the tool regularly. The risk is not the business case, it is the adoption rate.
The cheaper route: not everyone needs a full seat
For recurring, clearly bounded tasks the API is often far cheaper than a seat. Mid-tier models currently sit around USD 2 per million input tokens and USD 10 to 12 per million output tokens; smaller models cost a fraction of that. Anthropic works through an example in its own documentation showing that processing 10,000 support tickets with a small model comes to roughly USD 37.
For a business that means the suggested support reply, the pre-capture of receipts or the quotation summary do not need to run through twelve full seats. They run through an automated process costing a few euros a month. We show how that is built in our overview of AI automation use cases for SMEs.
What do other companies spend?
Solid budget figures for German mid-sized companies are scarce, and the available studies are not comparable because they survey different populations. An overview:
- Bitkom (n = 604, companies with 20+ employees): 36 percent use AI (surveyed in 2025); 29 percent are increasing their AI budget, 60 percent keeping it flat. In the follow-up survey of 14 September 2026 (n = 603) it is 57 percent.
- Federal Statistical Office, ICT survey 2025: 26 percent of all companies use AI; 23 percent among those with 10 to 49 employees, 57 percent from 250 employees up.
- KfW Research, Fokus Volkswirtschaft No. 554, 29 July 2026: 20 percent of mid-sized companies use at least one AI technology; what matters is not company size but the level of digitalisation.
- ifo Institute, May 2026: 54.5 percent of companies use AI in business processes.
The spread from 20 to 54 percent mostly shows how differently AI usage gets defined. For your own planning, benchmarking against the average is of limited value anyway. The more useful question is which of your processes eats how many hours.
Public funding: what is still running
Germany’s go-digital programme ended on 31 December 2024 and no longer accepts applications. Two routes remain open.
First, KfW’s ERP digitalisation loan. It is aimed at freelancers, sole traders and SMEs, has no minimum amount, and depending on the tier offers a repayment grant of 3 or 5 percent of the disbursed loan, capped at EUR 200,000. Integrating artificial intelligence into value-creating parts of the business is explicitly eligible, as is related staff training. Important: the application goes through your bank and must be filed before the project starts.
Second, the Mittelstand-Digital centres and their AI trainers. Those offerings are free of charge and vendor-neutral. For a first step without a budget, that is the obvious route.
Frequently asked questions
Is the free version not enough?
For individual research, yes. As soon as company data is involved, no. Business plans differ mainly on whether your inputs are used for training and whether there is any administration. That is a data protection matter, not a comfort feature.
Should we pay annually or monthly?
Monthly for the first quarter, annually after that. The annual contract saves around 20 percent, but the saving is worthless if you realise after two months that the tool does not fit.
How many licences do we really need?
Fewer than you think. Start with the roles that write, research or analyse a lot. Four to six convinced users deliver more than twenty indifferent ones. You can always add seats.
What does a custom AI assistant on our own data cost?
That depends entirely on the state of your data and cannot be sensibly quoted as a flat figure. As an order of magnitude for a first, tightly scoped project, many companies budget a low five-figure sum, with data preparation usually taking more time than the technology.
Are there hidden costs?
The most common ones: a required base licence nobody accounted for, prices quoted net of VAT, exchange rate exposure on US dollar billing, and surcharges for regional data processing. Check those four before you sign.
Conclusion
AI is cheaper to buy than most people expect and more expensive to introduce. Treat the licence as the smaller part of a budget that explicitly includes training and support, and check which tasks need an automated process rather than a seat. If you would like to run the numbers for your own business, start with the AI check, look at our AI tools services and managed service, or get in touch. For the selection itself, choosing the right AI stack helps; for the rollout, rolling out ChatGPT in your company.
