Whether AI pays off for companies is hardly the question any more in 2026. According to Bitkom, 57 per cent of companies in Germany with 20 or more employees now use artificial intelligence. The open questions are where it delivers, what getting started costs and who in the team can work with it confidently. If you only buy and distribute a licence, you have a tool in-house, but not yet a use case.
This article is for everyone who wants to use artificial intelligence in their company and is responsible for turning licences into real usage. It sorts ten use cases by entry level and metric. It shows where the costs of AI solutions actually arise, what Article 4 of the EU AI Act requires after the Digital Omnibus and in what order to introduce AI in your company.
AI in the company in 2026: most use it, hardly anyone exploits it
Adoption is no longer an open question. In the Bitkom survey of September 2026 (603 companies in Germany with 20 or more employees, surveyed in July and August 2026), 57 per cent use AI. A year earlier, the figure was 36 per cent, two years earlier 20 per cent. A further 38 per cent are planning or discussing its use; only 4 per cent say AI is not an issue for them.
Depth is a different story. No company using AI says it exploits the potential fully, and only 3 per cent say they use it to a fairly large extent. 28 per cent make little use of it, 59 per cent none at all. AI is most often used in customer contact, for example for handling enquiries (72 per cent), and in marketing and communication (54 per cent). Internal knowledge management, controlling and sales have grown significantly.
Where AI is running, it pays off: 66 per cent of users see their competitive position improved, 48 per cent have significantly accelerated internal processes. The gap lies in enablement. 66 per cent of companies rate their employees' AI skills as low. That is expensive, because untrained use rarely gets beyond trial and error. And it is risky, because AI literacy is a legal requirement.
What “using AI in the company” actually involves
Behind “using AI” are four components that belong together. Naming them individually shows where the work begins:
- Use cases: defined tasks in which AI makes a measurable contribution. Without them, the deployment remains a solution looking for a problem.
- Enablement: employees who know when a tool is suitable, how to use it and where its limits lie. This is the part that is most often missing.
- Tools and integration: the specific AI solutions and their connection to existing systems and data.
- Framework: rules for handling data, approvals, documentation and responsibilities.
The four tracks run in parallel. An introduction that only serves the third is the standard way to fail: licences keep running, usage trickles away. The value is created at points 1 and 2; purchasing comes afterwards.
10 use cases: where AI delivers for companies today
The most common order in practice is the wrong way round: first select a tool, then look for uses. The reverse makes sense. A use case is suitable as a starting point if three conditions apply. The task is recurring, its effort is measurable, and the consequences of an error are limited or can be controlled through an approval step.
The ten cases below usually meet these criteria. Each is assigned to one of three entry levels. Level 1 runs with standard tools such as ChatGPT, Microsoft Copilot or Google Gemini in the browser, without integration. Level 2 is a pilot connected to your own data. Level 3 is an integrated solution in production. What each level costs and how long it takes is covered in the section on costs.
| Use case | Entry level | Metric |
|---|---|---|
| 1. Customer enquiries: reply drafts and chatbot | 1 (drafts), 3 (chatbot) | Handling time per enquiry |
| 2. AI phone assistant | 3 | Share of calls resolved without a callback |
| 3. Knowledge assistant with RAG | 2 to 3 | Search time per question |
| 4. Sorting and answering the inbox | 1 | Handling time per email |
| 5. Reading documents and invoices | 2 to 3 | Share of documents without manual correction |
| 6. Quotations and tenders | 1 to 2 | Preparation time per quotation |
| 7. Marketing copy and variants | 1 | Time to an approved draft |
| 8. Meeting minutes and tasks | 1 | Follow-up time per meeting |
| 9. Workflows between systems | 2 to 3 | Lead time per process |
| 10. AI agents | 3 | Share of processes completed without intervention |
1. Customer enquiries: reply drafts and chatbot
According to Bitkom, customer contact is the most common area of use. The low-risk starting point is reply drafts for standard enquiries that an employee checks and approves. The approval step catches errors; the metric is handling time per enquiry. The next stage is a customer service chatbot that answers recurring questions around the clock and hands over to a human when it is unsure. It needs a well-maintained knowledge base; otherwise it answers wrongly instead of not at all.
2. Phone: AI phone assistant
An AI phone assistant takes calls, answers standard questions, records requests and forwards them. It fits where many similar calls come in, for example about opening hours, appointments or order status. The metric is the share of calls resolved without a callback. The cost driver is integration: the phone system, calendar and CRM need to be connected. The article Voicebots for businesses compares providers, costs and data protection.
3. Internal knowledge: knowledge assistant with RAG
According to Bitkom, the use of AI in internal knowledge management doubled within a year, from 11 to 22 per cent. A knowledge assistant answers questions from manuals, contracts or tickets and cites the source. Technically, it is usually based on retrieval augmented generation (RAG): the model first searches your documents and then formulates the answer. The critical factor is the data connection rather than the model. The metric is search time per question. The RAG knowledge assistants page shows how they are set up and how they work.
4. Inbox: sorting and answering emails
The inbox is the easiest way into level 1. AI summarises long emails, sorts them by urgency and writes reply drafts in your tone. This needs no integration project, just a template, a good prompt and a cross-check by a human. Customer data belongs in an approved company tool and never in a private account. The metric is handling time per email. A step-by-step guide with a sample prompt is in the whitepaper 5 AI workflows for SMEs (in German).
5. Documents and invoices: extracting data
In controlling and accounting, AI use rose from 17 to 25 per cent according to Bitkom. The typical case: reading incoming invoices, delivery notes or forms, checking fields and passing them to the ERP. Intelligent data extraction replaces manual typing; unclear documents are still checked by a human. The metric is the share of documents processed without manual correction. For larger volumes, level 3 is an integrated document processing solution with an approval step.
6. Sales: quotations and tenders
In sales, AI use rose from 5 to 14 per cent according to Bitkom. Obvious applications are first drafts of quotation texts and summaries of long tenders: which requirements are mandatory, which deadlines apply, where is information missing? This saves preparation time; price and commitment remain human decisions. At level 1, a standard tool with a template is enough. At level 2, the AI also accesses previous quotations and product data. The metric is preparation time per quotation.
7. Marketing: copy and variants
Marketing and communication is the second most common area of use at 54 per cent. AI delivers drafts for newsletters, social media posts, product copy and ad variants. Facts, tone and approval stay with humans, because AI writes drafts, not finished results. The metric is the time to an approved draft. The article AI in marketing 2026 shows which marketing automations can be set up quickly.
8. Meetings: minutes and tasks
Writing minutes and following up on tasks takes time in every team. AI turns notes or transcripts into a summary with decisions, open points and owners. Clarify beforehand whether conversations may be recorded, and make sure all participants know. The metric is follow-up time per meeting. The workflow runs at level 1 with Copilot, ChatGPT or Gemini and is also described in the whitepaper.
9. Process automation: workflows between systems
As soon as a task touches several systems, for example a form, the CRM and email, the individual case becomes a workflow. AI automation connects the steps and uses the language model where text needs to be understood or generated: classifying the enquiry, extracting data, preparing the reply. The metric is lead time per process. The article AI automation for SMEs explains where this pays off immediately. Solutions like this are implemented as process automation with AI workflows.
10. AI agents: the advanced stage
AI agents go one step further than a chatbot. They act independently, obtain quotations, compare prices or prepare decisions. According to Bitkom, among companies that use AI or are planning or discussing its use, 11 per cent already deploy agents and 29 per cent are planning to. An agent needs reliable data, connected systems and clear guardrails. It is therefore rarely suitable as a starting point. It builds on cases 3, 5 and 9; more on this under AI agents as digital employees.
What is not suitable as a starting point is just as important: tasks with a high potential for damage if errors occur, legally sensitive decisions about people, and processes without a measurable result. If you start there, you risk an expensive lesson and burn internal acceptance for the next round. The right cases can be selected systematically with an AI use case identification.
AI solutions for companies: standard tool, integration or own model
Which AI solution fits follows from the use case. There are three routes:
- Standard tools such as ChatGPT, Microsoft Copilot or Google Gemini cover level 1. They are available immediately and work with what you put in and, depending on the licence, with the data sources the provider connects itself, such as emails and files in Microsoft 365 or Google Workspace. Without your own integration, they cannot reach the ERP, the phone system or the logic of your business processes. What matters are business licences with proper data protection terms and clear rules on which data may go in.
- Integrated solutions connect a large language model (LLM) via an interface to the CRM, ERP, phone system or document storage. This is the norm at levels 2 and 3: chatbot, knowledge assistant, document processing, workflow.
- Self-hosted models such as Llama or Mistral pay off when strict data protection requirements demand it or when, at high and sustained volumes, token costs justify your own hardware. The article on digital sovereignty describes what a European stack looks like.
The same applies to data protection on all three routes: personal data needs a legal basis, and with an external provider you usually need a data processing agreement. The article GDPR and AI shows what this looks like in practice with ChatGPT and other tools.
What does AI cost in a company?
The answer depends less on the model than on data and integration. In the Bitkom survey, AI users mainly cite infrastructure such as cloud and computing power (51 per cent), data preparation (50 per cent) and integration into existing systems (41 per cent) as major cost items. Subscriptions and licences are a major cost item for only 21 per cent, on a par with external consulting. Training is cited by 19 per cent, token consumption by 8 per cent.
Bitkom President Ralf Wintergerst sums it up as follows (our translation): “If you only buy a licence, you have bought AI, but you haven't changed anything yet.”
For planning, this means that a level 1 use case mainly costs licences and training time. From level 2, the data situation, depth of integration and target system determine the price. Production solutions add running costs, namely API and token costs for the language models used, plus operation and monitoring. Companies are investing nevertheless: according to Bitkom, 45 per cent of companies are spending more on AI in 2026 than in the previous year.
| Step | Project type | Typical range | Duration |
|---|---|---|---|
| Upfront | AI workshop and use case assessment | low four-figure range | 1 to 2 weeks |
| Level 1 | Standard tool in the browser, one workflow | Licence per user per month (examples below the table) plus training time | one test week per workflow |
| Level 2 | Pilot project / proof of concept | from a few thousand euros | 2 to 6 weeks |
| Level 3 | Production use case (chatbot, voice agent, automation) | low to mid five-figure range | 6 to 12 weeks |
| Ongoing | Operation and further development | monthly according to usage: API and token costs plus monitoring | ongoing |
Two German list prices per user per month, excluding VAT and time-limited promotional discounts (as of 10 October 2026), show what the level 1 licence costs. Microsoft 365 Copilot Business costs €18.20 with annual billing as an add-on to a Microsoft 365 Business plan. At Google, Gemini is part of the Workspace plan itself: in Gmail, Docs and Meet from Business Standard at €13.60 with an annual contract, including the email mailbox and storage. Microsoft 365 Copilot Chat is included at no extra cost in eligible Microsoft 365 subscriptions.
The ranges for workshop, pilot, production use case and operation are taken from nextlevels' cost information on AI consulting and on the AI agency (as of July 2026). They are a market-standard guideline, not a price list. What a project actually costs depends on the data situation, depth of integration and target system. Before the rollout, it therefore makes sense to run a proof of concept that shows with your data whether a use case works before the larger budget is committed.
If an AI function is to become part of your own software, the project calculator gives you a first order of magnitude. A reliable fixed price is set after the analysis, for example as part of AI consulting.
The obligation behind it: Article 4 of the EU AI Act after the Digital Omnibus
Article 4 of the EU AI Act has applied since 2 February 2025. It is addressed to providers and deployers of AI systems. Under Article 3, a deployer is any person or body using an AI system under its authority, except where it is used in the course of a personal non-professional activity. That includes the SME with Copilot licences. The obligation applies regardless of the risk class, i.e. also well below the level of high-risk AI systems.
With the Digital Omnibus, the EU has rewritten the article. Regulation (EU) 2026/1744 of 8 July 2026 was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. Instead of “ensuring” a sufficient level of AI literacy, providers and deployers now take measures to “support the development” of their staff's AI literacy. Nobody has to guarantee a specific level of AI literacy; the new text states this explicitly.
The Commission and the Member States are to support providers and deployers, in particular SMEs, in doing so. The major postponements in the package concern the high-risk obligations: for systems under Annex III, they apply from 2 December 2027, for systems under Annex I from 2 August 2028. Article 4, by contrast, already applies.
Three points are often reported inaccurately. First, Article 4 prescribes no training format and no certificate. According to the European Commission's AI literacy FAQ, an internal record of training and other measures is sufficient. Second, national market surveillance authorities have been supervising compliance since August 2026. In Germany, under the KI-MIG (the German AI market surveillance act, in force since 29 July 2026), this is in principle the Federal Network Agency (Bundesnetzagentur). Third, the catalogue of fines in Section 15 KI-MIG contains no offence relating to Article 4.
The risk is therefore indirect, but real. If damage occurs due to a lack of AI literacy, for example through unchecked output or leaked customer data, the lack of enablement falls back on the company. If you train on the use case and document it, you meet both the old and the new version. The article EU AI Act for SMEs summarises what else the AI Act requires in 2026.
Getting started: introducing AI in the company in five steps
A structured start follows a sequence, because later steps build on earlier ones. The order is part of the method:
- Take stock. Where is AI already being used, including unofficially via private accounts (shadow AI)? That is the real starting point. An AI potential analysis combines this inventory with an assessment of your processes and takes one to three weeks, depending on the size of the company.
- Prioritise use cases. Select two or three cases from the list above by frequency, measurability and consequences of errors. Better a few that deliver than a broad catalogue that commits nobody.
- Set the framework. Which data may go into which tools? Who approves? How is usage documented? These rules are the core of your AI governance and at the same time fulfil part of the Article 4 obligation.
- Train on the use case. Not everyone needs the same thing (see the next section). Training starts from the prioritised case, with the tools and data the team will work with the next day.
- Measure and decide. Every use case gets a metric and a deadline in advance. Then you decide: roll out, refine or replace.
The most common mistake in this sequence is skipping step 3. Rolling out without a framework produces exactly the uncontrolled use that Article 4 addresses and merely postpones the governance problem. If you want support with steps 1 to 3, the process analysis and AI strategy provides the right framework.
What AI training needs to achieve
A lot has changed in training. According to Bitkom, 70 per cent of companies train their employees in using AI: 11 per cent train all of them, 22 per cent the majority and 37 per cent selected employees. 25 per cent offer no training at all, down from 43 per cent the year before. Among AI users, 91 per cent provide training. The open question is therefore mainly what training needs to achieve.
Effective AI training is not a one-off lecture on “What is AI?”. It covers three levels of competence, and not every role needs all three in the same depth:
- Basic technical understanding: what generative AI can and cannot do, why a language model produces errors (hallucinations) and what distinguishes a good prompt from a bad one. Everyone who uses AI needs this level.
- Legal and organisational framework: which data may be processed, what the internal framework stipulates and when a human remains responsible. Everyone who uses AI in a work context needs this level.
- Depth of application in the use case: hands-on use in the prioritised use case. This level is role-specific and belongs to the introduction itself, not to a general webinar.
Whether training works or is merely ticked off is almost always decided at the third level. General AI knowledge is quickly conveyed and just as quickly forgotten. Enablement comes from the specific use case. Formats for this are described under AI workshops and team enablement. Record the date, participants and content of every training session; that is your evidence for Article 4.
Frequently asked questions about AI in the company
How can companies use AI? Recurring tasks with measurable effort and controllable errors are best suited: answering customer enquiries, extracting data from documents, searching internal knowledge, drafting texts and minutes. Standard tools in the browser are the way in. Integrated solutions connected to the CRM, ERP or document storage follow once a pilot reaches its metric.
What does AI cost in a company? As a market-standard guideline (as of July 2026), a workshop with use case assessment is in the low four-figure range, a pilot starts at a few thousand euros and a production use case is in the low to mid five-figure range. Running API and operating costs come on top. According to Bitkom, the biggest cost items are infrastructure, data preparation and integration; licences play a smaller role.
Do employees have to be trained in AI? Anyone who uses AI systems professionally has had to take measures on their staff's AI literacy under Article 4 of the EU AI Act since February 2025. Since the Digital Omnibus, this means supporting the development of that literacy. Training is the usual route. According to the European Commission, no certificate is needed; an internal record is sufficient.
Which AI solutions are suitable for SMEs? To start with, standard tools such as ChatGPT, Microsoft Copilot or Google Gemini with clear data rules. For customer service, documents and knowledge search, integrated solutions that connect a language model to the CRM, ERP or document storage. Self-hosted models mainly pay off with strict data protection requirements or high, sustained volumes.
How long does it take until AI is running productively in a company? A workflow with a standard tool can be tried out within a test week. A pilot typically takes two to six weeks, a production use case with integration six to twelve weeks. How fast it actually goes depends mainly on the data situation and depth of integration.
Conclusion
Using AI in the company means bringing four things together: use cases, enablement, tools and framework, in that order of priority. The state of play in 2026 is clear. The majority use AI, hardly anyone exploits its potential, and since February 2025 AI literacy has been an obligation under Article 4 of the EU AI Act, which the Digital Omnibus has worded more softly but not abolished.
The pragmatic start is small: take stock, prioritise two or three use cases from the list, set a framework, train on the use case, measure. If you want to start right away without a project, you will find five ready-made workflows in the whitepaper 5 AI workflows for SMEs (in German).
If you want to know where AI has the biggest leverage in your processes, the AI potential analysis is the next step. AI consulting supports strategy, pilot and implementation. An AI agency brings use case experience from other mid-sized companies and prevents training and tool selection from drifting apart. The article Choosing an AI agency shows how to recognise a genuine implementation partner.