In short: Start with a single workflow that comes up often, takes a lot of time and produces a result a person can easily check. Test a solution for it for two to four weeks and decide beforehand how you will measure success. Rules for data protection, labelling and training are part of it from the start, because the transparency obligations of the EU AI Act have applied since 2 August 2026.
This is the question managing directors ask us most often today. This guide takes you from a concrete problem to a solution that works in day-to-day business.
How widespread is AI in Austrian companies?
Almost a third already use it. According to Statistics Austria (in German), 30% of companies in Austria with 10 or more employees used AI in 2025, while the EU average was 20%. Among companies with 10 to 49 employees, the share was 26%, according to the STATreport “ICT Usage in Enterprises 2025” (in German).
If you have not started yet, you are still in good company. Of the companies that do not use AI, 77% have not even considered it yet. The obstacles they name most often are a lack of in-house expertise (15%), data protection concerns (11%) and legal uncertainty (11%).
Why do AI projects fail most often?
Mostly because of people and processes, rarely because of the technology. In a 2024 survey by Boston Consulting Group of 1,000 executives from 59 countries, 74% of companies had yet to show tangible value from AI. Around 70% of the difficulties in implementation had to do with people and processes, around 20% with technology and around 10% with the AI algorithms themselves. The study surveyed executives worldwide, not small businesses specifically.
A false start begins with the sentence “We need some kind of AI now.” That is not a goal but the wish for a tool before the problem has been named. AI is not an end in itself. Sometimes it is the cheapest solution to an expensive problem. So do not ask where AI could be used. Ask which process costs you a disproportionate amount of time or money, even though it plays out in a similar way again and again.
How do I get started in practice?
Work through seven steps and keep the first attempt small.
- Find the problem. List the processes that come up often, run in a similar way each time and take a lot of time.
- Sort. Rate each candidate by value and feasibility. The value is the time or money the solution saves. A case is feasible if the data is available, the risk is low and the effort stays reasonable. Start with the case that scores well on both, and leave low-value cases aside for now.
- Choose the route. Decide whether to buy a ready-made tool, connect a language model or have something built.
- Clean up the data. Make sure the documents you need are current and well organised, and that people can get to them.
- Test small and measure. Run a pilot over two to four weeks and define the metrics beforehand.
- Set the rules. Settle data protection, labelling, training and review by a person.
- Assign responsibility and expand. Name one person who drives the topic, collects feedback and decides on improvements. You do not need a dedicated AI team for this. Expand only what has proven itself in the pilot.
Which task makes a good first use case?
A task that comes up often, runs in a similar way each time, involves language or text and produces a result a person can easily check before it causes any harm. A tool that makes a rare task easier saves hardly anything.
Good first use cases are drafting quotes, sorting emails and preparing replies, extracting data from invoices or delivery notes, first draft answers to customer enquiries, and summarising minutes or research. Not suitable as a starting point is anything that becomes legally binding without human review, or that is safety-critical.
Should I buy, connect or build?
Buy what already exists and build only what sets you apart from others. The three routes differ in effort, cost and control.
- Buy. Often the right AI feature is already part of software you use anyway, or there is a ready-made tool. This is quick and costs little, but it barely adapts to your particular needs.
- Connect. A language model is built into your existing workflow through an API, for example as a small internal tool. This lets you fit the workflow more closely to your needs without developing a complete application.
- Build. A custom solution pays off if the process is core to your business, very specific or has to draw on your own documents and data. This is the most demanding route, but it gives you full control.
The first route is the most common. According to Statistics Austria (in German), 56% of the companies in Austria that use AI bought ready-made commercial software or systems.
What do I need to consider about my data?
A model can only answer as well as the material it is given. If a tool is supposed to answer questions from your own documents and these are scattered across ten folders, outdated or unstructured, even a good model will give poor answers. So plan time for cleaning up the data and making it accessible.
How do I test whether AI pays off for us?
With a single, narrowly defined use case that you test for two to four weeks and measure against metrics set in advance. Four metrics are enough to start with.
- How much time does the solution save per task?
- What is the error rate?
- What does one run cost?
- Do employees accept the tool?
Without these figures, all you have at the end is a gut feeling, and you cannot base an investment decision on that.
What does the EU AI Act require of companies that only use AI?
As long as your use does not count as a high-risk application, there are essentially two obligations: AI literacy and transparency. For AI systems with minimal risk, only the AI literacy obligation applies, according to the AI Service Desk of the Austrian regulator RTR (in German). The dates come from the RTR timeline (in German).
| Date | What applies |
|---|---|
| 2 February 2025 | Prohibited AI practices may no longer be used, and the rules on AI literacy apply. |
| 2 August 2026 | The transparency obligations under Article 50 must be met. |
| 2 December 2027 | The requirements for high-risk AI systems under Annex III apply. |
What does AI literacy mean?
You have to take measures that build your employees' knowledge of how to work with the AI tools you use. Since the amendment by the Digital Omnibus of 27 July 2026, companies have to support the development of this literacy but do not have to guarantee a particular level, according to RTR (in German). RTR recommends a written internal AI policy, recurring training and a record of who was trained and when. According to Statistics Austria, only 27% of small companies that use AI have such a policy so far.
What do I have to label?
Since 2 August 2026, people must be able to tell when they are talking to a chatbot. According to RTR (in German), this obligation falls on the provider of the system. As a deployer, you must disclose when images, audio or video are so-called deepfakes. You must also label AI-generated texts that inform the public on matters of public interest. This does not apply if a person has reviewed the content of the text and someone holds editorial responsibility for it.
How high are the fines?
The maximum fine of up to €35 million or 7% of worldwide turnover in the previous year applies to prohibited practices, according to RTR (in German). For breaches of other obligations, it is up to €15 million or 3%. For small and medium-sized businesses, the lower of the two amounts applies in each case.
What else do we recommend?
Language models can sound convincing and still be wrong. So have a person check anything with consequences before it is used. Sensitive content belongs in a protected environment, European if possible, and not in some random tool from the internet.
Conclusion
To get started, you need neither your own data team nor a large project. You need a single tedious, repetitive process, a narrowly defined pilot, honest measurement and a solid set of basic rules.
Fragon Studios supports mid-sized companies from the first idea to a solution in production, with a focus on concrete benefit rather than buzzwords. If you want to find out which process in your company will pay off fastest, get in touch.
Frequently asked questions
Do I need a dedicated AI team to get started?
No. You need one clearly named person who drives the topic, collects feedback and decides on improvements. For the first use case, an AI feature in software you already use is often enough.
How long should a first AI pilot run?
We recommend two to four weeks for a single, narrowly defined use case. Decide beforehand how you will measure success, for example by the time saved per task, the error rate, the cost per run and acceptance in the team.
Do I have to train my employees in AI?
Yes, to an appropriate extent. The EU AI Act requires companies that use AI to take measures that promote the AI literacy of their staff. The RTR AI Service Desk recommends a written policy, recurring training and documentation of that training.
Do I have to label AI-generated content?
In certain cases, yes, and it has applied since 2 August 2026. You must disclose deepfakes. You must label AI-generated texts on matters of public interest if no person has reviewed their content and no one holds editorial responsibility. With a chatbot, it must be clear that you are talking to an AI.
Sources
- Austrian companies among the EU leaders in the use of artificial intelligence (in German), press release of 24 June 2026, Statistics Austria
- STATreport “ICT Usage in Enterprises 2025” (in German), Statistics Austria
- 20% of EU enterprises use AI technologies, 11 December 2025, Eurostat
- AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value, 24 October 2024, Boston Consulting Group
- AI Act timeline (in German), RTR AI Service Desk
- AI Act: obligations for deployers (in German), RTR AI Service Desk
- AI Act: AI literacy (in German), RTR AI Service Desk
- AI Act: transparency obligations (in German), RTR AI Service Desk
- FAQ on the AI Act (in German), RTR AI Service Desk
