When I notice people talk about Generative AI, they so often focus on prompting, LLMs, chatbots, and proofs of concept.
But what actually matters if you want to build something that works – or has to work more than a year is a bit more. This is not a complete list, just what came to mind in a minute:
- Try classical automation first. Not every problem needs GenAI.
- Integrate it into a process. Solve a real problem, not just a demo.
- Data privacy. Compliance isn’t optional—GDPR, access controls, and all.
- Security. Protect against prompt injection, data leaks, and model attacks.
- Betriebsrat (works council). In Europe, you can’t skip this if you are processing employee data. Involve them early.
- Observability. Log, monitor, and trace everything. You can’t fix what you can’t see. (But mind the requirements of the works council and data protection)
- Robust data pipelines. Your ETL must be reliable. Garbage in, garbage out. And you don’t want to be busy fixing data loads all the time.
- Model selection. Open-source? Proprietary? Cost vs Capabilities … Just picking the largest and latest model very often isn’t the wisest choice.
- Model half-life. Performance degrades. Plan for re-evaluation and updates. Models are dicontinued rather quickly.
- Costs. Tokens are just the start. Think maintenance, infrastructure, and long-term spend.
- Testing. Treat it like mission-critical software. Test rigorously.
- The other 95%. The AI is just a small part. APIs, UX, and infrastructure matter as well – or even more.
- APIs. If your app needs to talk to other systems (well because it should DO some work FOR you). Design for integration.
- Scalability. Will it work for 5 users? Or 10,000?
- Documentation. The EU AI Act (and common sense) demands it.
- Interface & UX. If users hate it, they won’t use it.
- Address the fear. People worry about losing their jobs. Be transparent about the impact.
GenAI isn’t magic—it’s engineering. Focus on the basics, or fail fast.
Leave a Reply