The videos all look much the same: in twenty minutes, a personal AI assistant takes shape that answers emails, books appointments, sorts invoices and dims the lights on the side. Like Jarvis from the films. It looks easy, and that is exactly the problem.
What the videos leave out
- Everyday reality. A demo runs on hand-picked examples. In real operation, the cases nobody thought of turn up. More on this in our article on the AI demo trap.
- Access. An assistant that answers your emails needs access to your inbox, your calendar and often your accounts. Whoever can control it can do a great deal.
- Instructions from outside. An email can contain text that the assistant takes as an order. A person would ignore it; a badly built AI carries it out.
- Mistakes. AI is sometimes wrong, and with complete conviction. Who notices before it reaches the customer?
- Cost. Every request to a language model costs money. With an assistant that is constantly reading along, it adds up.
- Maintenance. Interfaces change, services get rebuilt. What works today needs someone to fix it tomorrow.
- Data protection. Customer data going to a service in the USA is no small matter for a business in Austria.
What works instead
One agent, one task. Instead of an assistant that is supposed to do everything, a clearly defined process: sorting enquiries, preparing quotes, matching receipts. With firm limits, with a log that shows what it has done, and with a person who approves before anything goes out.
Where sensitive data is involved, the AI can also run on your own hardware rather than with a cloud service. We clarify that beforehand, not afterwards.
Starting small pays off
We have been using AI in our own work for years, every day. What really saves us time is not an all-rounder, but individual processes that now run by themselves. A Jarvis can grow out of that. But not in twenty minutes.