By now, almost every large company has run AI training. Almost none can say what came out of it. That's not a coincidence – it's a property of the format itself.
The core problem: knowledge without a reason
In a traditional training session, people learn something they don't need at the moment they learn it. The brain treats that information exactly as it should: it discards it. What's missing is the trigger – a real task that can't be solved without the new knowledge.
The question isn't whether the training was good. The question is what gets done differently on Monday morning.
| Criterion | Traditional training | Hackathon |
|---|---|---|
| Learning mode | Listening, taking notes | Building under time pressure |
| End result | Slide deck, attendance list | Working prototype |
| Relevance to daily work | Examples from the industry | Your own process |
| Visibility in the company | Attendance list | Demo in front of leadership |
| Measurability | Satisfaction survey | Time saved per use case |
| Half-life | Weeks | The prototype survives |
Three patterns you've probably seen
1. The tool tour
Someone demonstrates features for two hours. Participants see impressive demos and walk away with the feeling that "someone should really do something with this." What's created is admiration, not competence.
2. The prompt collection
A catalog of 50 templates gets handed out. Two weeks later, nobody uses anything beyond the three prompts they would have found on their own anyway. Templates don't replace understanding why a prompt works.
3. The pilot project with no owner
After the training, a use case gets chosen, written into a backlog, and never prioritized. The most common time of death for an AI initiative falls somewhere between the initial excitement and week three.
What measurably works better
- A real task instead of a sample task: teams work on their own actual bottleneck.
- Time pressure as pedagogy: a single day forces simplification and prevents endless debate.
- Mixed teams: domain expertise and tech affinity at the same table, not in separate tracks.
- A public demo: people listen differently in the morning when they have to present in the afternoon.
- A named owner per result before anyone leaves the room.
The uncomfortable consequence
This format takes more effort to prepare and is less comfortable for everyone involved. It can't just be rubber-stamped, because at the end it's obvious whether anything was actually built. That's exactly why it works. Learning that can remain without consequence, does.