Published on
AI Accelerates Many Things. Above All the Illusion That the First Draft Is Right
When it comes to access to language models, the differences between companies have narrowed. Tools and basic understanding are available. The demanding part begins afterwards, in daily work.
So we decided early not to treat this as a pure tool question but to look at the effects on how we work and on our culture. Everyone here has access to at least one language model. At the same time we deliberately invested time in building a shared baseline: what are sensible use cases? Where are the risks? Which guard rails always apply?
Because I observe two effects at once. On one hand efficiency rises: text, summaries and first analyses appear faster. On the other, the willingness to really work through a result falls. Where people used to cross-check or recalculate briefly, today they accept more quickly. The output gets used as though it had already been reviewed.
There is a psychological effect on top. When everyone can quickly do a bit of AI, a feeling emerges that you should be able to do all of it from the start. That creates pressure, and pressure rarely produces good learning curves.
Prompting, standards and review loops mean extra work at the beginning. That extra work is precisely the price of long-term reliability.
What helps is a low-barrier format for exchange. We share prompts regularly, show examples, and talk about wrong assumptions and about how we check results. Plausibility, the logic of the numbers, sources and assumptions all stay subject to review.
And culture above all means this: it is accepted for someone to say they are not sure. Rather than hiding uncertainty behind output.
Perhaps that is the difference between introducing AI and using AI. It is not speed or budget that decides, but whether an organisation demands and enables learning curves.