Technical due diligence, AI evaluation and governance.
This failure only produces false negatives. Systems that over-claim get caught early. A system that stays silent can be used for years while looking prudent.
Hire for the layer that does not commoditise: judgement about what to build versus buy, and the ability to evaluate whether what an AI system returns holds up.
If you want a quality metric to change a team's behaviour, it has to sit on the critical path of something they already care about. In software, that path is the pipeline.
In any system with AI, the design question is not which model do I use, but which part of this problem does not need a model at all. It is almost always more than it looks.
It sounds less interesting than talking about architecture. It is, however, where it gets decided which startups scale with margin and which run out of it, even when the product works.
A considerable share of the real operational risk is not in the model. It is in how the model reaches the machine where it runs, and what controls exist over that path.
When someone asks you for a date to add AI to a process, they are asking two different things: what it costs to build, and what it costs to find out whether it is possible at all.
23 incidents are enough to see patterns and not enough to estimate base rates. If that distinction does not travel with the figure, the analysis produces misinformation with your name on it.
The question when evaluating an agent is not what it can do, but what happens when it is interrupted midway. Partial failure is the scenario no demo shows and the only one that matters in production.
Presenting a winner without saying how far ahead it is of the runner-up hides exactly the information the decision-maker needs. The precise weight at which the recommendation would flip.
"Switzerland has no AI law" is both true and misleading. The territorial scope of the European regulation reaches whoever puts a system on the EU market, wherever they are incorporated.
Over- and under-forecasting do not cost the same. If you evaluate forecasting tools on MAPE, you are rewarding whichever best optimises a metric that treats both errors as equivalent.
No articles on that topic yet.