Enterprise AI beyond the prototype
A demonstration proves a model can produce an answer. It says nothing about whether anyone will act on that answer, which is the only thing that decides whether the system was worth building.
Insights
An institutional publication rather than a blog. Research notes, technology positions and company announcements.
A demonstration proves a model can produce an answer. It says nothing about whether anyone will act on that answer, which is the only thing that decides whether the system was worth building.
The happy path is the easy half. An automation is judged on what it does with the twelve percent of cases that do not fit, and most are designed as though those cases do not exist.
Collecting telemetry is a solved problem. Turning a stream of positions and sensor readings into something a dispatcher can act on before the shift ends is not.
Integration is scheduled last and estimated shortest, and it is where most of the real difficulty lives. Treating it as a feature of the system rather than a task at the end changes what gets built.
Systems are rarely retired because they stopped working. They are retired because changing them became more expensive than replacing them, and that threshold is set by decisions made early.
Saying a person reviews the output is not a design. Where the person sits, what they see, and what happens to their correction are architectural decisions with very different costs.