ARCHITECTURE
AI needs an architecture around it
The model is one component. The surrounding system still needs engineering.
From components to systems
Working with radar, automated tests and interconnected systems taught me to ask how components behave together. An impressive component is not the same thing as a dependable system. The same distinction is useful when adding language models to real work.
Put boundaries where they can be tested
AI-supported work needs explicit inputs, permitted actions, review responsibilities and failure handling. Conventional code can enforce some boundaries. Tests can exercise known cases. People still need to decide whether the output is suitable for the task and what to do when it is not.
Make assumptions and limits inspectable
A structured workflow does not make a model infallible. It can make assumptions and failure points easier to inspect. That is the engineering objective: understand which parts are controlled, which are probabilistic and which decisions remain with an accountable person.