Over the last year I led four AI proofs-of-concept — Smart Hiring, the Hand & Stone chatbot, a Hertz trip planner, and VRIZE OneView. Different domains, same lesson: a demo is not a system.
A PoC's job is to kill risk, not to impress
The point of a proof-of-concept is to find the one thing that will sink the real build — early, cheaply, on purpose. For us that was almost always retrieval quality and failure behaviour, never the base model.
The architecture that kept repeating
- Ground it. RAG over a curated, versioned corpus beats a bigger context window you can't trust.
- Constrain it. Tools and structured outputs turn "the model said something" into "the system did something."
- Observe it. Log prompts, retrievals and outcomes from day one. You can't improve what you can't see.
Prompt engineering is interface design
A prompt is an API contract written in prose. Treat it like one: version it, test it against fixtures, and review changes. The Master in AI Architecture work mostly reinforced this — the durable skill is systems thinking, not prompt trivia.
Where this is going
The interesting frontier isn't bigger models; it's reliable agents inside real enterprise workflows — the same migrations, reconciliations and audits I've always worked on, with a smarter assistant in the loop.