I didn't get interested in AI because I wanted a machine to write clever paragraphs.
I got interested because language models changed the interface between messy human intent and structured software.
That is a serious shift.
It also creates an entirely new class of problems that disappear in demos and become painfully obvious in production.
The demo is the easy part
You type a question. The model gives an impressive answer.
Great.
Now put that system inside a business.
Who is allowed to ask? What information can it see? What happens when the model invents a citation? What if its answer changes after an update? Who can stop it from taking an action? Can you reconstruct the evidence behind a decision?
Suddenly the problem is no longer prompt engineering.
It's architecture.
What I want from a system
I want boundaries I can explain.
A defined source of evidence. Access controls that survive an enthusiastic user. A record of what happened. Clear handling of uncertainty. A way to refuse an action that hasn't been authorized.
Those aren't glamorous requirements.
They're the difference between a compelling prototype and something an organization can responsibly depend on.
Why I care
My background cuts across marketing, products, automation and engineering. That combination makes me impatient with tools that look magical until someone asks who owns the failure.
I don't need an AI system that performs confidence.
I need one that knows when its evidence is weak, when its authority stops and when a human needs to make the call.
If we build that properly, AI becomes infrastructure.
If we skip it, we've just automated the production of very convincing problems.