The hard part isn't the model — it's the rules it can't get wrong
In a regulated process, the AI is the easy part. Faithfully encoding the rules nobody is allowed to break is the whole job.
I built a platform that drafts, reviews, and manages government contracts with a set of specialized agents, checked live against FAR, DFARS, VAAR, and HHSAR. The thing everyone expects to be hard — getting the model to write competent contract language — turned out to be the easy part.
The hard part was the rules. In a regulated domain, “mostly right” is a failure. A clause that’s 95% correct is still non-compliant, and non-compliant means the work gets thrown out. So most of the real effort wasn’t prompting — it was the unglamorous job of capturing exactly what the regulations require, where they conflict, and which ones win, in a form the system can apply every single time without drifting.
That reframed how I think about AI in serious work. The model is a capable but careless intern. The value you add isn’t the intern — it’s the guardrails that make its output trustworthy enough to ship. The leverage lives in the rules, not the model.
A close cousin of this idea shows up in The most useful tool sometimes refuses to answer: sometimes the most valuable thing the system does is not the confident answer.