If you cannot state the failure mode you are most afraid of, you are not ready to measure the system.
The demo is not the system
A prototype that works on curated inputs tells you almost nothing about what happens when the workflow runs unattended against your real data.
June 18, 2026
evaluationproductionMost AI pilots die in the gap between a demo that works and a system that runs. The demo is built against inputs someone chose. The system meets whatever arrives on a Tuesday.
Write the acceptance test first
Before any build work starts we agree a threshold: which cases must pass, at what rate, and what happens to the ones that do not. A sprint that ends in “not yet, and here is precisely why” is a useful outcome. A sprint that ends in an unmeasured demo is not.
Exceptions are the product
The happy path is the easy quarter of the work. The rest is deciding what the system does when a document is malformed, an API is down, or the answer is genuinely ambiguous — and who it escalates to.