Evaluation is the job. But evaluation requires knowledge. You cannot judge outputs if you don't know what good looks like.
This is where many teams fail. They build the system, run some tests, look at outputs, and ask themselves: "Is this right?" And they realize they're not sure. They don't have the knowledge to judge.
Evaluation requires three distinct types of knowledge. Each serves a different purpose. All three are necessary.
To judge whether an AI's output is correct, you need to understand the underlying domain logic. Not at expert level. But enough to recognize when something is obviously wrong.
If you're building an AI that recommends materials for engineering applications, you need to understand why certain materials are appropriate — the constraints, the trade-offs, the reasoning that makes an answer right or wrong.
If you're building an AI that extracts data from invoices, you need to understand what an invoice contains, how the numbers relate to each other, what "matching a purchase order" actually means.
Without domain knowledge, you're pattern-matching against vibes. You might catch outputs that look strange. You'll miss outputs that look plausible but are subtly, dangerously wrong.
The goal is not to become a 20-year expert. That's not realistic, and it's not necessary.
The goal is calibration. Enough knowledge to:
This takes investment. A couple of hours for a quick feature. Dozens of hours if this is the core product you're building. The time matters less than the density — two hours reviewing real examples with an expert explaining their reasoning beats two days reading documentation.
You cannot outsource this. You cannot wait for the expert to evaluate everything. You need enough knowledge to run the first loops yourself, to iterate quickly, to know when to escalate. If you're building it, you must be able to judge it — at least at a first-pass level.