Building AI has changed. The core skill is no longer designing architectures or writing clever code. Anyone can apply a prompt, call a model, and generate output. The barrier to producing something has collapsed.

This changes what separates good work from bad work.

Generative AI, by definition, generates a result. Not the result. Not necessarily the right result. It produces output that looks plausible, sounds confident, and might be completely wrong.

The same is true for pre-packaged AI functions - extraction APIs, classification endpoints, summarization services. You call them, you get output. But without knowing what the output should look like, you're just trusting the black box. You've delegated judgment to a system that has none.

Without evaluation, you have no idea if your system works. You just have output and hope.

The new core skill is knowing whether the output is right.

Calling an AI — whether it's a prompt to a language model or a call to a packaged function, is trivial. The API returns something. It always returns something.

The question is: Is that something correct? Is it good enough? Is it dangerously wrong in subtle ways?

If you don't know what the right output looks like, you can't answer these questions. You're shipping code you cannot judge.

The builder's primary job is now:

1. Assess results - Is the AI performing correctly? How often? On which cases?
2. Diagnose failures - Which outputs are wrong? What patterns emerge? Where does the system break?
3. Improve - What's causing the failures? How do we fix them?

You cannot do steps 2 and 3 without step 1. Everything downstream depends on evaluation. It is the foundation.

Evaluation is the forcing function.

When someone says "we're building an AI to help with procurement," that's a direction. It's not a problem definition.