The question every AI feature has

1. The question

Every AI feature in production has the same implicit question hanging over it.

"But is it actually working?"

You can phrase it nicer. "How's the quality?" "Are users happy with the summaries?" "Is the classifier good enough?" Same question. Somebody on your team is going to ask it. Probably your PM. Probably soon.

The honest answer for most teams, most of the time, is "I don't know." They don't have a way to find out. They have launch metrics (clicks, retention, complaints) but those are downstream. By the time clicks drop, the feature has been wrong for a week.

2. Why the question is hard to answer

For a traditional feature, "is it working?" decomposes into things you can check.

Does the API return 200? Does the data save to the database? Does the page load in under 500ms? Do the unit tests pass? You have answers. You wrote the assertions yourself.

For an AI feature, the question doesn't decompose the same way. The API returns 200. The page loads. The model gave you back a string. Is the string right? That's a different question, and your unit tests don't answer it.

You can't assertEqual your way to "is this summary good." The model is generative. The space of acceptable outputs is large. The space of unacceptable outputs is also large. You're now in measurement territory, not assertion territory.

3. What this course calls evals

An eval is the method for answering "is it actually working" with something other than vibes.

That's it. That's the whole frame. An eval is a way to take that fuzzy product question and turn it into a number, or a table, or at minimum a structured opinion that another engineer can read and agree with.

You'll see the word used to mean a lot of things — frameworks, dashboards, vendor products, benchmarks. Most of those are tools that help you run evals. They're not what an eval is. We'll come back to that in the next unit.

4. Why most teams skip this

Three reasons, in roughly this order.

They don't know it's a category. Evals as a discipline are newer than most of the engineers reading this. If you trained as an engineer before 2023, evals weren't part of the curriculum. Nobody told you this was a thing you were supposed to do.

They think it's a framework to adopt. Evals get talked about like they need infrastructure. They don't. The smallest useful eval is a markdown table you maintain by hand. Tools come later, when the table gets too big to read.

They confuse evals with "AI testing." They look at their feature, decide "we can't really test this," and stop. Evals aren't testing. They're measurement. The point is to get a number that goes up or down between Tuesday and Friday, not to assert a fixed truth.

5. What's coming

The rest of this chapter defines what an eval is, shows you the smallest useful one (five rows, three columns), explains why eyeballing doesn't work, and ends with the mindset shift this whole course is trying to install.

If you only remember one thing from chapter 1, remember this: an eval is the structured answer to "is it actually working." Everything else in this course is mechanics.