One of the hardest parts of building AI products is deciding whether an improvement is actually an improvement.

A model can produce better answers according to an evaluation score, but that doesn't necessarily mean users get more value from the product.

I think about AI product evaluation as a three-layer chain:

Model quality → User value → Business value

Each layer answers a different question.

1. Model quality: Did the AI get better?

Start with the behaviour of the AI itself.

Depending on the product, this could mean:

• Accuracy
• Relevance
• Factuality
• Instruction following
• Hallucination rate
• Tool-call success
• Structured-output validity

For the travel webapp, for example, I might evaluate whether recommendations satisfy the user's constraints and whether the generated rationale is relevant.

But this is only the first layer.

A higher evaluation score doesn't automatically mean a better product.

2. User value: Did the user actually benefit?

The next question is:

Did the improvement change what the user could accomplish?

For the travel webapp, better recommendation quality might lead to:

• More plans saved
• Fewer regenerations
• Fewer dismissals
• Higher satisfaction
• More users progressing towards booking

This is where AI evaluation starts becoming product evaluation.

For example:

Model quality improves → recommendations become more relevant → users regenerate less.

The important metric isn't necessarily the model score.

It's the user behaviour that the model improvement was supposed to influence.

3. Business value: Did it matter to the product?

Finally:

Did the improvement create meaningful business value?

Depending on the product, this could be:

• Conversion
• Retention
• Revenue
• Cost reduction
• Activation
• Engagement

For the travel webapp:

Better recommendations → more saves → more booking intent → potentially higher conversion.

But there can also be a negative trade-off.

Suppose a new model improves recommendation quality by 5%, but increases inference cost by 80%.

That's not automatically a good product decision.

The PM needs to evaluate:

incremental user value vs. incremental cost.

The evaluation chain

I use this simple mental model:

Model metric

↓

User behaviour

↓

Business outcome

For example:

+10% recommendation relevance

→ +6% save rate

→ +3% booking conversion

Now we have a credible product story.

Compare that with:

+10% evaluation score

That tells us the model changed.

It doesn't yet tell us whether the product improved.

The key PM question

The most important question isn't:

“Did the model get better?”

It's:

“What user behaviour should improve if the model gets better, and what business outcome should that ultimately influence?”

That question helps prevent a common AI product mistake: optimising what is easy to measure rather than what actually matters.

A strong AI evaluation system therefore connects three layers:

Model quality — Is the AI doing its job?

User value — Does that improve the user's experience or outcome?

Business value — Does that create value for the product?

Good AI evaluation doesn't stop at measuring the model. It closes the loop between model behaviour, user behaviour, and product outcomes.