When we talk about the cost of an AI feature, the conversation usually starts with:

How much does each inference cost?

That's important, but it's incomplete.

I think every AI feature has three costs:

Inference cost + UX cost + trust cost

And a PM should consider all three before deciding whether an AI capability is worth shipping.

1. Inference cost

This is the obvious one.

Every model call has a cost:

• Tokens
• API calls
• Tool calls
• Retrieval
• Compute
• Retries and fallbacks

A feature that requires five model calls instead of one may be significantly more expensive to operate, but sometimes paying more is worthwhile if the user value is substantially higher.

So the question isn't simply:

“Can we make this cheaper?”

It's:

“Is the additional intelligence worth the additional computation?”

2. UX cost

AI also costs the user time and attention. A feature might be technically impressive but introduce:

• 30-second waits
• Too many clarifying questions
• Long responses
• Unnecessary steps
• Unpredictable behaviour

For example, an agent might produce a better travel plan, but if it takes 90 seconds to do so, users may abandon before seeing the result. That's a UX cost.

Sometimes the right product decision is to sacrifice a little model quality for a dramatically faster experience.

3. Trust cost

This is the cost that's easiest to overlook. Every incorrect AI output spends some amount of user trust. A recommendation that is slightly irrelevant may be harmless. A fabricated flight price, incorrect visa requirement or confidently wrong financial answer is much more damaging.

Trust can also be affected by:

• Unexplained decisions
• Inconsistent answers
• Hallucinations
• Overconfident language
• Unexpected actions

And trust is difficult to recover once lost.

The PM trade-off

This gives me a simple way to think about AI product decisions:

More intelligence

→ potentially higher inference cost

→ potentially higher UX cost

→ potentially higher trust cost

The best solution isn't necessarily the most intelligent one.

It's the one that creates the most user value for an acceptable cost across all three dimensions. For example, in the travel webapp, I might choose deterministic logic for budget validation rather than asking an LLM to reason about whether a trip is affordable.

That can simultaneously: reduce inference cost + reduce latency + increase trust.

That's a much better optimisation than simply finding a cheaper model.

The framework

Before shipping an AI feature, I ask:

Cost	Question
Inference	What does this cost us to run?
UX	What does this cost the user in time and complexity?
Trust	What happens when the AI gets it wrong?

AI product optimisation isn't just about making AI cheaper. It's about making intelligence worth its total cost.

That's the trade-off I think AI PMs should be designing for.