AI Feature Pricing Is Broken and Most CPOs Don't Know It Yet

The math that nobody modelled at pricing time

Per-seat flat pricing was built for software where the marginal cost of one more user doing one more thing is essentially zero. That assumption held for a long time. It does not hold for AI inference.

When a power user runs daily feedback synthesis, generates three roadmap drafts, and processes meeting transcripts through your AI layer, they are triggering real compute costs on every action. Those costs compound across your user base. The $12 to $59 per user per month spread you see across PM tools today was set before AI usage was material in the numbers. None of those price points were modelled against what a team of ten PMs actually consumes when AI is embedded in their daily workflow.

Productboard offering Spark free in public beta is the clearest current example of this dynamic. It is a smart customer acquisition tactic. It is also transferring inference cost directly onto the vendor at a point when nobody yet knows what the steady-state usage curve looks like. The math works until it doesn't, and the inflection point is not a pricing question — it is a usage question, and usage is the one variable the vendor does not control.

As Shreyas Doshi has pointed out, the hardest product decisions are the ones where the short-term signal and the long-term signal point in opposite directions. Right now, AI bundling sends a strong short-term signal — adoption, engagement, competitive positioning. The long-term signal is quieter, and it lives in your gross margin.

There are three viable responses, not ten

I've seen CPOs treat this as a finance problem to solve later. My experience is that by the time it becomes a finance problem, you are solving it under duress — with churn risk on one side and investor scrutiny on the other. The window to design your way out of this is now, not after the margin wall.

The three responses that actually work structurally are these.

Usage-based pricing tied to AI actions. You price the AI layer on consumption — credits, queries, synthesis runs — rather than per seat. This is the cleanest model from a unit economics perspective because cost and revenue move together. The challenge is that it introduces variability that buyers often resist, especially in enterprise procurement.

A hard capability tier with a margin buffer. AI features live behind a premium plan priced with enough headroom to absorb inference cost at scale. The tier has to be real — not a feature flag that enterprise clients negotiate away — and the price point has to be modelled against actual usage, not assumed usage.

A platform bet. You accept that AI inference is subsidised, because AI drives enough retention and expansion revenue that the lifetime value math still works. This is not a bad model, but it requires deliberate architecture: you need to know which AI interactions actually move retention and which are just engagement theatre. Most teams cannot answer that question yet.

Each of these requires a deliberate architecture decision made before you hit the wall, not after.

The conversation to have before the board has it for you

The uncomfortable truth is that most AI pricing decisions made in 2023 and 2024 were made to win the market moment, not to survive at scale. That was probably the right call then. It is not the right call to leave unchanged now.

What I'd bring to your next planning conversation is one direct question: at our current AI adoption rate, what does our gross margin look like at two times the usage? If you can't answer that with real numbers, the pricing model isn't finished — it's deferred.

Pick one of the three structural responses. Model it against your actual inference costs. Test it with a segment before you need to reprice under pressure.

The vendors who design their way through this now will arrive at the inflection point with something tested. The ones who treat it as tomorrow's problem will arrive with something broken.