The Unit Economics of AI
This is the part of AI product management that SaaS instincts get most wrong, and it's the term in need × model behavior × economics that most PMs are weakest on. If you build AI features without understanding inference economics, you'll ship something users love and the business can't afford. I've watched it happen. The unit economics really are different, and you have to internalize how.
#Inference cost is not SaaS margin
In classic SaaS, the marginal cost of one more user is close to zero. You've already built the software, so serving another customer is rounding-error cheap. That's why SaaS runs 80 to 90% gross margins. AI breaks this. Every single inference costs real money, because the model has to run and burn compute every time a user does the thing. That marginal cost doesn't conveniently vanish at scale.
AI products commonly run roughly 50 to 60% gross margins against SaaS's 80 to 90%. If you price and plan an AI product on SaaS-margin assumptions, your unit economics are upside down, and you won't notice until usage scales. Orchestration (multi-step agents, retrieval, chained calls) can add another 30 to 60% to the per-request cost on top of the base model call. The agent that delights users in the demo may be three to five model calls per interaction once it's in production. (Treat the margin and cost figures here as directional industry rules of thumb. They vary widely by product and shift fast as model prices fall.)
#The levers a PM controls
The good news: a lot of cost is a product decision, not just an infra one. These are PM levers, and owning them is part of the job now:
| Lever | What it does | The trade-off |
|---|---|---|
| Prompt / response length | Fewer tokens in and out = lower cost | Terser context or output vs quality |
| Model selection | Smaller/cheaper model for easy tasks | Cost vs capability (route by difficulty) |
| Concurrency / batching | Smooth load, batch where latency allows | Throughput vs responsiveness |
| Caching | Reuse results for repeated inputs | Freshness vs cost |
| Retrieval scope | Send only relevant context | Precision vs completeness |
The highest-leverage move is usually routing: send the 80% of requests that are easy to a cheap, fast model, and reserve the expensive model for the hard 20%. That's a product decision about quality thresholds, exactly the kind of call an AI PM owns, and exactly what evals let you make safely.
#The counterintuitive trend
Cost per token is collapsing, on the order of 10x cheaper per year. So AI gets cheaper, right? No. Total spend is going up, because reasoning models burn far more tokens per task and we keep handing AI bigger jobs. Cheaper unit, more units, bigger total. Don't assume falling token prices bail out your margins, because consumption is outrunning them. Plan for the bill you'll have at scale, not the one the calculator shows you today.
#Pricing implications
Per-seat SaaS pricing is structurally mismatched with usage-driven cost. When cost scales with tokens and price scales with seats, your heaviest users quietly destroy your margins. The market is moving toward base + usage (a platform fee plus consumption), or credits, or outcome-based pricing. The PM has to connect the pricing model to the actual cost driver, which is Business Acumen and Models and viability (The Four Big Product Risks) made unavoidably concrete. A flat-rate AI product with unbounded usage is a business-design error, not a feature.
#My take
You don't need to run the finance model yourself, but you do need to reason about it: what does this feature cost per use, what drives that cost, and does the price still cover it as usage grows? An AI PM who can't answer that is missing a full third of AI product sense. The economics aren't a constraint you hand off to finance. They're a design dimension you build around from day one.
#Continue Reading
- The AI Product Manager for economics as a core part of AI product sense.
- Building AI Products for designing within cost constraints.
- Business Acumen and Models for the business-model competency this extends.
- The Four Big Product Risks for viability risk, made concrete.
- Writing Evals for AI Products for how evals let you route cheaply without losing quality.
- The 70% Problem for the quality-vs-cost trade-offs in the last mile.