For thirty years, marginal cost on a software feature was basically zero. The PM job was demand-side (engagement, retention, conversion), and the CFO worried about supply. That deal is over. In 2026 every AI feature I ship has a cost line that moves with usage, and if the PM does not own it, nobody does.
Why it lands on the PM
Token costs, retrieval costs, model routing, vendor markup, and latency-driven retries move gross margin by 20 to 40 percentage points on a product. That is the difference between a venture-backed business and a venture-killed one. The CFO sees a number at month end. The engineer sees a prompt at the character level. Only the PM sees the full customer job end-to-end, and only the PM can say "this entire flow costs too much and needs to be rebuilt." Cost is a product decision that happens to show up on the finance ledger.
The numbers that should ruin your afternoon: AI-first SaaS runs 55 to 70 percent gross margin against traditional SaaS at 78 to 85. A B2B app processing 50M tokens per enterprise customer per month spends $500 to $2,000 in raw inference before anything else. Companies still on per-seat run 40 points lower margin than those on hybrid usage pricing. And a model like Sonnet 4.5 doubles in price above the 200K token input threshold, so a PM who does not know that threshold ships a feature that silently halves margin the first time a customer sends a large document.
The three levers every PM can pull
- Model routing. Not every task needs the flagship. Classification, structured extraction, simple rewrites, and format conversion often work on a 10x cheaper model with comparable quality. Build a routing table: for each surface, what model, what fallback, what eval threshold triggers a swap. Review it quarterly; about once a quarter you find a surface over-routed to the flagship for no reason other than it shipped that way.
- Prompt and context hygiene. Most production prompts are 3 to 10x longer than needed. Cutting one from 3,000 tokens to 600 cuts cost on that surface by about 80 percent with no quality loss. Open the prompt, cut ruthlessly, run the eval, keep the cuts that survive.
- Caching and early exit. Prompt caching cuts cost 50 to 90 percent on repeated prefixes. Early-exit logic (do not call the LLM when a rule or cached answer will do) cuts cost 100 percent on the hit. Identify the top 20 percent of inputs by volume and design for cache hits or rule shortcuts there.
The dashboard I watch
Cost per successful action by surface, its 7-day trend, latency p50/p95/p99, token volume input versus output, cache hit rate, model distribution, and the cost of failed actions. If it is not on your wall, you are guessing. I ran the ugly version as a Notion page for six months before anyone built a real dashboard, and the ugly version was better than nothing. This is why pricing is the natural next decision: cost and pricing are two sides of the same question, covered in Pricing for AI Products.
Do this this week
Pick your most-used AI surface, find its cost per successful action over the last 30 days, and if you do not have that number, figuring out why is your week. Find the top 10 percent of requests by cost, spot what they share, and apply one lever. My first time doing this at Smartcat I found a 4,200-token prompt three PMs had padded over six months, cut it to 800 at the same eval score, and the feature got 70 percent cheaper overnight.