There is a number attached to every AI feature your team ships, and most PMs cannot name it. Not roughly, not at all. For twenty years that was fine, because software had no meaningful marginal cost and another user was another row in a database. That arrangement is over. Every model call, every retry, every agent loop shows up on an invoice, which means you own a P&L line now whether or not anyone updated your job description.
The five numbers
Fluency here does not mean accounting. It means five numbers you can produce in a hallway. First, the gross margin of your area, yours, not the company number. Second, cost per outcome of your main workflow, the cost to deliver one unit of the thing the customer buys, because per-user averages hide everything when usage variance is enormous. Third, your two biggest cost drivers and, critically, which product decision moves each one, since a driver you cannot connect to a decision is trivia. Fourth, the payback or LTV to CAC shape of your segment. Fifth, the compute bill trend over the last two quarters, because "growing faster than revenue" is the one that ends up in a hard conversation with your CFO, and you want to see it a quarter before they do. When I started asking PM candidates for these five from their current role, the pass rate was depressing, and the ones who passed were, without exception, already operating at the next level.
The working session
You cannot get cost per outcome from a dashboard. You build it by sitting with your finance partner for 90 minutes. Agree the outcome unit in one sentence. List every cost that fires when the workflow runs once: model calls by model, retries and fallbacks, tool and API calls, vector search and storage, human review minutes at loaded cost. Pull one real month of volume including retries. Divide and sanity-check against the actual bill, and if it does not match, find the leak, which is usually retries or a workflow nobody remembered was running. Name the two biggest drivers and one product decision that moves each. The second-order effect is always the same: finance starts treating product as a partner instead of a cost center.
Product decisions are margin decisions
This is the shift that makes the numbers useful. Model routing (sending the easy 80% to a smaller model) can transform unit cost, and knowing which requests are easy is product knowledge. Caching repeated context is often the bulk of token spend, and prompt structure decides cacheability. Workflow design matters because an agent that takes seven steps where four would do costs nearly double for the same outcome. Confidence thresholds decide when to escalate to the expensive path. Each shows up in gross margin within a quarter, and none of them belongs to procurement.
Run the 4-week plan: week one, get your five numbers and estimate cost per outcome badly on one page; week two, run the working session; week three, find one margin lever and size it before building; week four, put cost per outcome, its trend, and the lever you are pulling into your next exec update. For reading, Berman and Knight's Financial Intelligence is the only book you need, and your own finance partner is the best free course available. Send them the ask this week: 45 minutes a month, starting by rebuilding cost per outcome together.