If you own an AI product’s P&L, someone in finance has asked what it costs, and the honest answer was a range. Not because nobody looked. Every place you could look answers a different question. The cloud bill is a total with no unit behind it. The vendor’s pricing page times a token estimate stops being true the first time usage shifts. Finance has the spend but not the drivers, and engineering has the drivers but not the spend. Nobody can say what one more subscriber costs.
FinOps consulting for AI products closes that gap. I build the unit economics, the model your team runs before the next feature is built, and the reporting path finance runs without engineering. Then I hand it over. Most FinOps consulting services are built for compute: rightsizing, reservations, tagging. Inference doesn’t behave like compute. Its cost moves with usage, model choice, and prompt design, and it keeps moving after the pricing page is public. That is the workload this practice is built for, whether the product is in production or about to ship. The rest of what I get brought in to do is on the home page.
What you get
- Cost per request, cost per active user, and cost per subscriber for your top AI products, delivered in the first month and broken out by model and by feature, so a pricing decision has inputs instead of a range
- A cost model product teams run before a line of code is written, covering model selection, token splits, caching efficiency, and the break-even point as usage scales
- A cost driver breakdown that names which models, which training runs, and which inference paths actually move the bill, so the fix targets the right system instead of the easiest one to blame
- A reporting path finance runs without waiting on engineering: pacing, budget, forecasting, and anomaly detection inside a tool FP&A already opens
- A baseline forecast engineering and finance both sign before the next budget cycle, built on the unit that maps to usage rather than raw compute spend
How an engagement runs
Baseline. Unit economics for your top AI products: cost per request, per active user, per subscriber. At the end, you have numbers nobody in the org has produced before.
Attribution. Cloud and model spend correlated against product usage data, so a cost spike traces to a system or a feature instead of a guess.
Pre-build model. The calculator product teams use to price the next feature before engineering commits to an architecture. If you are still pre-launch, this is where the engagement starts.
Reporting path and handoff. A self-serve dashboard finance runs on its own, replacing the export-and-email cycle, plus the baseline forecast both sides sign. I leave when your team can run this without me.
What a FinOps consultant for AI has actually built
A Fortune-500 financial data company with eight figures of annual AI and cloud spend. Built cost per request, per active user, and per subscriber for the top three AI products, plus the calculator product teams now use before scoping a feature. Finance replaced a manual cost-tool pull with a self-serve reporting path it runs without engineering.
A media and streaming company with nine figures of spend across five cloud platforms and 600+ engineering teams, mid-merger. Built the consolidated forecast and cost model that let two companies’ cloud spend report as one number. Driver-based forecasting landed within 5% of actual.
Who it isn’t for
- A single internal AI tool with no P&L attached to it. A cloud cost review handles that; this practice is built for products with a pricing decision behind them.
- A team that wants the model built and then kept by me. This is scoped: baseline, attribution, pre-build model, reporting path, handoff. If you need FinOps as a service running year-round, that is a different search.
- A team that wants the number softened for a board deck. The forecast is the forecast, and the value is in finance and engineering trusting the same one.
Start
If you have an AI product in production, or one about to ship, and a number you can’t answer cleanly, that is the starting point.
Email nick@nparr.com with the product and the number you can’t answer. I’ll reply with whether a two-week scoping makes sense.