AI unit economics: what every AI dollar buys.

By Justin Moore · Founder & CEO, COGScontrol · June 12, 2026

AI unit economics is the practice of measuring what each unit of business activity — a user, an interaction, a customer — costs to serve with AI, and what it returns. It divides fully loaded AI costs — model tokens plus supporting cloud infrastructure — by business metrics such as monthly active users and revenue to produce per-unit costs and margins.

The discipline has moved from abstraction to line item with unusual speed. Uber exhausted its 2026 budget for AI coding tools by April, and its chief operating officer was candid that the link from tokens to shipped value “is not there yet” — what the company wants, he said, is to draw a direct line from spend to useful features. Drawing that line, unit by unit, is the quantitative core of AI Value Management, and it starts with four formulas.

What are the core AI unit economics formulas?

Four formulas do most of the work: cost per monthly active user, cost per interaction, contribution margin per customer, and AI-attributed gross margin. Each sets a fully loaded cost figure against a business metric, and each answers a different question.

The affordability headline. It tells you what it costs to keep one user active for a month — the figure to set against per-seat or per-user pricing. It rises when users do more, which is not automatically bad news, so read it alongside the next metric.

The engineering-economics gauge. It isolates the cost of a single unit of AI work — a query, a completion, an agent run — and is the most sensitive of the four to model swaps, caching, and prompt design. When a team switches models or reworks a pipeline, this is the number that should move.

The pricing test. Average revenue per user, minus the AI cost attributable to that customer, should be comfortably positive in every tier. When the heaviest users push it negative, the pricing model is subsidizing precisely the customers who cost the most to serve.

The board-level number. It treats AI serving costs as cost of goods sold and shows what the AI-enabled business actually keeps from each dollar of revenue — the figure investors will increasingly ask for alongside the blended margin.

All four stand or fall on the numerator. Fully loaded means model and token charges plus the attributable share of cloud infrastructure — vector databases, orchestration compute, storage, networking — not model list prices. ServiceNow CFO Gina Mastantuono has said that AI reasoning is less than 10% of her company’s cost to serve; the rest is everything around it. Deciding what belongs in that numerator is a discipline of its own — see our guide to what counts as AI COGS.

Why did AI break SaaS unit economics?

Because inference reintroduced marginal cost to software. A SaaS product serves its ten-thousandth seat for close to nothing; an AI product pays for every token, so cost scales with usage rather than with seats sold — and the customers who love the product most are the most expensive to keep.

The margin consequences were visible early. Andreessen Horowitz observed in 2020 that AI companies often run gross margins in the 50–60% range, well below the 60–80%-plus benchmark for comparable SaaS businesses. That was before agentic workloads: an agent that plans, calls tools, retries, and checks its own work can consume many multiples of the tokens of a single chat exchange, with no human in the loop to feel the cost accumulating.

The 2026 reporting reads like a genre. TechCrunch describes a CTO’s engineer who ran up $40,000 in token charges in a single month — the CTO, in a story relayed by Faros AI’s chief executive, admitting he genuinely did not know whether to stop him — and Priceline facing a four-to-five-fold cost increase at its Cursor renewal. OpenAI’s head of enterprise says customer conversations have turned to a blunter question: we are spending so much — what visibility do you have?

None of this means the established disciplines failed. FinOps remains genuinely good at what it was built for — rates, commitments, discount instruments, and the engineering accountability loop around cloud efficiency — and the FinOps Foundation has been frank that token spending has pushed some members into what it called existential crises. The gap is not effort but denominator: cloud tooling reports cost per service, while pricing, margin, and budget decisions need cost per customer. Cost tools tell you what you spent. AI Value Management tells you what it bought. The comparison is drawn in full in FinOps vs AI Value Management.

What do good AI unit economics look like?

Direction beats absolutes. Published benchmarks for AI unit costs are scarce and heavily workload-dependent, so the healthiest signal is per-unit cost falling — or at least holding flat — while volume grows. The table below summarizes the patterns worth a CFO’s attention.

The a16z range is a reasonable anchor — an AI-attributed gross margin in the 50s is common, and one that climbs toward SaaS territory as the product matures is the goal — but the trend line matters more than the starting point. Reading these signals together, and turning them into a narrative a board will accept, is covered in measuring the ROI of AI initiatives.

A worked example

The numbers below are illustrative — a hypothetical company, not customer data. Call it Meridian: a B2B software firm whose product includes an AI assistant, serving 1,200 customer accounts and 90,000 monthly active end users.

Meridian spends $48,000 a month on model usage across its providers and a further $14,000 a month on attributable infrastructure — vector database, orchestration compute, storage, and egress. Its fully loaded AI cost is therefore $62,000 a month.

• Cost per MAU. $62,000 ÷ 90,000 MAU = $0.69 per active user per month.
• Cost per interaction. Users ran 2.4 million assistant interactions this month. $62,000 ÷ 2,400,000 = $0.026 — about 2.6 cents per interaction.
• Contribution margin per customer. Revenue is $144,000 a month across 1,200 accounts, so ARPU is $120. Attributable AI cost per account is $62,000 ÷ 1,200 = $51.67. Contribution margin: $120 − $51.67 = $68.33 per account — AI consumes 43% of ARPU.
• AI-attributed gross margin. ($144,000 − $62,000) ÷ $144,000 = 56.9% — inside the 50–60% range a16z observed for AI businesses, and well short of the SaaS benchmark.

What does Meridian’s CFO do with this? First, note that excluding the $14,000 of infrastructure would have reported a 66.7% gross margin — a ten-point flattering error, and the most common one. Second, watch the ratios move together: if Meridian ships agentic features, interactions per user will climb, and cost per MAU can rise even while cost per interaction stays flat. That is a pricing question, not an engineering one — and it is only visible if both metrics are tracked.

How do you instrument AI unit economics?

Three capabilities: attribution of every AI dollar, import of the business metrics that form the denominators, and reconciliation frequent enough to trust. A spreadsheet can manage this for one provider and one product line; it stops scaling the month a second team starts shipping.

Attribution. Every provider invoice line needs to be classified — by cost center, P&L category, product line, environment, and project — under rules that reapply retroactively when the org chart changes, with an audit trail behind every number. Without attribution, fully loaded cost per product is a guess.

Metrics and reconciliation. Revenue, headcount, DAU/MAU, and transaction counts need to live next to the cost data, imported by CSV or API, so the formulas compute from a single source. And the cost side should reconcile to provider invoices daily — usage can double inside a billing period, and a month-end surprise is a failure of instrumentation, not of forecasting.

This is the work COGScontrol productizes: it aggregates spend across OpenAI, Anthropic, AWS Bedrock, Azure OpenAI, and Google Vertex AI alongside AWS, Google Cloud, and Azure infrastructure, normalizes it into one ledger reconciled to invoice every 24 hours, and computes the metrics in this guide against imported business data — the full capability set is on the features page. Subscriptions are fixed; the price is never a percentage of AI spend.

To pressure-test your own numbers first, the free AI unit economics calculator implements every formula in this guide. For the broader operating discipline these metrics belong to, start with what AI Value Management is.