AI Unit Economics: A CFO Framework for Gross Margin and Cost to Serve

AI companies often inherit software metrics and then discover that the cost structure does not behave like conventional software. The product may carry variable model charges, retrieval and storage costs, human review, implementation labour, or unusually expensive edge cases. A single blended cloud number cannot explain which customers or workflows are economically attractive.

The CFO problem is not to invent an exotic new accounting system. It is to build an operating view that connects usage, delivery cost, customer value, and cash.

Start with the economic unit

The right unit is the smallest recurring activity that meaningfully connects what the customer receives to what the company spends.

Common units include:

• active customer or workspace;
• user seat;
• document, image, or minute processed;
• model request or token volume;
• workflow completed;
• agent task completed successfully; or
• business outcome delivered.

Requests and tokens are easy to count, but they may be weak commercial units. A workflow could require five requests this month and twenty next month after a model change. A customer buys the completed work, not the token. Finance should retain both views: the technical cost driver and the customer-facing unit of value.

Build cost to serve in layers

A useful model separates five layers rather than relying on one gross-margin percentage.

Layer	Typical costs	Management question
Model	API inference, fine-tuning, dedicated capacity	What does each workload cost at current quality?
Infrastructure	Compute, storage, retrieval, observability, data transfer	Which costs scale with usage and which are committed?
Data	Licensed data, labelling, enrichment, human feedback	Is data cost recurring, customer-specific, or product investment?
Delivery	Implementation, prompt or workflow configuration, human review	How much labour is required to make the product work for this customer?
Support and reliability	Direct support, incident response, customer-specific monitoring	Which accounts consume disproportionate operating attention?

This is primarily a management model. The mapping to financial-statement presentation should be reviewed with the company’s accounting advisers and applied consistently.

Separate three kinds of margin

One margin number cannot answer every question. Management benefits from three views.

Infrastructure contribution

Revenue less model and infrastructure costs. This shows the underlying computational economics and helps technical teams compare architecture or vendor choices.

Delivered gross margin

Revenue less all direct delivery costs, including implementation and human operations. This is the closer view of whether the current product can scale economically.

Customer contribution

Delivered gross profit less customer-specific support or success costs. This helps identify accounts whose contract value looks attractive but whose operating demands consume the margin.

The definitions should be written down. Changing the definition every quarter makes trend data almost useless.

Measure distribution, not only averages

AI usage is often uneven. A small group of customers, tasks, or prompts can drive most of the cost. Averages conceal that pattern.

At minimum, review:

• median and high-percentile cost per workload;
• gross margin by customer cohort and pricing plan;
• cost per successful outcome, not only per attempt;
• human-review rate and rework rate;
• contracted minimums versus actual usage;
• margin before and after implementation effort; and
• the customers with the largest absolute gross-profit leakage.

A 75% blended margin can coexist with a group of negative-margin enterprise accounts. Those accounts may still be strategically valuable, but management should know the size and reason for the subsidy.

Connect product quality to cost

The cheapest model is not automatically the best economic choice. A lower inference cost may create more failed tasks, retries, human review, or customer churn. Conversely, using the most capable model for every request may add cost without improving outcomes.

The finance model should therefore sit beside the evaluation system. For each important workflow, connect:

cost per attempt × attempts per successful outcome + review cost = cost per successful outcome

That equation creates a common language for finance, product, and engineering. Model-routing decisions can then be evaluated against both quality and contribution margin.

Use scenarios before negotiating contracts

Before approving a material customer contract, test at least four cases:

1. expected usage and expected quality;
2. high usage with the existing architecture;
3. lower automation with more human review;
4. a vendor-price or model-mix change.

Then compare the contract’s price floor, usage allowances, service requirements, implementation commitments, and renewal mechanics with the cost range. This is especially important when a customer requests unlimited usage or extensive custom work.

A monthly CFO dashboard

The monthly review does not need dozens of metrics. A strong starting set is:

• revenue and delivered gross margin by product;
• cost per successful core workflow;
• model and infrastructure cost as a percentage of revenue;
• human-review and exception rate;
• implementation hours for new customers;
• top five margin-positive and margin-negative accounts;
• forecast versus actual volume and cost; and
• the two or three product decisions with the largest economic impact.

Each metric should lead to an owner or a decision. If it never changes behaviour, it is probably reporting noise.

The goal is decision quality

Unit economics are not an investor-performance ritual. They tell management where product value is being created, where delivery complexity is being subsidised, and which technical or commercial changes can improve the business.

For AI companies, the most useful finance function does not stand outside the product system. It translates that system into an economic model the whole management team can use.

This article is general business information, not accounting, tax, legal, investment, or securities advice. Financial-statement classification should be confirmed with appropriately qualified advisers.