AI Unit Economics: Measure Cost per Successful Workflow

Start with one question: what does it cost to deliver a successful unit of work, and does the contract pay enough for the full customer obligation? For an AI company, a request, a completed workflow, and a customer account can produce different answers. Finance needs all three views, with a clear bridge between them.

This is Amafi Advisory’s practitioner method for a U.S. AI or technology management team. It is a management analysis, not a rule for classifying expenses in financial statements. The company’s accountant should determine that presentation.

Define the customer unit and the technical drivers

Choose a unit the customer recognizes, such as a document processed to an agreed standard or a workflow completed. Define what counts as completed, how failed attempts and rework are recorded, and which customer and period receive the result. Keep the raw technical drivers beside it: attempts, model calls, input and output tokens, retrieval, storage, and review time.

This separation matters because a single completed workflow may require several attempts. OpenAI’s API pricing distinguishes input, cached input, and output token charges; Amazon Bedrock pricing also describes on-demand and provisioned capacity models. Both sources were accessed November 3, 2026. They illustrate possible billing structures, not a universal cost mix or a recommendation to use either provider. The unit and allocation choices below are Amafi Advisory’s practitioner judgment.

Build a cost bridge from invoices to customers

For each monthly close, connect five records rather than relying on one blended cloud expense:

1. Usage: Count completed workflows, failed attempts, retries, and model or cloud usage by product and customer where the data permits.
2. Provider charges: Match metered usage to invoices. Record credits, minimum commitments, reserved capacity, and invoice timing separately so a low-usage month does not look artificially cheap.
3. Human delivery: Record review, implementation, and direct support time against the customer or workflow that required it. Keep shared product development outside the customer delivery view unless there is a defensible allocation.
4. Allocation: Assign shared capacity using a stated driver, such as workload volume or reserved capacity. Show unallocated or idle capacity as its own line rather than hiding it in a successful workflow average.
5. Reconciliation: Tie the management totals to the general ledger and name every difference in period, allocation, or accounting classification.

The result should show both cost per successful workflow and cost to serve each customer. A customer with few workflows may still require substantial implementation and support. Conversely, a high-volume customer may consume expensive model capacity even with little human attention.

A worked monthly example

The numbers below are invented for arithmetic only. They are not an Amafi client result, market benchmark, or target margin.

• A product completes 10,000 workflows after 12,000 attempts in a month.
• Metered model and cloud charges attributable to those workflows total $3,000. Direct human review costs $900.
• The company also pays $1,100 for committed capacity attributable to this product, including unused capacity. It records that amount once, separately from the metered charges.
• The management delivery view is therefore $5,000, or $0.50 per completed workflow. Dividing by 12,000 attempts would answer a different question and understate the cost of a successful workflow.

If those workflows earned $12,000 of revenue in the same period, the initial delivery margin is $7,000 / $12,000, or 58.3%. That is an incomplete contract view. Suppose the same customers also needed $800 of implementation work and $200 of direct support recorded outside the initial $5,000. The management cost to serve becomes $6,000, leaving $6,000 / $12,000, or 50%. These added amounts are also invented; the two percentages are different definitions of the same example, not reported gross margins.

The two views answer different management questions:

• Workflow delivery, 58.3%: Is the product’s model, capacity, and review route efficient for completed work? The calculation excludes implementation and direct support.
• Full customer delivery, 50%: Does the contract cover the work required to serve these customers? This adds implementation and support to the same month’s cost population.

Before using either view for a price decision, verify that the revenue, completed workflows, provider charges, and human work cover the same customers and period. Record any deferred implementation revenue or cost timing separately; ask the accountant how the amounts belong in the financial statements. If a capacity invoice is already included in metered charges, remove the overlap rather than adding it twice.

Keep reported and management margins distinct

Reported gross margin follows the company’s accounting classifications. A management delivery margin can include costs to serve a customer that sit elsewhere in the accounts. Neither figure should silently replace the other. A short bridge can show reported cost of revenue, directly attributable costs recorded elsewhere, allocation adjustments, and the resulting management delivery cost. Document the allocation method and any change to it.

Use the bridge to ask which decision changes: revise included usage, change a model route, charge for implementation, narrow a service promise, or accept a temporary subsidy with an explicit owner and review date. See the AI pricing strategy guide for contract guardrails and the KPI definition framework for keeping the margin measure comparable in a management pack.

Review the distribution and the contract

An average can hide expensive workflows or customers. Review cost per completed workflow by use case, customer, and plan, along with retry rate, review time, committed capacity use, implementation effort, and support demand. Investigate outliers before changing the standard price or model configuration. The cheapest model per request may require more retries or review, so compare cost per successful result at an agreed quality level.

For a material contract, run expected and high-usage cases against included usage, overage terms, implementation scope, support obligations, and any capacity commitment. A monthly fractional CFO engagement can maintain this review cadence; a defined corporate finance project can establish the initial cost bridge and pricing analysis. To discuss the underlying decision and available data, use the contact form.

Common questions

What should an AI company include in cost of revenue?

For management analysis, trace model usage, infrastructure, data, human review, implementation, and directly attributable support to the product or customer. Then reconcile that cost-to-serve view to the accounts. An accountant should determine financial-statement classification.

Why can reported gross margin differ from delivery margin?

A management delivery view may include customer-specific costs recorded outside accounting cost of revenue. Keep reported gross margin intact, identify each added or removed cost, and show the reconciliation rather than relabeling either measure.

Which unit should an AI company measure?

Measure a completed customer workflow or another unit tied to customer value, and retain attempts, tokens, and provider charges as internal cost drivers. Also review economics by customer and contract because implementation and support may not scale with workflow count.

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