Unit economics
AI unit economics measure the cost and value of an AI-enabled workflow per defined business unit. They translate a total AI bill into decision-ready measures such as cost per resolved case, value per approved analysis, or margin per AI-assisted transaction.
The unit should represent accepted work or a meaningful business outcome. Cost per token and cost per request can explain technical consumption, but they do not show whether the organization received useful work.
How are AI unit economics calculated?
Two basic measures are:
Cost per accepted unit = attributable AI-enabled workflow cost / accepted completed units
Net value per unit = attributable financial benefit per unit - attributable cost per unit
The cost boundary may include licenses, model and infrastructure usage, retrieval systems, integration, monitoring, human review, rework, and allocated implementation costs. Use the boundary that matches the investment decision and state it explicitly. AI token cost and AI inference cost describe the technical consumption inside that boundary, and AI total cost of ownership widens it across the full lifecycle.
The FinOps Foundation defines unit economics as connecting technology spending with the value it creates. Its guidance distinguishes technical units from business units and notes that generative AI measurement can progress from cost per token toward outcome-oriented measures such as cost per assist or case deflected (FinOps Unit Economics).
How do you choose the right unit?
Start with the workflow owner's completion event. The unit of work should have a clear definition, acceptance rule, source record, and treatment for duplicates, reopened work, or partial completion.
Choose a unit that can change a decision. Cost per model call may help engineers optimize routing. Cost per accepted case may help operations compare AI-assisted and non-AI workflows. Net value per customer or transaction may help finance decide whether to expand the investment.
Avoid combining materially different work without segmentation. A routine request and a complex escalation may have different cost, time, and quality requirements. Report raw units and any normalization method so stakeholders can inspect the comparison.
How do cycle time and quality affect unit economics?
Cycle time may fall while cost per accepted unit rises if the workflow creates more retries, review, or correction. Conversely, a more expensive model may improve acceptance enough to reduce total unit cost.
Track quality guardrails with the unit metric. Useful measures can include correction rate, escalation rate, acceptance rate, defect severity, or another workflow-specific standard. This prevents a team from optimizing the denominator by declaring unfinished or poor-quality outputs complete.
The FinOps Foundation's AI cost forecasting guidance identifies cost per unit of work as a useful measure and pairs cost forecasting with speed and response quality (FinOps AI cost forecasting).
How do unit economics support investment decisions?
Unit economics make scale assumptions visible. If volume grows, leaders can model whether marginal cost, review capacity, and quality remain acceptable. They can also compare vendors or workflow designs using the same accepted unit rather than incompatible billing metrics.
For an AI payback period, unit economics can estimate recurring net benefit when volume is known. For AI ROI, they provide the operational bridge between total investment and realized financial outcome.
An AI readiness assessment should confirm that cost allocation, completion data, and quality records are available before the team promises unit-level measurement.
Common unit-economics mistakes
Common failures include using tokens as the business unit, omitting implementation or human-review cost, mixing accepted and unreviewed outputs, comparing different workflow scopes, and treating estimated value as realized value.
A useful unit metric must remain traceable. Record the formula, sources, allocation rule, period, owner, and caveats. If the organization cannot reproduce the number, it should remain a working estimate rather than an executive performance claim. An AI measurement framework records the source, owner, formula, period, and limitation for each metric.