AI unit economics should be calculated around a delivered unit of value, such as an accepted support resolution or a completed analysis. Counting only tokens per request ignores retries, abandoned sessions, storage, external tools and human review. A clear worksheet identifies which costs scale with usage and which remain fixed.
The contract to define first
Define the billable or operational unit, then allocate model calls, retrieval, tool execution and review expense to it. Separate provider charges from engineering overhead and state the allocation method. Compare cost per accepted result with cost per attempted request so quality failures cannot disappear behind a low average inference bill.
A worked scenario
A hypothetical assistant handles ten requests but only eight meet the acceptance criteria. The cost of the failed attempts still belongs in the cost per accepted outcome. If a human reviewer finishes two tasks, include the relevant review assumption rather than describing those outcomes as fully automated.
An implementation and review sequence
- Define the delivered unit and acceptance condition
- Include failed attempts and human completion
- Separate variable costs from stated overhead allocations
Failure cases worth testing
| Condition or input | Expected behavior |
|---|---|
| Requests retry repeatedly | Their cost remains in the outcome ledger |
| A human completes an escalated task | The review assumption is included |
| Quality thresholds change | The unit-economics definition is versioned |
Download the illustrative acceptance checklist (JSON)
Tradeoffs and limits
Not every cost can be attributed precisely at an early stage. Use transparent approximations and sensitivity ranges instead of reporting a misleadingly exact margin from incomplete data.
Use the result to make a decision
The practical output is straightforward: allocate inference storage support and retry costs per active account. Keep the assumptions with the artifact so a product owner, implementer or prospective partner can review the same decision. A useful handoff states the supported scope, the evidence collected and the remaining dependency instead of treating a successful demonstration as proof of every operating condition.
Continue with related implementation detail
Related guide: Usage-based billing for AI SaaS: make every charge explainable
Cloud credits versus cash runway
References and current platform requirements
Microsoft for Startups: current programme overview
Google Cloud: startup programme
Platform capabilities, programme terms and access conditions can change. Check the linked primary references against the version and account you actually use; the patterns above do not establish a provider approval, certification or universal compatibility guarantee.
Discuss the workflow
Discuss the product scope and post-credit operating assumptions; AptixLabs does not guarantee grants or funding.
Explore the relevant AptixLabs service
What is the right unit for AI delivery cost?
AI unit economics should be calculated around a delivered unit of value, such as an accepted support resolution or a completed analysis. Not every cost can be attributed precisely at an early stage.
Which failure cases should be included in the review?
Include these explicit cases: Requests retry repeatedly: their cost remains in the outcome ledger; A human completes an escalated task: the review assumption is included; Quality thresholds change: the unit-economics definition is versioned.