Agent Unit Economics & Cost per Successful Task
What does a verified useful outcome cost after failures and human work are included?
Reference notes by Min · Reviewed October 15, 2026
Definition
Agent unit economics relates a defined cost pool to a meaningful unit of delivered work. Cost per successful task divides the costs of all relevant attempts by the number of outcomes meeting an explicit acceptance standard.
The unit should reflect the decision the organization is making. Cost per model call can identify infrastructure waste, while cost per correctly resolved case helps compare service designs. Neither measure is sufficient alone: a low per-case cost can coexist with unacceptable errors or poor access for difficult cases.
What it covers
- Cost boundary
- Include model calls, tools, infrastructure, retries and attributable human work.
- Outcome
- Define successful, failed, escalated and abandoned tasks consistently.
- Time horizon
- Separate marginal operating expense from setup and ongoing fixed costs.
- Decision
- Compare options on the same workload and quality requirements.
Define the acceptance window and population before reporting results. A case reopened the next day may not represent durable resolution. Report escalations and refusals separately; appropriate escalation can be good governance even when it does not count as an autonomous success.
In practice
Practical design patterns by Min.
- Record a stable task ID linking attempts, costs and verified outcomes.
- Show both variable cost and a separately labeled fully allocated cost.
- Compare with the current process, including review and rework on both sides.
An operational calculation can be written as total attributable cost divided by accepted outcomes over the same period. Keep failed attempts in the numerator. If there are no accepted outcomes, report the metric as undefined with the observed spending rather than displaying a misleading zero.
Research notes
Source-grounded details and editorial applications. Follow each link for the original document.
Relate spending to a meaningful unit
The FinOps Foundation distinguishes technical efficiency measures from business unit measures. Applying that distinction to accepted agent outcomes is the editorial approach used here.
Benchmark efficiency is narrower than business value
Wang and colleagues evaluate efficiency and effectiveness on GAIA using a cost-of-pass metric. Their results do not establish the total operating cost or return of an enterprise deployment.
Working example
A cheaper attempt can produce a more expensive outcome
Illustrative monthly data: system A costs $600 for 1,000 attempts and delivers 750 accepted outcomes; B costs $500 and delivers 500.
- Calculate A at $0.80 and B at $1.00 per accepted outcome.
- Include failed-attempt spending in each numerator; do not divide only successful-run costs.
- Add comparable review and operating costs before deciding whether either system creates value.
Why the controls matter
The example compares equal attempt volumes but different accepted outputs. It does not imply that A is automatically preferable: the systems may serve different case mixes or produce different harms. Compare matched task classes and inspect critical failures before acting on the aggregate number.
Common pitfalls
- Calling a generated answer a successful business outcome.
- Counting time saved as cash savings without a realistic capacity or spending change.
- Excluding retries, reviewer effort or failed tasks from the cost comparison.
Maintain separate views for short-run optimization and investment decisions. Setup, evaluation, security review and ongoing maintenance matter to total cost of ownership. Use a stated time horizon and avoid spreading fixed costs across an optimistic future volume that has not materialized.
Questions for operational review
Use these questions to decide what evidence your team needs. They are editorial prompts, not a compliance checklist.
- What independently establishes an accepted outcome?
- Which expenses and failed attempts are included?
- Does the comparison hold for difficult task classes?
Articles linking here
References
Primary documents reviewed for this entry. Examples and implementation patterns are editorial synthesis.
- FinOps Foundation: Unit Economics ↗Accessed October 15, 2026
- Wang et al.: Efficient Agents (2025 preprint) ↗Accessed October 15, 2026