Agent products can look inexpensive when measured only by model tokens. Real workflows also use search, browser automation, external APIs, retries, storage and human review. The most useful unit-economics metric is therefore total cost per successful task: what the company spends to produce one outcome the customer actually accepts.

Define success clearly

A task is not successful just because the model returned text.

Use business completion such as booked meeting, reconciled invoice or accepted report.

Include every model call

Planning, tool selection, verification and repair all consume tokens.

Count hidden retries as part of the workflow.

Add tool costs

Search APIs, document parsing and browser services may charge separately.

These costs can exceed language-model spend in some agents.

Measure retry economics

A cheap model that fails often may cost more after repeated attempts.

Track cost per first-pass success and cost after recovery.

Include human review time

Approvals and manual corrections have real labor cost.

Agent economics should show when automation is only shifting work to reviewers.

Segment by workflow complexity

Simple extraction and multi-step procurement should not share one average.

Different task classes need separate margins and pricing.

Track latency beside cost

A cheaper workflow that takes ten minutes may deliver less customer value.

Cost optimization should not ignore time-to-outcome.

Use quality-adjusted success

Some outputs technically complete but require large user edits.

Correction rate can reduce the effective success metric.

Compare model-routing strategies

Route simple steps to cheaper models and expensive reasoning to stronger ones.

Measure whether end-to-end success actually improves unit economics.

Price for outcome variability

A flat subscription may work when task cost is stable.

Highly variable agent workflows may need usage limits or premium tiers.

Watch external API inflation

Tool providers can change price independently of model vendors.

Maintain cost attribution by dependency.

Use successful-task cost in product reviews

It creates a shared metric for finance, engineering and product.

Teams can discuss one workflow outcome instead of debating isolated token rates.

Gross margin by modality shows why blended AI cost can mislead. Agent task economics applies the same discipline to multi-step automation.

Management review 1: AI agent cost per successful task

Management should connect AI agent cost per successful task to cash timing, contribution margin, customer behavior and the assumptions used in the operating forecast. A metric is most useful when it has a clear owner, source system and review cadence rather than appearing only in a monthly spreadsheet after the underlying decision has already been made.

Scenario analysis should include a base case, a downside case and the operational action attached to each outcome. That makes the model useful for pricing, hiring and infrastructure decisions instead of turning it into a passive reporting exercise. Revisit assumptions whenever product mix, payment terms or model costs change materially.

Management review 2: AI agent cost per successful task

Management should connect AI agent cost per successful task to cash timing, contribution margin, customer behavior and the assumptions used in the operating forecast. A metric is most useful when it has a clear owner, source system and review cadence rather than appearing only in a monthly spreadsheet after the underlying decision has already been made.

Scenario analysis should include a base case, a downside case and the operational action attached to each outcome. That makes the model useful for pricing, hiring and infrastructure decisions instead of turning it into a passive reporting exercise. Revisit assumptions whenever product mix, payment terms or model costs change materially.

Management review 3: AI agent cost per successful task

Management should connect AI agent cost per successful task to cash timing, contribution margin, customer behavior and the assumptions used in the operating forecast. A metric is most useful when it has a clear owner, source system and review cadence rather than appearing only in a monthly spreadsheet after the underlying decision has already been made.

Scenario analysis should include a base case, a downside case and the operational action attached to each outcome. That makes the model useful for pricing, hiring and infrastructure decisions instead of turning it into a passive reporting exercise. Revisit assumptions whenever product mix, payment terms or model costs change materially.

Management review 4: AI agent cost per successful task

Management should connect AI agent cost per successful task to cash timing, contribution margin, customer behavior and the assumptions used in the operating forecast. A metric is most useful when it has a clear owner, source system and review cadence rather than appearing only in a monthly spreadsheet after the underlying decision has already been made.

Scenario analysis should include a base case, a downside case and the operational action attached to each outcome. That makes the model useful for pricing, hiring and infrastructure decisions instead of turning it into a passive reporting exercise. Revisit assumptions whenever product mix, payment terms or model costs change materially.