Unit Economics

Unit economics answers a plain question: when the product adds one user, one task, or one contract, does it actually make money?

Traditional SaaS often treats gross margin as a relatively stable company-level number. Agent products cannot, because marginal cost is not zero. Inference, tools, sandboxes, external APIs, retries, and human review all happen per task.

Three Units To Analyze

Agent products need at least three economic units:

Unit	Question it answers
Task	Does each successful outcome make money?
User	Which users contribute margin, and which consume it?
Contract / organization	Is the customer worth serving, and does pricing or quota need adjustment?

Averages hide problems. A customer can pay a large contract but still have poor margin if most usage comes from high-cost, low-success workflows.

Task-Level Margin

Task-level margin is the foundation:

gross_margin_per_successful_task =
  revenue_per_successful_task
- cost_per_started_task / success_rate
- review_cost_per_task
- expected_fallback_cost

This deliberately uses “per successful task,” not “per started task.” Failed tasks still consume cost but may not be billable or valuable.

The cost fields come from Cost Model, but the metrics layer cares most about derived fields:

• cost_per_started_task
• cost_per_successful_task
• gross_margin_per_successful_task
• review_minutes_per_successful_task
• retry_cost_per_successful_task

User-Level Margin

Under the same subscription price, users with different intensity can produce very different margins. The following is a synthetic example showing distribution shape, not a pricing recommendation:

User type	Monthly tasks	Task pattern	Margin intuition
Light user	Few tasks	Low cost, low review	High margin, but value proof may be weak
Medium user	Steady tasks	Cost and value are balanced	Best baseline for pricing and retention
Heavy user	Many tasks	More likely to trigger high tokens, long runtime, and review	May create the most value, or consume the most margin

Heavy users are not automatically bad users. The question is whether their tasks are high value, whether they pay for overage, and whether routing and workflow design can lower cost. The problem is not “heavy usage”; it is a mismatch between heavy usage, pricing, and cost structure.

Contract-Level Margin

B2B agent products often have contract-level cross-subsidies: a few teams inside one organization use the agent heavily while others use it lightly. Looking only at total organization revenue can misread health.

A contract-level dashboard should break out:

• Task volume by team and workflow.
• Success rate and review time by workflow.
• Cost per successful task by workflow.
• Whether overage is billable.
• Whether a small number of users consume most of the budget.

At renewal time, margin distribution explains more than total ARR.

Pricing Responses

When unit economics deteriorate, price increases are not the only response.

Strategy	Good fit
Improve model routing	High-capability models are used for low-risk tasks
Compress tool results	Tool outputs are pushing up history cost
Add task caps	Subscription users have unbounded heavy usage
Charge overage	Heavy usage creates clear value and the customer will pay for increments
Redesign HITL	Human review costs exceed token costs
Exclude workflows	A task type is hard to verify, failure-prone, and low value

Pricing is often the last step. Many margin problems start in product and engineering: task scope is too broad, tool outputs are too large, success criteria are unclear, or routing is too coarse.

Do Not Report Only Average Margin

Average gross margin hides the tail. Distributions are more useful:

• By user percentile: P50 / P75 / P90 / P99 cost and margin.
• By task type: which workflows make money and which lose money.
• By model route: whether high-capability models are used for high-value tasks.
• By customer organization: whether contract-level margin risk exists.

For an agent product, “average gross margin is 70%” is less convincing than “P90 heavy users are still margin-positive.”

Cross-section Connections

• Cost field definitions: economics/cost-model
• Budgeting and ROI adjustment: economics/controls-and-roi
• Pricing mechanisms: pricing
• User-segment operations: playbooks