From Token Cost to Workflow Economics

For an AI SaaS customer, the important economic unit is often larger than the model call that appears on the invoice. This chapter builds a workflow-level account of an AI deployment: how task fit affects performance, how deployment cost extends beyond tokens, how reliability creates expected loss, and how those elements combine into net value. The result is a model the reader can use to compare an AI service with the process it changes, rather than with a benchmark or a single cost line.

The model comes from Agentomics, which treats human and artificial agents as productive resources in an organizational workflow. That framing is useful because an AI agent has two profiles at once. It has capabilities that determine what work it can perform, and it has economic attributes that determine what the organization must spend and risk to use it. A pricing decision that sees only one profile will systematically miss part of the deployment.

An Agent Has a Capability Profile and an Economic Profile

Traditional software is often evaluated as a tool with defined functions. An AI agent can also reason over context, invoke tools, maintain memory, execute parts of a workflow, and adapt its behavior to changing conditions. The relevant economic question is not whether those capabilities are impressive in isolation. It is whether they fit the task for which the customer is paying and whether the surrounding organization can use the result.

Agentomics represents each workflow stage by a requirement profile. A stage may require domain knowledge, contextual reasoning, low latency, uncertainty handling, policy compliance, security sensitivity, or human-verifiable justification. The AI agent has its own capability profile: perhaps strong retrieval, tool use, speed, or narrow domain expertise, but weaker performance under ambiguity or distributional change. The gap between the two profiles matters before any token is consumed.

The paper introduces a coverage measure, written as μ(d, m), to represent how much of a stage’s requirements an agent’s capabilities cover. The notation is less important than the relationship. An agent can be highly capable in general and still be poorly matched to a particular task. Poor fit can lower the probability of a correct result and increase the effort required for supervision, correction, or re-execution. A customer may therefore pay the same API price for two tasks while receiving very different economic performance.

Stage performance is represented by a success probability. In the paper’s general formulation, that probability can depend on the agent, the task, the degree of capability coverage, the context, and environmental uncertainty. The paper also gives a simple multiplicative specification in which baseline agent performance is adjusted by coverage. It presents that specification as a convenient special case, not as a law of AI behavior. In a real AI SaaS deployment, the relationship would need to be estimated from the tasks and conditions that actually matter.

The agent’s economic profile is equally important. A market price can represent a subscription, license, or API charge. Operational cost can include computing, infrastructure, monitoring, communication, latency, verification, supervisory effort, and other expenses incurred when the agent is assigned to a stage.