Cost per Task

Cost per task measures what it costs an AI system to get a piece of work done. A single task often involves more than one model call: an agent might plan, search, call tools, check its work and revise, while a support assistant might retrieve documents, draft a reply and run a safety check. Cost per task adds up everything involved, including tokens across all calls, tool and retrieval charges, and the cost of failed attempts that had to be repeated. Related measures include cost per query, cost per conversation and cost per resolution, each defined around the unit of work that matters for a given application.

Cost per task is often the most useful single measure in AI economics. Token prices and per-request costs describe inputs, but organizations care about outputs: how much it costs to resolve a ticket, process an invoice or answer a question well. For IT financial management, cost per task allows fair comparison between models and designs, connects AI spending to business value and provides the basis for budgets, pricing and targets.

Definition of Cost per Task

Cost per task is the total cost of all AI resources consumed to complete one defined task, divided across successfully completed tasks or across all attempted tasks, depending on the purpose.

The task must be clearly defined, such as a customer conversation that ends without escalation, or an invoice with all fields extracted correctly. A precise definition is what makes the measure comparable over time and between options. Changing the definition, for example to count partial resolutions, should be recorded so that trends are not misread.

What It Includes

Model costs include all input and output tokens across every model call in the task, including reasoning tokens and calls to helper models such as routers or classifiers.

Supporting costs include embeddings, vector database queries, tool calls such as web search, and any charges for code execution or other services used during the task.

Failure costs include tokens spent on attempts that failed or were retried, and, where relevant, the cost of human review or correction. Including these gives a truer picture than counting only successful calls.

Variations

Cost per attempted task divides costs by all tasks started, while cost per completed task divides by tasks finished successfully. The second is higher when failure rates are significant and is usually more meaningful.

Some organizations calculate a fully loaded cost per task, adding shares of platform, engineering and support costs, which is useful for business cases and pricing.

Tracking the distribution of cost per task, not just the average, reveals expensive outliers, such as agent tasks that loop. These outliers are often where the largest savings lie.

Importance of Cost per Task in IT Financial Management

Cost per task connects AI spending to the work it produces. It shows whether an application is becoming more or less efficient, independently of changes in volume.

It allows fair comparison between models and designs. A model with a higher price per token may have a lower cost per task if it needs fewer calls or fails less often, as shown in cost per token.

It is also the natural link to value. Comparing cost per task with the value of each task, such as the labor cost it replaces, shows whether an AI use is economic.

Budgeting and Pricing

Budgets for AI applications are best built from expected task volumes multiplied by cost per task, rather than from total spending last year.

Companies selling AI-enabled products use cost per task to set prices and usage limits, ensuring each task, on average, earns more than it costs.

Cost per task targets give teams a clear goal for optimization and a way to show progress. Targets should be set together with quality targets, so that savings do not come at the expense of results.

Agents and Variability

Agent tasks vary widely in cost, because the number of steps and the size of the context window differ from task to task. Averages can hide a long tail of very expensive tasks.

Monitoring high percentiles and setting limits on steps or spending per task keeps this tail under control.

Cost per task is central to evaluating AI agent pricing models, which often charge per task or outcome.