AI agents: how to measure cost per completed task in your operation

To measure the cost of an AI agent, divide the operating cost of the workflow by the number of completed and accepted tasks in the same period. Include technology usage, human review and rework. Track the initial investment separately when assessing expansion.

This answers a concrete question: how much does it cost to deliver a usable quote, record a correct order or confirm a valid appointment? An ended conversation may leave work unfinished. A fast response may require correction. Follow the output the team can actually use in the next step.

Define the output that counts

Choose a unit of work with a beginning, an end and an acceptance rule understood equally by operations and the implementation team.

• Logistics: a prepared quote, with all required data and calculations checked against commercial rules.
• Manufacturing: an order handed to production, with customer records, items and terms validated and accepted by the destination system.
• Automotive retail: a service appointment with vehicle, location, service and time confirmed, without duplicates.

These are examples of measurement design, not customer results. The process owner should adapt acceptance criteria to the business.

The destination system needs to confirm completion. Anthropic distinguishes between what an agent says it did and the final state of the system. Evaluation should verify the actual outcome. Source: Anthropic.

Include the cost of the whole workflow

Cost per completed task = operating cost of the workflow during the period ÷ completed and accepted tasks during that period.

Include applicable costs:

• Models, infrastructure and tools allocated to the workflow.
• Telephony or messaging, where used.
• Time spent reviewing, correcting and handling exceptions.
• Recurring support and maintenance attributable to the process.

A task requiring three attempts incurs costs on all three. Count its completion once. A cancelled request may also consume resources: include that cost and record cancellation separately.

Choose a consistent allocation rule for shared expenses. Otherwise, an apparent improvement may simply reflect an accounting change. Keep implementation investment on a separate line. Mixing it with a single week's consumption makes comparison difficult. An expansion decision should include both recurring costs and additional investment.

Example: cheaper technology can mean a more expensive operation

Imagine two designs for the same workflow, evaluated during the same period with requests of comparable complexity. All values below are hypothetical, in Brazilian reais.

• Design A: R$2,000 for technology and channels plus R$6,000 for review, rework and support. Total: R$8,000. With 800 accepted tasks, the cost is R$10 per completed task.
• Design B: R$3,500 for technology and channels plus R$2,500 for review, rework and support. Total: R$6,000. With 1,000 accepted tasks, the cost is R$6 per completed task.

Design B costs more in technology but less per accepted delivery in this example. Approval still requires checking quality, turnaround time, incoming volume and implementation investment.

HappyRobot's recent content similarly discusses evaluating model costs together with process outcomes. This reference informs the evaluation question; it does not imply that Magical provides HappyRobot's platform features. Source: HappyRobot on costs.

Track quality and turnaround alongside cost

Keep a short operational dashboard:

• Cost per accepted delivery: the cost of completing the work.
• Completion rate: the share of eligible requests that reach acceptance.
• Rework rate: deliveries requiring correction.
• Time to acceptance: the wait for the customer or the next team.
• Human intervention: where people take over and why.

Separate routine work from exceptions. A recurring logistics route and a quote with exceptional commercial terms may require different effort. In automotive retail, confirming an available appointment differs from resolving a mismatch between service, vehicle and schedule.

Also define errors that prevent expansion. An incorrect order or an unauthorized commercial commitment should not disappear inside a favorable cost average.

Compare versions under similar conditions

Before changing the model or workflow, record the current version and acceptance rule. Compare periods or groups with similar request mixes, making differences in channel, location, shift and complexity visible.

Where volume and conditions allow, a controlled experiment can help distinguish the impact of a change from operational fluctuations. HappyRobot describes A/B testing with a primary metric defined before analysis. Source: HappyRobot on experiments.

At low volume, treat conclusions as preliminary. Review individual cases and gather evidence before extrapolating across the company. An improvement at one location may depend on data or practices not yet available elsewhere.

Give freed capacity a purpose

The RBR Transportes project illustrates the value of measuring a concrete task: the case describes freight quotes produced in seconds, compared with a manual spreadsheet workflow reported at around 30 minutes, with costs and margin in the same process. That outcome belongs to the project's scope; the case does not provide the cost inputs required to calculate this article's metric.

In your operation, freed time may support more requests, shorter queues or better commercial analysis. Record which change actually happened. Available hours do not automatically represent lower expenses or additional revenue.

If the goal is growth, track downstream results: proposals sent, orders accepted or services actually delivered, depending on the workflow. This connects evaluation to the outcome that justified the project.

Decide the next step with operations

Expansion makes sense when outputs meet acceptance criteria, total cost fits the operation and exceptions have owners. If human review absorbs the benefit, investigate where work returns. Missing data, integration or unclear rules may be the cause.

The solution may combine AI agents, process automation and custom development, depending on the work that needs to improve.

Want to identify the right measure for your process? [Talk with a Magical partner](https://magicalsolutions.ai/#conversa) about your operation, the opportunity and a pilot that lets you evaluate results before expanding.