What does a voice agent cost? Calculate cost per completed task

Cost per task: define the counted outcome

Define the denominator before adding expenses. A received request, an answered transfer and a confirmed booking are different outcomes. To calculate completed-task cost, choose a verifiable outcome and count only instances meeting that criterion in the period. Keep unsuccessful-attempt costs in the total because they consume resources too.

Record eligible requests and exclusions. Comparisons change meaning when the agent handles simple cases and the team receives only difficult ones.

Count each cost component once

Include attributable platform, telephony, integration and human-work costs. Separate initial implementation from recurring maintenance. Avoid counting components already included in contracted charges twice.

Tigy tracks usage and credits by workspace. Consult current billing for your account; this article does not fix prices or credit-to-minute equivalence.

Reconcile measurement sources

Reports aggregates the selected interval; Runs details conversations; Billing shows cycles and credit history. Check agent, workspace, period and timezone before combining figures.

A calendar month may differ from a billing cycle. State the interval and cost allocation when source dates differ.

Interpret cost alongside quality

Divide attributed costs by verified completed tasks. If none completed, the indicator lacks a useful denominator: investigate the pilot rather than claim efficiency.

Track failures, repeat contacts and team effort alongside the value. This is operational analysis rather than guaranteed savings.

An illustrative calculation with an explicit denominator

Assume a fictional pilot with 100 eligible requests, 80 verified outcomes, BRL 120 technical consumption, BRL 30 telephony, BRL 50 allocated maintenance and BRL 100 human work. Total cost is BRL 300; completed-task cost is 300 / 80 = BRL 3.75. Received-request cost is BRL 3.00, a different measure.

These values are teaching assumptions in Brazilian reais, not Tigy prices, results or savings. Maintenance allocation depends on the selected period. Include unsuccessful attempts in pilot cost and verify outcomes at the destination without excluding difficult calls to improve averages.

Compare previous service using similar requests, equal intervals and identical cost components. Record excluded expenses. Measure changed human effort instead of assuming every ended call removed staff work.

Separate implementation from recurring operation

Separate initial configuration and development from recurring review and maintenance.

If allocating setup costs, disclose horizon and volume assumptions. Small pilots and larger operations have different overheads.

Distinguish observed consumption from forecasts and avoid converting illustrative arithmetic into savings promises.

Define a completed task before dividing costs

Cost per task depends on what you spend and what counts as completion. For appointment service, collecting interest is different from reserving a confirmed slot. In support, providing guidance does not establish that the issue was resolved. Choose a definition verifiable in the responsible system or through an explicit operational criterion. Otherwise, a team may declare efficiency while increasing the number of people who must call again.

A fictional calculation makes the difference visible. If a period costs 900 currency units and ends with 300 received requests, cost per request is 3. If only 180 reach the operation's defined confirmed outcome, cost per completed task is 5. Neither figure is inherently wrong; they answer different questions. Name the denominator so contact cost cannot be presented as resolution cost. These values illustrate accounting, not Tigy prices or promised savings.

Define treatment of partial service. A conversation identifying an intention and transferring to staff can create value without completing the primary task. Record that outcome separately and calculate its cost when useful. Do not combine prepared handoffs and automated completions as equivalent results. Also classify abandonment, technical failures, and contacts with no recognized intention. If these cases disappear from the denominator and their costs disappear from the numerator, the metric loses connection with real spending.

Consider the observation window. An appointment may be created during a call and soon canceled because it was duplicated. A request may remain pending until staff completes it the next day. State when results are checked and who completed the task. That clarity supports period comparisons without concealing procedural changes. For investment decisions, show cost per contact, cost per confirmed outcome, and human involvement together, each with its own definition. Separating these measures also prevents teams from interpreting reduced human time as proof that all requested outcomes were achieved.

Choose the expenses included in analysis

An operational analysis must declare its boundary. Usage spending during calls is only one possible cost component. Depending on the question, include telephony, external services, integration maintenance, content review, and human recovery time. Compare equivalent categories across alternatives. Comparing a current team's fully loaded cost with just one technical automation component creates an artificial advantage even if each individual figure is correct.

Separate recurring and implementation expenses. Initial configuration and test preparation may be upfront costs. Policy review and incident monitoring continue after release. If implementation is spread across several months, record the chosen period and also show the calculation without that allocation when relevant. The choice changes the result and should remain visible. No single allocation period fits every operation; seasonal tasks and daily service have different exposure.

Use observed consumption and commercial conditions verified by the team. Tigy's usage and billing documentation helps identify where to inspect the account, but it does not replace checking values applicable to your environment. Do not infer current prices from an editorial example. For external systems, determine whether tool calls incur charges and whether unsuccessful attempts count toward consumption. A repeated lookup may produce no additional result while still consuming resources.

Measure human time with a simple rule. A sample can record review and recovery minutes by task, including checks of uncertain operation states. Multiply that time by the organization's chosen labor cost and state the assumption. Do not treat every transfer as the complete cost of one avoided human interaction: the agent may reduce initial data collection while staff still completes the case. Label estimates and show a range instead of an apparently exact single number. Maintain separate totals for directly observed invoices and modeled labor, making later revisions possible without confusing cash expenditure with allocated operational effort.

Compare similar tasks and investigate extremes

An overall service average mixes complexity. Giving an address, checking an order, and correcting a profile require different resources. Group outcomes by intention and relevant difficulty. If one month contains more simple questions, average cost may fall without any agent improvement. If the next contains more exceptions, the average can rise despite better performance within every task. Contact composition must accompany the metric.

Investigate unusually long conversations before concluding that pricing is responsible. Causes may include repeated identification, lengthy explanations, silence while a tool fails to respond, or repeated attempts without a defined alternative. Each needs a different correction. Shortening the opening may improve presentation but will not fix a slow external lookup. Removing confirmations may shorten calls while increasing write errors. Assess the gain alongside completion and quality rather than duration alone.

Use a fictional example for each segment. A simple lookup may cost less and have high completion. Profile changes may cost more because they require confirmation and validation. Do not automatically eliminate the second segment if it delivers a valuable outcome and reduces rework. The choice follows operational purpose. For infrequent exceptions, human routing may be more appropriate than a complex integration, but include time staff spends repeating initial collection.

When testing a change, observe the same task set and a comparable window. Record source updates, tool changes, or changes in human handling. Track cost per confirmed outcome, repeat contact, and correction requirements. Lower consumption accompanied by more callbacks can simply move expenditure to another channel. Look for the procedure's total effect. A useful number guides an action and remains valid when the person's next interaction is considered. Keep difficult cases visible rather than removing them from reports, since that would improve the appearance of cost while leaving the operating burden unchanged.

Turn the calculation into an operating decision

When presenting the analysis, state which change the figure supports. High cost per task may suggest narrowing initial scope, improving a slow integration, or revising answers that cause repeat calls. The figure alone does not determine the action. Add conversational and external-stage evidence connecting the metric to a cause. If the conclusion depends on an unmeasured assumption, make it explicit and choose a small verification exercise.

Define the next test's boundary: included tasks, observed spending, and an outcome that must improve without increasing errors. Compare before and after under identical rules. The team should be able to explain why it expanded, adjusted, or stopped a task. That explanation is more useful than a generic savings promise because it describes what automation actually delivers in the operation's context. Assign a review date and an owner to each decision, particularly when projected savings depend on reduced human recovery work that has not yet been observed.