Aircall charges an origination fee on every outbound AI Voice Agent attempt, even when the call never connects. That rule exposes the problem with treating a platform's per-minute figure as the cost of a call. A failed attempt can consume money and capacity before it produces one useful outcome.
Start with a call-level ledger rather than a monthly invoice. Give every conversation one identifier, then attach its telephony duration, transcription minutes, language-model tokens, synthesized characters, platform time, tool calls, retries, and transfer charges. That baseline matters because the hidden price of an AI voice stack grows across services that meter different units.
Containment needs the same discipline. An agent can avoid a human transfer by giving the wrong answer, losing the caller, or ending early. Count that as failure, not cheap automation, and keep unsafe or policy-required transfers in their own valid-outcome category.
Your useful metric is cost per correctly resolved call. Add every billable cost generated by a cohort, including unsuccessful attempts, then divide that total by the number of verified successful outcomes. The denominator must exclude calls that merely connected, reached voicemail, or produced an incomplete backend action.
Segment the calculation by intent before averaging anything. Appointment confirmation, collections, sales qualification, and technical support have different talk times, tool use, transfer policies, and acceptable outcomes. One blended number lets an efficient workflow hide a costly one.
TTS cost follows what the agent says, not simply how long the caller stays connected. LLM cost follows the tokens processed across turns, including repeated context and tool results. A quiet or slow call can therefore carry a different component mix from a brisk call of the same duration.
Warm transfers deserve separate records. Depending on the platform, a proxy agent can accrue talk-time charges while it calls the human, followed by a distinct transfer rate after the legs merge. Some services also bill voicemail as normal call time, so a campaign with poor answer rates can look inexpensive per connected call while wasting money across the attempt pool.
Fixed fees belong in the ledger too. Allocate phone-number rental, platform subscriptions, reserved capacity, recording storage, monitoring, and support across the calls that used them. Keep engineering labor outside the vendor subtotal, then add it as a separate operational layer when manual review or correction is required.
Billing granularity changes the answer. Exact-second pricing rewards shorter calls directly, while per-attempt fees remain even when no conversation begins. Preserve raw units and the converted charge so a later contract change does not corrupt historical comparisons.
Record the final state and the reason behind it. Useful states include resolved, safely transferred, caller abandoned, no answer, provider failure, tool failure, policy refusal, and unresolved completion. A retry should point back to the original task so three attempts do not masquerade as three unrelated customers.
Inspect the distribution, not just the mean. Median cost shows the ordinary call, while the upper tail exposes long conversations, repeated tool failures, looping prompts, and transfers that took too long to connect. Track cost by successful outcome alongside failure rate, transfer rate, and repeat-call rate.
Reconcile the ledger against every provider invoice before trusting it. Missing request IDs, time-zone boundaries, taxes, credits, minimum charges, and delayed usage records can all create gaps. Mark unreconciled spend explicitly instead of spreading it across successful calls and pretending the allocation is precise.
Compare cohorts using the same outcome definition and accounting window. That separation prevents a healthy appointment agent from hiding an expensive support flow, while showing whether the next fix belongs in call logic, tool reliability, retry policy, telephony, or the voice stack itself.
Start with a call-level ledger rather than a monthly invoice. Give every conversation one identifier, then attach its telephony duration, transcription minutes, language-model tokens, synthesized characters, platform time, tool calls, retries, and transfer charges. That baseline matters because the hidden price of an AI voice stack grows across services that meter different units.
A completed call is not always a successful call
A disconnected call is finished, but it may not be resolved. Define success from the business event your system was meant to produce, such as a confirmed booking, collected payment, verified address, completed support action, or safe transfer. Do not trust a polite goodbye or the model's own claim that the task succeeded.Containment needs the same discipline. An agent can avoid a human transfer by giving the wrong answer, losing the caller, or ending early. Count that as failure, not cheap automation, and keep unsafe or policy-required transfers in their own valid-outcome category.
Your useful metric is cost per correctly resolved call. Add every billable cost generated by a cohort, including unsuccessful attempts, then divide that total by the number of verified successful outcomes. The denominator must exclude calls that merely connected, reached voicemail, or produced an incomplete backend action.
Segment the calculation by intent before averaging anything. Appointment confirmation, collections, sales qualification, and technical support have different talk times, tool use, transfer policies, and acceptable outcomes. One blended number lets an efficient workflow hide a costly one.
Every provider bills a different slice of the call
Telephony may be prorated by active seconds, while another service meters audio minutes, text characters, tokens, or complete attempts. Some platforms also impose a minimum charge on failed outbound calls. Convert each record into actual dollars at the rate that applied when the event occurred.TTS cost follows what the agent says, not simply how long the caller stays connected. LLM cost follows the tokens processed across turns, including repeated context and tool results. A quiet or slow call can therefore carry a different component mix from a brisk call of the same duration.
Warm transfers deserve separate records. Depending on the platform, a proxy agent can accrue talk-time charges while it calls the human, followed by a distinct transfer rate after the legs merge. Some services also bill voicemail as normal call time, so a campaign with poor answer rates can look inexpensive per connected call while wasting money across the attempt pool.
Fixed fees belong in the ledger too. Allocate phone-number rental, platform subscriptions, reserved capacity, recording storage, monitoring, and support across the calls that used them. Keep engineering labor outside the vendor subtotal, then add it as a separate operational layer when manual review or correction is required.
Billing granularity changes the answer. Exact-second pricing rewards shorter calls directly, while per-attempt fees remain even when no conversation begins. Preserve raw units and the converted charge so a later contract change does not corrupt historical comparisons.
Outcome-level records expose the expensive failures
Join provider records with the application's outcome log, not with a spreadsheet assembled at month-end. The call identifier should follow the session through the carrier, speech services, model, orchestrator, and business tools. When a vendor cannot accept your identifier, store its returned request ID beside yours.Record the final state and the reason behind it. Useful states include resolved, safely transferred, caller abandoned, no answer, provider failure, tool failure, policy refusal, and unresolved completion. A retry should point back to the original task so three attempts do not masquerade as three unrelated customers.
Inspect the distribution, not just the mean. Median cost shows the ordinary call, while the upper tail exposes long conversations, repeated tool failures, looping prompts, and transfers that took too long to connect. Track cost by successful outcome alongside failure rate, transfer rate, and repeat-call rate.
Reconcile the ledger against every provider invoice before trusting it. Missing request IDs, time-zone boundaries, taxes, credits, minimum charges, and delayed usage records can all create gaps. Mark unreconciled spend explicitly instead of spreading it across successful calls and pretending the allocation is precise.
Compare cohorts using the same outcome definition and accounting window. That separation prevents a healthy appointment agent from hiding an expensive support flow, while showing whether the next fix belongs in call logic, tool reliability, retry policy, telephony, or the voice stack itself.
