Research-backed calculator guide

Cost per Successful Voice Call: Replace Cost per Minute With an Outcome Denominator

Learn how to calculate AI voice agent cost per successful call by combining all-in call spend, escalation cost, and a verified resolution denominator.

Why cost per minute is an incomplete answer

A per-minute rate answers, “What does this platform meter while a call is active?” It does not answer, “What did it cost to finish one customer job?” That distinction matters when the agent spends six minutes failing, transfers to a human, or resolves the request but causes the caller to ring back the next day.

Start with four mutually visible events in your ledger:

1. Connected attempt: the call reached the agent and created billable or operational work.
2. Contained attempt: the call ended without a human transfer. This is an operational outcome, not proof of success.
3. Verified resolution: the caller’s defined intent was completed, with no disqualifying repeat contact inside your chosen window.
4. Escalation or repeat: a human, second call, refund, rework step, or other recovery path was needed.

The distinction prevents a common accounting error. If 1,000 calls connect and 700 end without a transfer, it is tempting to use 700 as the denominator. If 80 of those callers return about the same issue, only 620 may be successful under a seven-day repeat-contact rule. Using 700 would understate cost by treating silence at the end of a call as a completed task.

The denominator also makes quality economically visible. An agent with a lower rate per minute can be more expensive per completed task if it talks longer, repeats prompts, or sends more cases to people. Retell’s published worked comparison similarly says the cost of a resolved outcome should include the machine attempt and the human work created by failure.^[1]

Define “successful” before calculating

Write the success rule before choosing a vendor or tuning a prompt. For a support line, a defensible rule might require all of the following: the requested action was completed, the relevant system returned a successful result, the call did not require a human handoff, and the caller did not reopen the same issue within seven days. For lead qualification, success may instead be a verified qualification record, not a pleasant conversation.

Keep containment rate and resolution rate as separate columns. Containment is useful for capacity planning because it predicts how many calls reach a human queue. Resolution is the economic denominator because it tracks whether the customer job was completed. Report both rates by intent, language, and inbound or outbound direction. A blended rate can hide an expensive failure mode in a high-volume intent.

Use the same rule when comparing human-only and AI-first handling. If the AI requires a repeat call but the human queue is judged only at hang-up, the comparison is biased. The rule also needs an observation window. A seven-day window is a practical example, not a universal standard; billing disputes, appointments, and sales leads may need different windows.

Build the numerator from actual meters

The numerator begins with every connected attempt, not only successful calls. Include the metered voice-agent minutes and any separate speech, model, telephony, recording, storage, analytics, number, concurrency, or platform charges that belong to the call period.

Vendor quotes must be dated. On the Retell pricing page checked on November 2, 2026, the pay-as-you-go page displayed a broad AI voice-agent range of $0.07–$0.31 per minute. Its detailed example displayed $0.04 per minute for GPT 5.1, $0.055 per minute for voice infrastructure, $0.015 for platform TTS, and $0 for telephony in that selected configuration, while the detailed table lists telephony choices and add-ons separately.^[2] Treat those figures as a dated vendor quote, not as a universal market rate.

Telephony can be a separate bill. Twilio’s US pricing page, marked current as of August 2026 when checked, lists pay-as-you-go local outbound calls at $0.0140 per minute, local inbound at $0.0085 per minute, and number rental at $1.15 per month. It also lists feature charges such as recording and transcription.^[3] Your stack may bundle those lines or meter them independently, so map each invoice row to the call ID before using a blended rate.

Speech components can also use different units. Deepgram’s pricing page checked on November 2, 2026 lists streaming STT per minute, TTS per 1,000 characters, and Voice Agent connection time per minute. Its standard Voice Agent pay-as-you-go rate is $0.075 per minute, while BYO combinations show lower rates but move some components into your own ledger.^[4] Do not add a character rate to a minute rate without measuring generated characters.

Fixed charges should be allocated deliberately. A simple monthly allocation is monthly fixed voice costs ÷ connected attempts, but that is only meaningful when the period has enough volume. Number rental and reserved concurrency may be capacity costs rather than marginal call costs. Show both marginal cost per attempt and fully allocated cost per successful call so finance can see whether a change affects unit economics or only capacity.

Worked cost example

Assume a US inbound support queue with 1,000 connected calls in one month. These are illustrative planning inputs, not a vendor quote:

• Average billed AI and telephony time per attempt: 4.0 minutes.
• All-in variable voice-agent rate: $0.11 per minute.
• Verified resolution rate: 68%.
• Loaded human escalation cost: $6.00 per failed call.
• Monthly fixed number, monitoring, and concurrency allocation: $180.
• Repeat-contact rule: same intent reopened within seven days removes the original call from “successful.”

The AI variable cost is 1,000 × 4.0 × $0.11 = $440. At 68% resolution, 680 calls are successful and 320 create expected escalation work. Human recovery cost is 320 × $6.00 = $1,920. Add the $180 fixed allocation and total period cost is $2,540. Divide by 680 verified successes: $3.74 per successful call, rounded to cents.

The same operation looks very different under weaker quality. Hold all inputs except resolution at 55%. The expected successful calls fall to 550, escalations rise to 450, and total cost becomes $3,320. Cost per successful call becomes $6.04. A $0.11 per-minute headline did not change; the denominator did.

A useful sensitivity check is to hold the 68% resolution rate constant and reduce the variable rate from $0.11 to $0.09. Variable cost falls by $80, so cost per success falls from $3.74 to about $3.62. By contrast, raising resolution from 68% to 75% while keeping the $0.11 rate lowers expected escalation cost to $1,500 and produces 750 successes; the result is about $2.83 per success. The exact result depends on whether fixed allocation and human handling time truly remain constant, but the direction shows why quality measurement deserves first-class cost data.

A practical measurement workflow

First, create one row per connected call with a stable call ID. Store billed seconds or minutes, model and voice, telephony direction, transfer flag, post-call outcome, and any retry or repeat-contact ID. Do not infer minutes from a rounded dashboard if the invoice is second-level.

Second, label outcomes after a consistent review window. Use automated post-call extraction as a candidate label, then audit a sample against recordings, transcripts, CRM actions, and repeat contacts. A vendor’s “successful” field can help operationally, but it is not automatically your finance definition.

Third, attach human work. A transfer can create a new telephony leg, queue time, handle time, after-call work, and sometimes a supervisor or specialist action. Use the loaded cost of that actual path. Do not apply one human cost to every failed call if intents have materially different handle times.

Fourth, reconcile the result to invoices. The AI API invoice reconciliation calculator can help match usage rows to the period total. For model and voice changes, use the Model pricing comparison page to keep units and tiers distinct. For a broader business case, the AI observability ROI calculator is useful when evaluation and monitoring costs are part of the deployment.

Finally, segment before optimizing. Report successful-call cost by intent, channel, call direction, language, model, and transfer reason. A single monthly average can improve while a valuable billing intent gets worse. Track connected, contained, resolved, transferred, repeated, and abandoned counts together.

What to exclude or qualify

Do not present a $3.74 example as an industry benchmark. It is a transparent scenario built from stated assumptions. Replace the $0.11 rate with your invoice-derived rate, replace $6.00 with your loaded recovery cost, and use an outcome label your operations team can reproduce.

Exclude sales tax unless it is part of the business cost you intend to compare. Exclude one-time implementation work from marginal call cost, but show it separately for payback analysis. Include it in a fully loaded first-year cost if you are deciding whether to launch. Keep vendor promotional credits, free minutes, and temporary discounts separate from steady-state rates.

There are also measurement limitations. Resolution may be delayed or partly inferred. Repeat contacts can be hard to join across phone numbers or channels. A transferred call can be successful after human help but not an AI-only success. State whether your metric is “AI-resolved success” or “AI-assisted successful outcome”; both can be useful, but they answer different questions.

FAQ

Is AI voice agent cost per successful call the same as cost per resolution?

Usually, yes, if “successful call” is defined as a verified resolution. The labels differ by team. Write the success rule, observation window, and treatment of transfers and repeats before calculating. If a successful call includes a human-assisted transfer, call it AI-assisted cost per successful outcome rather than AI-only resolution cost.

What resolution rate should I use?

Use the rate measured on your own call mix, not a generic vendor benchmark. Calculate it as verified successes divided by connected attempts, and report it by intent. Apply a repeat-contact window that matches the work, such as seven days for many support issues. If the label is immature, publish a range and show sensitivity rather than a false point estimate.

Does a contained call count as a successful call?

Not automatically. Containment means a human did not take the call. Success means the caller’s intent was completed under your rule. A contained call that triggers a repeat contact is contained but not successful. Keep both metrics visible because containment predicts human queue load while resolution determines outcome economics.

Should transferred calls be included in the numerator?

Include their AI and telephony cost in the numerator because those costs were incurred. Whether the transfer contributes to the successful denominator depends on the metric. For AI-only resolution, it does not. For AI-assisted successful outcome, it can, but you must add the human leg and label the metric clearly.

How often should I recalculate the metric?

Recalculate monthly at minimum and after a model, prompt, telephony, voice, routing, or transfer-policy change. Re-run it by intent whenever call mix moves materially. A stable per-minute quote does not guarantee a stable outcome denominator; resolution and repeat-contact rates can drift independently.

Which rate should I take from a vendor pricing page?

Take the rate that matches your selected model, voice, telephony path, billing unit, plan, and date. Record whether the page quotes a bundle, an all-in platform rate, or a component. Then reconcile it to your invoice. The dated Retell, Twilio, and Deepgram pages illustrate why a voice stack may have several valid rates rather than one universal number.^{[2] [3] [4]}