Show your sources
Why these defaults
The preset describes an inbound service line. Each number below is either taken from a public source, calculated from public list prices, or marked as an assumption you should replace.
- Monthly call volume10,000 callsAssumption
- A mid sized inbound service line. Replace it with the count from your carrier call detail records.
- Human cost per call$8Sourced
- SQM Group reports an average cost per call of $8.60 across more than 300 North American call centers. The preset rounds down to stay conservative.
- Handle time7 minutesSourced
- The same SQM Group study reports an average handle time of 424 seconds, talk plus wrap up. It is shown for context and does not change the math.
- Automation (containment)35%Assumption
- Share of calls the agent finishes without a transfer. Set deliberately below typical chat figures because callers ask for a person more readily.
- Human review10%Assumption
- Contained calls that still need a callback or a quality listen.
- Retry rate10%Assumption
- Extra model turns from barge in, misheard intents and repeated prompts.
- Tokens per call12,000 in, 600 outAssumption
- About eight turns, each resending the system prompt and the growing transcript to the model.
- ModelGPT-5 miniSourced
- A small, fast model, because latency decides containment on a call. Rates come from the site model table, which cites the OpenAI pricing page.
- Other cost per call$0.12Calculated
- 4 agent minutes per call times about $0.0297 per minute: $0.0085 for a US local inbound minute on Twilio, $0.0077 list rate for Nova-3 streaming speech to text on Deepgram, and $0.0135 for text to speech at Deepgram Aura-2's $0.030 per 1,000 characters. The 4 minute call and 450 characters of agent speech per minute are assumptions.
- Build and maintenance$90,000 build, $6,000 per monthAssumption
- Telephony integration, evals on recorded calls, latency tuning and monitoring. Replace with a scoped estimate.
- Discount, ramp, decay12% per year, 3 months, 0%Assumption
- Kept equal to the general calculator so results stay comparable.
About this tool
What this AI voice agent ROI calculator answers
This AI voice agent ROI calculator estimates payback, monthly net savings and 36 month NPV for an agent that answers or places phone calls, with telephony, speech recognition, speech synthesis and model tokens priced per call. Use it before you sign a voice platform contract or staff a build, so the cost per minute sits next to the labour it replaces.
Voice changes the economics compared with chat. Every second of a call is metered by the carrier and by the speech providers whether the caller is helped or not, so a call the agent fails to contain still costs money before it reaches a person. The preset loads an inbound service line with cautious containment, and every number stays editable once you have your own recordings and carrier invoices.
How to use it
Work through the inputs in the order a finance reviewer would challenge them.
- Enter call volume. Enter monthly inbound or outbound call volume from your contact center platform or carrier call detail records.
- Set the human cost per call. Replace the cost per call with fully loaded agent cost divided by handled calls, including wrap up time.
- Set containment and review. Set automation to calls the agent finishes without a transfer, and review to the share that still needs a callback.
- Price each call. Multiply expected agent minutes per call by carrier, speech to text and text to speech rates, then enter the total.
- Size tokens and model. Estimate tokens from a typical call transcript, remembering that each turn resends the growing conversation to the model.
- Stress test and share. Read payback and the sensitivity table, then share the scenario link with the operations lead who owns the phone queue.
How the per call cost is built
The calculator keeps two cost lines. Monthly LLM cost is input and output tokens per call, priced from the model table, multiplied by call volume and by one plus the retry rate. Monthly other cost is the per call figure you enter multiplied by call volume. Labour saved is call volume times containment, reduced by the review share, times the human cost per call. Net savings subtract both cost lines and maintenance, and NPV discounts each month across three years.
Retries on a call come from barge in, misheard intents and repeated prompts. They inflate tokens but not minutes, which is why minutes live in their own field. Other cost is charged on every call the agent touches, contained or not, because the carrier and speech meters run either way.
Where voice agent ROI usually breaks down
The common failure is latency, not price. A caller who waits two seconds after every sentence hangs up or asks for a person, and containment collapses even when the answers are correct. Smaller, faster models often win on the phone for that reason, the opposite of the instinct to buy the largest model available.
The second failure is transfer leakage. A call that spends three minutes with the agent and then moves to a human carries the speech bill and the full human cost, so weak containment can cost more than no agent at all. The third is audio coverage: recognition accuracy measured on clean recordings rarely survives speakerphones, cars and heavy accents.
When the voice agent numbers hold up
The numbers hold for busy lines with a narrow set of intents: appointment booking and rescheduling, order and delivery status, payment reminders, account unlock flows and outbound confirmation calls. These have short scripts, clear success signals and an easy route to a person.
They hold poorly for complaints, retention offers, collections disputes and calls with legal disclosure scripts that change by state or country. Quiet lines struggle too, because the fixed build and maintenance cost never spreads across enough calls to pay back.
How this differs from voice platform calculators
Voice platforms usually quote one blended price per minute, which hides how much of it is carrier, speech or model margin. This page splits the bill so you can test swapping one provider at a time. The preset uses public list prices from the Twilio US voice pricing page, the Deepgram pricing page for speech to text and text to speech, and call benchmarks from SQM Group. Nothing here is a quote. Check each page again before budgeting, because speech pricing moves often and promotional rates come and go.
Pressure test your voice agent business case
Bring the scenario link to a call and walk through latency, transfer design and per minute spend before you commit to a platform or a build.
Book an architecture reviewFrequently asked questions
- What does an AI voice agent cost per minute?
- Most of the cost sits outside the model. At public list prices, carrier minutes, streaming speech to text and text to speech add up to roughly three cents per minute of agent talk time, while the LLM tokens for a short call often cost well under one cent. Platforms that bundle these usually charge more per minute in exchange for less engineering work. Enter your own blended figure in the other cost field.
- How is AI voice agent ROI calculated?
- ROI compares the labour saved with everything the agent costs to run. Labour saved is call volume times containment, reduced by the share of calls that still need human follow up, times the human cost per call. The calculator subtracts token cost, per call minutes and speech fees, and monthly maintenance, then divides build cost by the result to get payback in months.
- What containment rate should I assume for a voice agent?
- Start lower than you would for chat. The preset uses 35% as an assumption for an inbound service line with a handful of common intents. Callers interrupt, mumble and ask for a person more readily than chat users do, so measure containment on a pilot of real calls before you put a higher figure into a business case.
- Should transferred calls count as savings?
- Usually not. A call the agent handles for two minutes before transferring still incurs the full human handle time, and it adds carrier and speech cost on top. That is why the calculator charges the per call cost on every call while savings only come from calls the agent finishes. If transfers carry a summary that shortens human handling, lower the human cost per call to reflect it.
- Does this calculator work for outbound calling campaigns?
- Yes, with two adjustments. Outbound minutes are priced higher than inbound on most carriers, so recalculate the other cost per call with the outbound rate. Many outbound attempts end in voicemail or no answer, so count task volume as completed conversations and fold the cost of unanswered attempts into the per call figure. Check consent and calling rules for your market first.
- Can I share a voice agent scenario with my team?
- Yes. The share button writes every input, including the per call cost, into the link after the hash sign and copies it to your clipboard. Anyone who opens the link sees the same numbers in their own browser. Nothing is sent to a server, there is no account to create, and browsers never include the hash in requests to the site.
Related services and reading
From a scenario to a scoped build.