AI voice automation ROI is what an AI agent saves and recovers on phone calls, minus what it costs to run. Savings are staff time no longer spent on routine calls. Recovered revenue is enquiries you would have missed after hours or during overflow. Running cost is now mostly per-minute usage, roughly $0.07 to $0.31 a minute on the platforms I checked in October 2026.
Updated October 2026: replaced the old cost table with per-minute pricing read from vendor pricing pages on 6 October 2026, and removed figures I could not source. This is a calculation guide. If you are choosing between vendors, how to choose an AI voice agent covers that.
The ROI Formula
Monthly net benefit = labour saved + missed-call revenue recovered − monthly running cost
Payback period (months) = one-time setup cost ÷ monthly net benefit
Three inputs go in. Two of them you can measure from your own phone logs. The third, how much missed revenue the agent actually recovers, is an assumption, and as the example below shows, it is the one that decides the result.
What the Running Cost Looks Like in 2026
Self-serve voice agent platforms bill per minute of call time, and the total is the platform fee plus speech-to-text, the language model, the voice and telephony. These are the figures each vendor listed on its own pricing page on 6 October 2026, in US dollars:
| Platform | What the pricing page lists | All-in per minute |
|---|---|---|
| Vapi | $0.05/min hosting, plus speech-to-text ($0.0095 to $0.0099), language model ($0.0077 to $0.0452) and voice ($0.0146 to $0.0238) | $0.082 to $0.129, from its calculator at 1,000 minutes a month |
| Retell AI | $0.055/min voice infrastructure, $0.015/min telephony (US, via Twilio), $0.015/min on its own voices, language model priced by model | $0.07 to $0.31, as listed |
Two cautions. Telephony outside the US costs something different, and the language model you pick moves the total more than any other line. These are the cheap end of the market: setup, knowledge-base work, integrations and ongoing tuning are separate, and a managed build is quoted per project. My own pricing breakdown is in AI voice agent pricing in 2026.
A Worked Example
Every input below is an assumption I chose to show the arithmetic. None of it is a measured result from a real business. Replace each line with your own number.
- Call volume: 400 inbound calls a month, 3 minutes each, so 1,200 minutes of agent time.
- Routine share: 60% of calls are routine (hours, order status, appointment confirmations) and the agent resolves them. That is 240 calls, or 12 hours of staff time.
- Labour saved: 12 hours at a loaded $25 an hour = $300 a month.
- Missed calls: 30 calls a month go unanswered outside hours. The agent answers them, 15% become booked jobs, and the average job is worth $350. 30 × 0.15 × $350 = $1,575 a month recovered.
- Running cost: usage of 1,200 minutes at the top of the $0.07 to $0.31 range is $372, plus an assumed $150 a month for management = $522. At the bottom of the range, usage is $84.
- Monthly net benefit: $300 + $1,575 − $522 = $1,353.
- Payback: an assumed $1,500 setup ÷ $1,353 = about 1.1 months.
Now look at where the benefit comes from. $1,575 of the $1,875 gross benefit, or 84%, is recovered revenue. If the agent recovers nothing, the labour saving of $300 is less than the $522 running cost, and the agent loses $222 a month. The headline payback figure is only as good as your estimate of missed-call conversion, so measure that first.
To run your own version, use the AI automation ROI calculator.
How to Measure Your Own Inputs in 20 Minutes
- Count monthly inbound calls and average call length from your phone system logs.
- Log one week of calls by topic. The routine share is whatever you find, not a number from a vendor.
- Price the labour. Hours spent on routine calls times the loaded hourly cost of the person answering.
- Count missed calls. Calls that went to voicemail or rang out, times your close rate on inbound enquiries, times average customer value.
- Get a per-minute quote and add setup and management as quoted.
- Run the formula twice: once with your estimated recovery rate, once with zero. If the zero case is badly negative, the case rests on recovery, so check that number again.
What Moves Payback Up or Down
| Driver | Pushes ROI up when | Pushes ROI down when |
|---|---|---|
| Call mix | Most calls repeat the same 10 to 20 questions | Most calls need judgment or a negotiation |
| Missed calls | Many enquiries arrive when nobody can answer | Your phone is already answered at all hours |
| Value per customer | One recovered enquiry covers weeks of running cost | Low-value transactions need very high volume |
| Language model and voice | A smaller model handles your scripts well | You need the most expensive model to hold quality |
| Tuning | Someone reviews real calls and fixes the knowledge base | It is set up once and never reviewed |
Where I Have Seen It Work
Le Marquier is a French plancha manufacturer. Voxdonna built it a production voice agent for its support line, answering in French, English and Italian with automatic language detection. Its case study reports an 80% reduction in annual support cost. That is my own case study and an unaudited claim, so treat it as one data point rather than a benchmark for your business. The build lessons are in building a French voice agent for Le Marquier, and the customer outcomes are in the full case study.
What Voice Agents Do Badly
- Emotional or high-stakes calls. A customer with a serious complaint needs a person. The agent should triage and hand off, not resolve.
- Unpredictable queries. If every call is different, the agent needs far more tuning before it saves anything.
- The first weeks. Expect a tuning period where you read real transcripts and fix gaps. Do not judge ROI on week one.
An AI readiness assessment will flag these before you commit. The related guides AI voice agent vs call center cost and how AI voice agents handle customer support go further on cost comparison and call handling. For the wider picture on agents, see the AI agents hub.
Questions to Ask Before You Commit
- How does the agent learn my products and policies? A generic agent will not hold up on your specific questions.
- What happens when it cannot answer? Warm transfer, callback or voicemail each give the caller a different experience.
- What does "handled" mean in the report? Make sure calls that ended in confusion are not counted as resolved.
- Who tunes it after launch, and is that in the fee?
- Can I see a real case study with the method behind the number? A vague "up to X%" claim tells you nothing.
The Bottom Line
Run the formula with your own call logs, and run it twice, once with your recovery estimate and once with zero. If it only works with recovery, you now know which number to test before spending anything. If it works either way, the case is solid.
Frequently Asked Questions
What is AI voice automation ROI?
AI voice automation ROI is the labour cost an AI agent saves on routine calls plus the revenue it recovers from calls you would otherwise miss, minus what the agent costs to run. The costs are mostly per-minute usage now, roughly $0.07 to $0.31 a minute across the platforms checked in October 2026.
How quickly do small businesses see ROI from an AI voice agent?
It depends on one input: whether the agent recovers calls you would have missed. In the worked example on this page, payback is about 1.1 months, but 84% of the benefit is recovered revenue. With no recovered revenue, the labour saving alone does not cover the running cost.
What does an AI voice agent cost per month?
On self-serve platforms the usage cost is per minute. Vapi's calculator shows about $0.082 to $0.129 a minute all-in for 1,000 minutes, and Retell lists $0.07 to $0.31 a minute depending on the language model and voice. Setup and ongoing management are separate quotes.
How do I work out what missed calls cost me?
Pull a month of call logs. Count calls that went unanswered or to voicemail, multiply by your close rate on inbound enquiries, then by average customer value. That is your missed-call revenue. Use your own numbers, not an industry average.
Can an AI voice agent handle most of my calls?
That depends on your call mix, so log one week of calls by topic before you trust any handling-rate figure, including vendor figures. Le Marquier's case study reports an 80% cut in annual support cost after Voxdonna built its voice agent. That is the author's own unaudited claim.
Want Help Running the Numbers?
On a free 30-minute call I will go through your call logs with you, build the model above with your inputs, and tell you honestly whether an AI voice agent makes sense right now.
