AI voice agent ROI: an honest calculation for inbound calls
Writing freed agent hours down as revenue is the most common calculation mistake in AI projects. This guide covers four input figures, calculating three effects separately — answered calls, agent time and the queue — an illustrative calculation, the cost side, the shape of the ROI table, and checking the figures against a baseline.
The short answer
The ROI (return on investment) of an AI voice agent is built from three separate effects, and they must not be mixed: calls that used to go unanswered now get answered, agent time is freed, and those two changes affect a real business outcome — sales, bookings, retained customers. Only the third brings money; the first two are the road to it.
The most common mistake is writing freed agent hours straight down as revenue. Time saved becomes money only if labour cost actually falls or the time goes into other measurable work. This article builds a calculation model that keeps that distinction, step by step.
Four input figures
Before calculating, take these four figures from your own systems. If you have to estimate, mark it clearly in the table.
- Inbound call volumeCalls received per month, split into business hours and out of hours.
- Answer rateThe share of those calls that become a conversation with a person; the rest are busy lines, callers hanging up in the queue, or out-of-hours calls.
- Split by call typeThe share of calls the agent can fully answer, can collect details for and hand over, and will not touch.
- Outcome rateThe share of answered calls ending in a booking, order or sales interest, and its average value.
First effect: answered calls
An agent can answer several calls at once and does not depend on business hours. So the first effect is that part of the previously unanswered calls get answered. Be careful here: not every unanswered call is a potential customer. Repeat calls, wrong numbers and spam must be removed separately.
The money value of this effect is calculated with the outcome rate of the extra answered calls. If 20% of calls answered by agents used to end in a booking, do not automatically apply the same rate to the extra calls the AI answers — a caller at night may have a different intent from one during the day.
Second effect: agent time
If the AI answers repeated questions before they reach an agent, agent time is freed. That time is easy to count in hours: the number of those calls multiplied by the average talk time plus after-call work.
Third effect: queue and waiting
When simple questions stay with the AI, the agent queue shortens, waiting time drops and fewer callers may hang up while waiting. This matters for customer experience, but it is hard to put a direct money value on it. Show it as a separate line in the ROI table, or add only a measured reduction in abandonment to the first effect.
Illustrative calculation
Not real customer figures. A company receives 6,000 inbound calls a month and answers 4,800. After removing repeat and wrong calls from the 1,200 unanswered, 700 unique requests remain. In the pilot the AI answers 500 of them. The booking rate on answered calls is 15%, and the average margin per booking is 40 AZN.
Agent time: the AI fully answers 1,800 information questions a month, each averaging 3 minutes — 90 hours. If those hours prevent a planned half-time hire, that salary is recorded as a saving. If not, the 90 hours stay in the table as "freed time", not as money.
The cost side
- Platform subscription and per-minute call cost
- Phone line or SIP trunk costs
- Internal time to write scenarios, prepare the knowledge base and test
- Time of the person who reviews calls and makes corrections during the pilot
- Integration work — writing fields to the CRM, API connections
Internal time in the first month is usually higher than later. Show it as a separate set-up cost rather than a monthly cost, so the payback period is calculated correctly.
The shape of the ROI table
- Line 1 — measured extra outcomeExtra bookings, orders or sales × average margin.
- Line 2 — realised cost savingOnly costs that actually fall or are avoided.
- Line 3 — freed time (hours)Shown in hours, not money.
- Line 4 — costsMonthly costs and a separate set-up cost.
- ResultROI = (Line 1 + Line 2 − monthly cost) ÷ monthly cost; payback = set-up cost ÷ monthly net benefit.
How to check the figures
The most reliable method is comparison: record the four weeks before the pilot as a baseline, then compare pilot weeks for the same call types. Note outside factors such as season, advertising campaigns and price changes. If possible, switch the agent on for one line only, or only at certain hours, so a comparison group remains.
Common mistakes
- Recording every freed hour as a salary saving
- Counting every unanswered call as a lost sale
- Applying the outcome rate of agent-answered calls to the AI's calls
- Forgetting the set-up cost or charging it to one month
- Taking a vendor's general statistics as your own figure
Limits
An ROI model rests on assumptions, and the result is only as reliable as they are. Some effects — brand impression, a customer getting an answer at night — do not turn easily into money, and there is no need to force them. This is not financial advice; for an investment decision, apply your own finance team's methodology.
Where the figures come from in Vexvon
In Vexvon AI Call Center every call's transcript, one-sentence summary and the fields you defined in the scenario — for example "outcome: booking completed" — stay in the panel and are written to the customer card. That lets you pull the count needed for line 1 without listening to calls one by one. The agent answers several calls at once (the parallel-call ceiling depends on the plan) and works outside business hours; difficult calls are passed to a live operator.
To value an unanswered call separately, see the cost of missed calls; to choose call types for the pilot, see the selection matrix.
First step
Collect the four input figures for last month and fill in lines 1–4 of the table. Next to every estimated figure, write "assumption". More articles are in the voice agent strategy section; to check the calculation together, get in touch.