Most voice AI evaluations start with cost per minute. That framing almost guarantees the wrong conclusion. Here is a 4-step framework to model the actual economics, plus an interactive ROI calculator to run your own numbers.
Most voice AI evaluations start in the wrong place. They start with cost per minute. That framing almost guarantees the wrong conclusion, because cost per minute is not the economic driver in a call-heavy business. Revenue capture is.
If inbound calls represent high-intent demand at your business, then your telephony stack is a revenue conversion layer. The correct question is not "How much does voice AI cost?" The correct question is "How much revenue is currently leaking through the intake layer?"
To answer that, you need to model the economics of a call from first principles.
Here is a 4 step process to model that.
Related articles
You cannot meaningfully evaluate automation without knowing what a call is worth. And every inbound call sits inside a funnel.
Some callers are qualified. A portion of qualified callers convert. Each conversion generates gross profit. When you collapse that funnel into a single metric, you get gross profit per call.
It is a function of 4 variables:
If you do the math, the equation would be:
It is simply your sales math expressed at the call level.
Consider a mid-market operator handling 400,000 inbound calls per year. That is roughly 1,100 calls per day, which is common for regional insurance groups, healthcare networks, and home services aggregators.
Assume:
Gross profit per call is therefore:
0.5 × 0.4 × 400 = $80
At 400,000 calls per year, that translates to:
400,000 × 80 = $32,000,000 in annual gross profit flowing through the phone channel.
This is the number most companies never calculate. They track call volume and handle time, but not economic yield per call. Without that baseline, automation decisions default to cost containment instead of revenue capture.
No contact center captures 100 percent of high-intent demand. Leakage appears in predictable places like:
Abandonment rate is usually the only metric reported, but it understates economic loss. Some callers remain on the line yet convert at lower rates due to delay. From a financial perspective, that is still leakage.
For modeling purposes, assume effective leakage of 12 percent. This includes hard abandonment and soft decay in conversion caused by friction.
At $32 million in gross profit flowing through the channel, 12 percent leakage represents:
$32 million × 0.12 = $3.84 million in unrealized gross profit. Which can be attributed to intake inefficiency.
The range is realistic. Many mid-market contact centers operate between 8 and 15 percent abandonment, with materially higher rates outside business hours.
Voice AI agents do not need to eliminate leakage to be economically viable.
It only needs to reduce it incrementally.
Suppose voice automation reduces effective leakage from 12 percent to 9 percent. That is a 3 percent absolute improvement.
On the same call volume the recovered gross profit becomes:
Three percent sounds small. At scale, it is not.
Operationally, a 3 percent reduction can come from answering overflow instead of forcing callers into queue, providing coverage outside business hours, qualifying intent before routing, or eliminating scheduling bottlenecks. None of these require replacing a contact center. They require tightening the conversion layer.
Latency also matters here. If queue time increases abandonment or reduces close rate, response time becomes a conversion variable.
Most teams evaluate voice AI by comparing it to agent salaries. That misses the point entirely. The real question is how much revenue is leaking through your intake layer, and what fraction of that leakage can automation recover. When you model it that way, even small improvements can justify significant investment. David Casem, Chief Product Officer @ Telnyx
Cost must be evaluated against recovered gross profit, not against salary.
Assume the conversational system costs $0.08 per minute and the average interaction lasts three minutes. Variable cost per handled call is
0.08 × 3 = $0.24
If the system handles 200,000 calls per year, variable cost is:
Even after adding integration, monitoring, and operational overhead, a conservative annual cost of $100,000 is reasonable for many deployments running on a carrier-grade network.
Against $960,000 in recovered gross profit, the return multiple approaches 10x. You do not need unrealistic assumptions to justify the economics.
| Metric | Conservative | Base Case | Optimistic |
|---|---|---|---|
| Gross profit per call | $50 | $80 | $120 |
| Baseline leakage | 8% | 12% | 15% |
| Leakage reduction | 2% | 3% | 5% |
| Annual call volume | 400,000 | 400,000 | 400,000 |
| Recovered gross profit | $400,000 | $960,000 | $2,400,000 |
| Voice AI annual cost | $100,000 | $100,000 | $100,000 |
| ROI multiple | 4x | ~10x | 24x |
A model is only useful if it survives stress testing.
If gross profit per call drops to $50, leakage is 8 percent, and automation reduces leakage by 2 percent absolute, recovered gross profit becomes:
Against $100,000 in cost, that is still a 4x return.
Now test the other direction.
If gross profit per call is $120, leakage is 15 percent, and automation reduces leakage by 5 percent absolute, recovered gross profit becomes:
The upside expands quickly in high-margin verticals.
Where does this break down?
The model weakens when:
In those cases, incremental capture shrinks, and voice AI becomes a cost optimization tool rather than a revenue expansion lever.
The key is understanding which business you are running.
When teams focus exclusively on cost per minute, they implicitly assume calls are a liability.
In high-intent businesses, calls are demand. Marketing spend drives that demand. If 10 to 15 percent of inbound intent decays at the intake layer, the constraint is not acquisition. It is throughput and capture.
Voice AI is therefore not primarily a labor substitution tool. It is a mechanism for protecting and expanding the economic value already present in inbound traffic.
Many operators cannot answer three basic questions with confidence:
Without those inputs, automation is evaluated in isolation. With them, even small absolute improvements can justify material investment.
The math itself is straightforward. The discipline to model it consistently is less common.
That discipline, more than any model choice or vendor comparison, determines whether voice AI looks like an experiment or a scalable economic lever.
