"AI voice agents will replace BPOs" is a great headline and a bad business case. The truth is more nuanced: AI voice works on specific call types at sufficient volume, and on those the unit economics either work cleanly or they don't. Here is a model for assessing whether a given call workflow is a candidate, and the inputs you need to gather before anyone writes code.
The four call attributes that determine fit
- Repeatability. Does largely the same conversation happen thousands of times a month? AI does well. Is every call unique? AI struggles.
- Length and complexity. Short, structured calls with few branches suit an agent. Long calls with many turns and diagnostic back-and-forth raise error rates and cost, and human agents are often still cheaper end to end.
- Outcome clarity. Calls with a clear outcome — paid or didn't pay, booked or didn't book — are easy to track and price. Open-ended consultative calls are not.
- Tolerance for occasional escalation. If the cost of a wrong answer is high, you need a human in the loop. That changes the maths but doesn't kill it.
Good early candidates include appointment confirmations and reminders, early-stage payment reminders, order and delivery status, lead qualification against a short script, and after-hours reception.
The cost model
Compare cost per successful outcome, not cost per minute or per call attempted. On the human side:
human cost per outcome = fully loaded cost per agent hour ÷ successful outcomes per agent hour
"Fully loaded" means salary and benefits plus supervision, QA, training, attrition and replacement, seats, telephony and software. On the AI side:
AI cost per outcome = (variable cost per call × calls + escalation cost + fixed monthly cost) ÷ successful outcomes
Variable cost covers telephony minutes, speech recognition, model inference and speech synthesis. Escalation cost is the human time spent on calls the agent hands over. Fixed cost covers platform fees, monitoring and the people who review and tune the agent.
A worked example (hypothetical numbers)
Say a lender makes 60,000 payment-reminder calls a month, and suppose:
- Variable AI cost is ₹6 per call: ₹3.6 lakh.
- 20% of calls escalate to a human at ₹30 each: 12,000 × ₹30 = ₹3.6 lakh.
- Fixed costs — platform, monitoring, one reviewer — are ₹2 lakh.
- 30% of calls reach the right person and end in a payment commitment: 18,000 outcomes.
That is ₹9.2 lakh ÷ 18,000, or about ₹51 per outcome. If your human team produces the same outcome for ₹80, the AI wins. If it does so for ₹40, it doesn't — and the next lever is raising the success rate or cutting escalations, not negotiating cheaper minutes. Swap in your real figures; the structure is what transfers.
Two effects usually decide the result. First, fixed costs dominate at low volume, so small programmes rarely pay back the integration effort. Second, an agent can place calls at the hours customers are most likely to answer, which can raise right-party contact — but measure that in a pilot rather than assuming it.
The hidden costs nobody mentions
Voice AI is not a "set and forget" product. It needs:
- An ongoing tuning function — someone reviewing call samples and transcripts every week, updating prompts and adding test cases.
- Telephony partnerships (Twilio, Plivo, Exotel, Gupshup and others), number management and the operational overhead of running them.
- A human escalation team for the calls the AI shouldn't close on its own.
- Compliance work: recording notices, consent, calling-hour rules and do-not-call lists. In India, TRAI's rules on commercial communications apply; in the US, the FCC has ruled that AI-generated voices count as "artificial" voices under the Telephone Consumer Protection Act, so its consent rules cover AI calls.
- Integration with your CRM, payment or booking systems, so the agent can act and not just talk.
Account for these or your projected ROI will be wildly optimistic.
Run a pilot before you commit
- Pick one call type and write down exactly what counts as a successful outcome.
- Split the call list randomly between AI and human agents and run both for a few weeks.
- Compare cost per outcome, complaint rate, escalation rate and compliance findings.
- Scale only the call types where the AI wins on cost without losing on quality.
What it can't do well
Genuine sales conversations that depend on building rapport over a long call. Complex troubleshooting that depends on diagnostic exploration. Calls where the customer is in distress and needs empathy that doesn't come from a model. Don't pretend otherwise — set the AI on the calls it is good at and leave the rest to people. Where the agent needs to take actions across several systems, the design questions overlap with agentic systems more broadly.
FAQ
Should the agent disclose that it is an AI?
Yes. Beyond any legal requirement in your market, customers who find out later tend to trust the whole programme less.
What volume makes this worthwhile?
There is no universal number. Run the formula above with your own fixed costs; the break-even is the volume at which AI cost per outcome drops below your human cost per outcome.
How Velura Labs approaches this
Our AI Voice Call Center is billed per resolved call, so you pay for calls that closed, and we start with a unit-economics review so the maths is clear before anything is configured. For broader context, see our agentic AI vs RPA piece on the related back-office shift. Talk to us if you want a real ROI model for AI voice on your specific call types.
Our clients for this span US tech hubs (San Francisco, Seattle, Austin, New York), European markets (Paris, Milan, Rome), the Middle East (Dubai, Riyadh, Abu Dhabi) and India. Start a conversation from anywhere.