The pitch is that an AI voice agent replaces your call centre. It does not, and suppliers who say it does create disappointed clients about four months in. What it replaces is the share of call volume that was never worth a trained human in the first place.
Lacewing Technologies is an AI and software development company in Navi Mumbai, and voice is the service line where expectations and reality diverge most often. So this post is mostly about where the comparison breaks rather than where it flatters.
What does an AI voice agent cost per minute in India?#
All-in, roughly ₹4 to ₹12 per minute. That covers three things: the telephony leg, speech recognition and synthesis, and the language model doing the reasoning.
The spread is wide for a reason. A short, scripted confirmation call with a small model sits near the bottom. A long, open-ended support call with a large model, high-quality Indian-accent voice synthesis and several tool calls to your systems sits at the top, and can exceed it.
| AI voice agent | Trained human agent | |
|---|---|---|
| Cost basis | Per minute of call | Per month, fully loaded |
| Typical India cost | ₹4 – ₹12 per minute | Salary plus overhead, seat, supervision |
| Availability | 24/7, no rota | Shift-bound |
| Concurrency | Effectively unlimited | One call at a time |
| Ramp time for a new process | Hours | Weeks |
| Handles exceptions | Poorly | Well |
| Cost when idle | Zero | Full |
That last row is the one people miss. A human agent costs the same whether the phone rings or not, which is why a call centre sized for peak is expensive during every other hour of the day. An AI agent costs nothing at 3am when nobody calls, and costs nothing extra when forty people call at once.
Why is cost per minute the wrong number?#
Because a call the AI cannot finish costs you twice. You pay the per-minute cost of the AI attempt, then you pay the human cost of the call anyway, and you have spent some of the customer's patience getting there.
The number that decides the business case is cost per resolved call. Work it out like this: take the per-minute cost, multiply by average handling time, then divide by the share of calls the agent actually finishes. An agent that costs ₹8 per minute and finishes four calls in five is cheap. The same agent finishing one call in three is not, and no amount of prompt tuning fixes a call type that needs judgement.
This is why we scope voice projects by call type rather than by call volume. Volume tells you the size of the prize. Call type tells you whether you can win it.
Which calls does AI actually finish?#
The pattern is consistent across the deployments we run. AI does well when the task is bounded, the information needed is in a system it can query, and success is unambiguous.
- Answering calls outside working hours, when the alternative is voicemail
- Appointment confirmation, rescheduling and cancellation
- Payment and EMI reminders
- First-touch qualification within a minute of a form fill
- High-volume repetitive outbound where the script barely varies
- Order status and delivery tracking
It does badly, and predictably badly, on:
- Negotiation of any kind, including discount requests
- Genuinely angry customers, where the goal is de-escalation rather than information
- Anything requiring judgement about an exception to policy
- Complex technical troubleshooting with branching diagnosis
- Calls where the customer's actual question is not the question they asked
That last one is the interesting failure. A human agent hears "is my order shipped" and works out from the tone that the real issue is a wedding on Saturday. An AI agent answers the question that was asked, accurately, and loses the customer.
What does the realistic deployment look like?#
AI takes first contact and everything after hours. Humans take escalation and anything commercial. The handover carries a written summary so the customer does not repeat themselves.
Three design details matter more than model choice:
State that it is an automated assistant in the first sentence. Complaints drop, and it removes an argument you do not want to have later.
Make the route to a human obvious and short. A trapped caller is a lost customer. The measure of a good agent is not how few calls it hands over, it is how quickly it hands over the ones it should.
Write the handover summary automatically. The commercial gain from voice AI is often not the deflected calls at all. It is that every escalated call arrives with the customer's history and intent already typed up, which cuts the human's handling time on the calls that actually matter.
How do you test a voice agent before buying one?#
Vendor demos are run on clean audio, in a quiet room, on the call types the agent handles best. Your calls are not that. Testing properly costs a few days and prevents most of the disappointment.
Take fifty recordings of real calls from your own lines, chosen to match your actual mix rather than your ideal one. Include the bad ones: the noisy shop floor, the caller on a moving train, the person who switches between Hindi and English mid-sentence, the one who interrupts.
Then measure three things separately, because a failure in any one of them looks identical to the customer but has a different fix.
- Transcription accuracy. Did the system hear the words correctly? This is where accent and line quality show up.
- Intent accuracy. Given a correct transcript, did it work out what the caller wanted? This is a prompt and design problem.
- Task completion. Did it finish the job, including the system calls behind it? This is usually an integration problem.
A vendor who cannot run this against your own recordings before contract is asking you to buy on faith. A vendor who can, and shows you the failures rather than the wins, is worth talking to further.
What are the compliance questions before going live?#
Voice raises issues that text does not, because you are recording a person.
Speech is personal data, and processing it puts you inside the Digital Personal Data Protection Act, 2023. Practically that means three things to settle before launch rather than after: what you tell the caller at the start about recording and automated processing, what lawful basis you are relying on, and how long recordings and transcripts are retained before deletion.
None of this is difficult, but it is a policy decision rather than an engineering one, and it needs an owner in the business. Where a project stalls after build, this is usually why.
For outbound calling there is a second layer around unsolicited commercial communication and consent registration. If your use case is outbound marketing rather than servicing existing customers, resolve that before you build anything.
The honest summary#
An AI voice agent is a strong buy when your call mix is dominated by short, repetitive, bounded calls, and a poor one when your value is in the conversation itself.
Ask a supplier which of your call types they expect the agent to finish, and what happens to the ones it does not. A supplier who answers "all of them" has not run one of these in production.