Most discussions about AI voice agents eventually come down to one question:
“Does it sound human?”
It’s an important question.
But after analyzing more than one million AI-assisted business calls across India, we found that voice quality is only one piece of the puzzle.
A voice agent can sound incredibly natural and still perform poorly if it responds too slowly, calls at the wrong time, struggles with Hinglish, uses a bad retry strategy, or doesn't have a clear objective.
The interesting part isn't whether AI can make a phone call.
It can.
The interesting part is what makes that call successful.
Our analysis of production traffic on the DialNexa platform revealed several patterns worth paying attention to.
1. The first attempt tells only half the story
New outbound numbers achieved a 48% first attempt pickup rate across the first 1,000 leads. Then things changed.
As numbers accumulated more calls, pickup rates declined, with some call categories falling toward 20%.
That could easily lead a team to conclude that the campaign or AI agent isn't working.
But retries changed the outcome.
With appropriately timed retry sequences, some campaigns achieved more than 70% cumulative connectivity.
This suggests an important design principle for voice agents:
Don't optimize only for the first call. Optimize for the entire contact cycle.
A retry isn't simply another dial. Its timing, frequency, and relationship to previous attempts can directly affect the probability of eventually reaching the person.
**Latency is a UX problem, not just an infrastructure problem **
The median response latency in the dataset was under one second.
The 95th percentile was approximately 2.1 seconds.
At first glance, that might not seem like a huge difference.
In a normal application, a couple of seconds might be tolerable.
In a voice conversation, it's different.
A long happens in the middle of a sentence. The caller doesn't know whether the system is thinking, the network failed, or the call has disconnected.
That uncertainty changes the experience.
The analysis identified several contributors to latency spikes, including uncommon phrasing, additional network hops, and peak concurrency.
This is why voice AI teams should monitor p95 latency, not just average or median latency.
The median tells you what usually happens.
The tail tells you where conversations break.
3. India makes code-switching a first class engineering problem
Consider a conversation like this:
“Haan, I wanted to know what's the pricing after the free trial?”
The caller isn't choosing a language.
They're just talking.
English, Hindi, and Hinglish appeared among the most consistent languages in the analyzed dataset.
The data also showed an important architectural difference. Speech to speech approaches handled mixed language conversations more naturally than some traditional cascade pipelines, where speech recognition, language processing, and speech generation happen as separate stages.
With a cascade, every handoff can introduce another opportunity for transcription errors, pronunciation problems, or lost context.
For Indian voice applications, “supports Hindi” isn't necessarily enough. You need to ask:
Does it understand how my users actually speak?
**4. Clear objectives beat complicated conversations **
The strongest use case in the dataset was pre-sales lead qualification.
Webinar and event attendance followed.
There's a reason these use cases work well.
They have a defined outcome.
A qualification agent needs to determine whether a lead is interested and whether they're ready for a sales conversation.
An event agent needs to confirm attendance, answer basic questions, and remind the registrant.
The objective can be expressed in one sentence.
That's an underrated principle for AI agents:
If you can't clearly define what success means, the agent probably isn't ready for deployment. Open-ended conversations aren't impossible.
They're simply harder.
They require better memory, better orchestration, stronger guardrails, more robust handling of unexpected responses, and much more careful evaluation.
Starting with a structured workflow gives teams a much cleaner way to measure whether the system is actually working.
5. People don't necessarily hang up because they're talking to AI
One of the more interesting findings was the under-3% user-initiated drop-off rate.
That doesn't mean AI conversations are automatically good.
It means that once people answered, most didn't immediately hang up.
Natural voices helped.
But conversation design mattered too.
A rigid opening, repetitive questions, or an awkward response can make even an excellent voice model feel artificial.
There's another interesting detail: when callers asked whether they were speaking with an AI, confirming it was an AI agent could build trust rather than destroy it.
That leads to a useful product principle:
Don't try to hide the system. Make the system useful.
**6. Inbound and outbound are completely different problems **
The dataset contained roughly 16% inbound and 84% outbound calls.
But the two categories behaved very differently.
Inbound calls averaged around 13 minutes, and approximately 89% completed their defined objective.
The reason is straightforward.
An inbound caller has already expressed intent.
They're not being interrupted. They're looking for something.
The AI's job is to answer quickly, understand the request, and help them reach an outcome.
Outbound calling has a much harder starting point: the person didn't ask to receive the call.
That means businesses shouldn't combine inbound and outbound performance into one metric.
They solve different problems and should be optimized differently.
For campaigns targeting working professionals, three windows produced the strongest connectivity: 10 AM–12 PM
4 PM–6 PM
8 PM–9 PM
These shouldn't be treated as universal rules.
A campaign targeting students, retirees, healthcare workers, or night shift employees could behave completely differently.
The broader engineering lesson is more useful:
Connectivity isn't evenly distributed throughout the day.
Instead of distributing call volume uniformly, teams should build their own connectivity curves and use those curves to schedule campaigns and retries.
What does this mean for AI engineers?
The biggest takeaway from one million calls isn't that AI voice agents are ready to replace human conversations.
It's more specific than that.
Voice agents perform best when the system around the model is designed properly.
That means:
Monitor p95 latency, not just averages.
Design retries as part of the initial workflow.
Test code-switching, not just individual languages.
Start with structured, measurable objectives.
Treat inbound and outbound as separate systems.
Design calling schedules around actual user behavior.
Optimize the first few seconds of the conversation.
Don't sacrifice transparency just to make the AI appear human.
The voice model is only one component.
The real system includes the telephony layer, speech recognition, language model, speech synthesis, orchestration logic, retry engine, scheduling, observability, and conversation design.
That's where the difference between a convincing demo and a production-grade voice agent starts to become obvious.
The bigger lesson
A million calls give you enough data to stop arguing about whether voice AI sounds human and start asking a more useful question:
Does the system reliably accomplish what it was designed to do?
Our data suggests that success depends on a combination of factors rather than one breakthrough model.
Fast responses.
Good conversation design.
Reliable language handling.
Smart retries.
Appropriate calling windows.
Clear objectives.
And a voice that people are comfortable talking to.
The voice gets the call started.
The system determines what happens next.
That may be the more important lesson for anyone building voice AI for India's next billion conversations.