One of India’s first voice AI deployments to carry 40% of all inbound support
How Kissht put voice AI on its most sensitive inbound queue — loan-rejection queries and mandate/NOC cancellation — and reached 40% of total inbound volume within two months, resolving 88% of those calls without a human at roughly 30% lower handle time.
Industry
Digital lending
Use cases
Loan rejection queries, mandate/NOC cancellation
Languages
Hindi, English
Featuring
Suraj Shetty, Head of CX, Kissht (Ring) Shadab, Senior Product Manager, Kissht
Within the first two months, the bot was handling forty percent of our entire inbound volume.
Suraj Shetty · Head of Customer Experience
Achievements
40%
of total inbound volume within two months — all inbound, not a pilot cohort
88%
of those calls resolved with no human involvement
↓ 30%
lower average handle time than human specialists, with customer experience holding steady
Challenge
The queue where being right most of the time isn’t enough
When a loan application doesn’t come through, the customer who calls in isn’t looking for a policy document. They want a straight answer and clarity on what they can do next and they want it in the first minute, not the fifth.
A senior specialist gets that right nearly every time. But Kissht’s capacity on that queue was arithmetic: nine hours a day, as many customers as there were specialists on shift. Everything outside that became a queue waiting for the next working day. And “nearly every time” doesn’t scale cleanly to thirty thousand conversations a month — consistency, not capability, was the constraint.
How we built it
Two queues, drawn tight, and nothing else
2
queues in scope — loan rejection, mandate/NOC cancellation
500ms
turn latency, on Arrowhead’s own small language model
Same turn
handoff to a human on any mention of the RBI
The scope was drawn narrow on purpose and enforced hard. The agent handles two call types and nothing else. If a customer so much as mentions the Reserve Bank of India, the call hands off to a human inside the same conversational turn — no attempt to answer, no negotiation.
Inside those boundaries, the agent behaves like a senior specialist. It is integrated directly into Kissht’s dialer and CRM, so the customer’s record and application status are loaded before the call connects. Nobody identifies themselves twice. Nobody repeats their story.
The constraint that mattered most wasn’t accuracy. It was silence. That half-second is the difference between a customer feeling they are talking to a system and feeling they are having a conversation. On a call about a rejected loan application, it is also the difference that decides whether they stay on the line at all.
Results
40% of everything that comes in
The ramp was controlled, and then it wasn’t small any more. Within two months the agent was handling 40% of Kissht’s total inbound volume — not 40% of a carved-out test cohort, but 40% of every call that reaches the support line. Of those, roughly 88% close out without a human being involved at any point.
How we measured
Volume share is measured against total inbound calls across all queues. Resolution rate counts calls closed without any human involvement. Handle time is compared against the same queues handled by human specialists over the same period.
Within the first two months, the bot was handling forty percent of our entire inbound volume.
Suraj Shetty · Head of Customer Experience
Economics
Faster, without the trade-off everyone assumes
↓ 30%
average handle time vs. human specialists
~30%
saving on every call the agent resolves end to end
And that saving is calculated before counting the calls that never needed a person on them in the first place.
But the number Kissht’s CX team points to is the one that didn’t move.
The bot’s average handle time is running about thirty percent lower than our human agents.
Shadab · Senior Product Manager
Coverage
The hours nobody could cover
24/7
coverage
100%
inbound
~20%
of volume between 8 PM and 8 AM
Specialists work fixed hours. Customers with a rejected application don’t wait for the support floor to open — roughly a fifth of Kissht’s inbound volume arrives between 8 PM and 8 AM, and all of it used to become a next-day queue.
That volume is now answered under the same policy, the same scope and the same guardrails that apply at nine in the morning.
Around twenty percent of our inbound volume comes between 8 PM and 8 AM. Earlier, that became a queue waiting for the next day.
Suraj Shetty · Head of Customer Experience
Partnership
Built alongside, not delivered to
Both teams put the speed of the ramp down less to the model than to the working rhythm around it. Daily standups from day one rather than weekly reviews. Arrowhead’s prompt engineers listening in on live Kissht customer calls. Direct access to Arrowhead’s engineers and founders outside business hours, instead of a ticket queue and one assigned account manager.
We had daily standups with the Arrowhead team from day one. Not a weekly review — an actual working rhythm. Their prompt engineers were sitting in on our calls, listening to our customers, and working alongside our team. That is not how vendors usually work.
Shadab · Senior Product Manager
What's next
Next: the queue where a mistake costs money
We started with a queue where the risk was contained. And now we are moving into EMI and repayment journeys, where the stakes are much higher. We would not be doing that if the first one had not held up.
Suraj Shetty · Head of Customer Experience
The deployment in numbers
Share of total inbound volume, by month 2
40%
Resolved without human involvement
~88%
Average handle time vs. human specialists
~30% lower
Cost impact on bot-resolved calls
~30% saving
Customer experience
Held steady
Turn latency, down from ~1 second
~500ms
After-hours volume covered (8 PM – 8 AM)
~20%
Coverage
24/7
Call direction
100% inbound
Live use cases
Loan rejection; mandate/NOC
In build
EMI and repayment