Recovering EMIs with zero humans on the line

How a top-5 private bank ran promise-to-pay collections autonomously, in the customer’s own language.

Industry

Banking / lending

Use cases

EMI collections, promise-to-pay, callbacks

Languages

Hindi, English, Marathi

Achievements

↑ 20%

higher collection rate than human agents on any given bucket

ROI on every rupee spent on the bot, for write-off accounts

↑ 50%

higher connection rate than human agents

An EMI is four days late. The right call is firm but not threatening, in the language the customer is comfortable in, on the day they’re most likely to pay — and it has to sound that way every single time. A roster of agents can’t hold that line across a whole bucket: tone drifts, dates slip, and some accounts simply never get dialed. The bank wanted to know if one agent could.

Challenge

The tone has to hold across every account

Accuracy is non-negotiableTone is the hard partConsistency breaks at scale

Lending’s top of funnel is enormous and impatient. A team works the leads it can reach in a day, and the rest cool off — not lost to a competitor, just to a call that came too late or never came.

Reaching them wasn’t enough on its own. The agent had to answer real eligibility questions, hold a borrower through the awkward parts, and get them to a submitted application — without ever sounding like a script.

How we built it

The full recovery conversation, run the same way every time

The agent confirms who it’s speaking to, states the exact outstanding, and works toward a promise to pay. When the customer commits to a date, it checks the date makes sense and sets a callback for it. The payment link goes out on the call.

How firm to be, what to say once a promise is made, where the line is — all of it is fixed in advance. The agent doesn’t improvise on sensitive ground, so account number ten-thousand hears the same calibrated tone as the first.

A real call

Collections
(Pre-delinquency)

Collections
(Tamil)

Results

More recovered, and money back on accounts already written off

On the same Bucket X accounts, the agent reached customers 50% more often and collected 20% more than the human team. The number that changed the conversation was on the write-offs — accounts the bank had stopped expecting anything from. There, every rupee spent on the agent brought back four.

How we measured

Collection rate and connection rate are compared to the human team on the same Bucket X accounts, same period. ROI is computed on recovered value against bot cost for write-off accounts.

What's next

Built to scale across India’s languages

The agent now operates in 11+ languages across every state, and does far more than translate. Customers switch between Hindi and English, ask half-finished questions, and change their minds — and the agent follows all of it the way a person would.

Arrowhead AI calling bots

Speaks like a human.Performs like a machine.