The EMI that didn't land
In a typical collections review, two patterns show up again and again. An account moves to the next bucket because the borrower was never reached while the balance was still one EMI. Another generates a complaint because outreach continued after the payment had already posted.
The lender in this piece is made up. Call it Northline Finance, a mid-size NBFC with a two-wheeler book and a personal-loan book. The accounts are the ordinary ones. If you run recovery at a bank or an NBFC, all three are on this week's MIS: a bounce from Saturday, a borrower on day 18 who is stuck rather than refusing, and an account on day 40 that should have been taken off the dialler.
The bounce
Rakesh's two-wheeler EMI is ₹4,850. The NACH file went out on Saturday and came back the same day: insufficient balance. He did not see it. His salary hit the account on Monday morning.
Agent: Rakesh ji, Northline Finance se bol rahi hoon. Aapki two-wheeler EMI, ₹4,850, Saturday ko NACH se nahi kati. Account mein balance kam pad gaya tha. Main payment ka link bhej doon?
Rakesh: Haan, bhej do. Salary aaj aayi hai. Shaam tak kar deta hoon.
Half a minute. The loan, the amount, and the bounce reason were already on the screen, matched from his number. The call could open on the failed EMI and close on a link and a time someone can check on Monday evening.
This is the volume call. It sits in the first two or three days of SMA-0, when the account is overdue and still relatively easy to cure. The script is straightforward and the same for most borrowers, but the number of accounts is large. A Saturday debit cycle can put thousands of names on Monday's file, most of them people who will pay as soon as they know the debit failed.
In many setups, that same file is worked a few days later. Connect rates drop. Borrowers who do answer have often already seen a bounce SMS or a bounce charge, and some have used the salary they planned to put toward the EMI. An earlier call on Monday can avoid both problems.
A voice agent can make this call on Monday morning, in the borrower's language, for every bounce in the file. The human team hears about the ones that do not pick up, or that say something other than yes.
The tight month
Meena is on day 18. Still SMA-0, and no longer a bounce. Her personal-loan EMI did not go through, and a second debit will come back the same way.
Agent: Meena ji, is mahine ki EMI abhi pending hai. Payment mein koi dikkat aa rahi hai?
Meena: Salary late hai, aur bete ka hospital bill aa gaya. Poori amount is hafte nahi ho payegi.
Agent: Samajh gayi. Aap ₹3,000 is hafte kar sakti hain, aur baaki 28 tarikh ko? Main yahi note kar deti hoon, link bhi bhej deti hoon.
That second date is inside a window Northline's credit policy already allows: a part payment now, the rest before the month turns, no change to tenure, no waiver. The agent is reading the policy out loud, not inventing a plan. The note on the account has an amount and a date, which is what makes it checkable on the 28th.
A promise works for the book when it includes a date and an amount. Vague commitments are harder to follow up on and do not always translate into payment.
If Meena says even ₹3,000 will not happen, the call changes job. The agent stops offering. It writes down what she said, the hospital bill, the late salary, the amount she thinks she can do next month, and a person calls her with that conversation already attached. The person can open on the decision, because the story has been told once.
Many teams worry about putting this call on a voice agent. That concern is reasonable if the agent can go beyond approved options. It works better when credit policy defines the script: part payment, a date, then a handoff. Restructuring, settlement, or a change in tenure stay with a person who has authority to offer them, with the earlier conversation attached so the borrower does not repeat their story.
The call that should end
Imran is on day 40, inside SMA-1. The balance on the screen still shows one missed EMI. He paid it on Friday by UPI. The posting lagged, and the dialler did not know.
The first call of the morning does not reach him.
Agent: Namaste, kya main Imran ji se baat kar sakti hoon? Northline Finance se call hai.
Woman: Main unki wife hoon. Boliye, kya baat hai?
Agent: Ji, yeh baat unhi se karni hogi. Main yahan koi detail nahi de sakti. Shaam 6 baje phir try kar loon?
The amount stays unsaid. So does the overdue. Whether it is a wife, a colleague, or a neighbour on the landline, the agent keeps the same boundary: no loan details to a third party. That detail is what often appears later in a complaint.
An hour later Imran answers his own phone.
Imran: Maine Friday ko UPI se de diya tha. Baar baar call kyun aa rahi hai?
Agent: Main dekh leti hoon. Haan, ₹6,200 Friday ko credit hua hai, system mein update late aaya. Main yeh call yahin band karti hoon.
When the credit is on the screen, the call is over. When it is not, she still stops asking for the money. She logs the claim, the amount he says he paid, and the time, and a person reconciles it before the next attempt. That reduces the chance of a second payment and a later refund dispute.
The same approach applies to other cases that should close quickly: a dispute, a request to stop calling, mention of a lawyer, or the wrong person on the line. Each outcome gets a clear note and the account is paused. Sending another payment link is usually not the right next step.
This is where recovery-conduct risk is highest: calling after payment has posted, sharing balance details with a family member, dialling outside permitted hours, or quoting settlement terms the caller is not authorised to offer.
A voice agent can handle these endings consistently — wrong party, already paid, stop calling — and log the result on the account before the dialler places another attempt. Human teams under volume and time pressure can find it harder to apply the same rules on every call.
What Monday's MIS should show
Vendor dashboards often highlight operational metrics: containment, connect rate, calls per hour. Collections leaders also need numbers that tie calls to the book. These five usually already appear in the weekly pack:
- Right-party contact. Did the call reach the borrower, not just someone who shares the phone? Speaking to a spouse about a loan is a common source of complaints.
- Promises kept against promises made. Promise rate shows how calls ended. Kept rate shows whether money arrived. A wide gap may point to script or policy issues.
- Roll rate by bucket. Did fewer SMA-0 accounts move to SMA-1? That is what the credit committee tracks. Higher connect rates alone do not help if accounts still roll forward.
- Complaints per thousand calls. Recovery gains matter less if complaints rise — for example, calls after payment or disclosure to a third party.
- Cost per rupee recovered. Low cost per attempt helps only if attempts are well targeted. Extra calls to paid accounts or after a stop request add cost through refunds, waivers, and complaint handling.
When voice AI is added to an existing list, it helps to review both operational and book metrics together. We wrote about that gap in banking here.
The bucket to start with
One bucket is enough. NPA is a poor place to begin: the script is no longer standard, the judgment is the work, and a missed nuance is expensive in both money and conduct.
Start where the call is already a script and the file is thickest. For most lenders that is the first few days after a bounce, while the account is still one EMI and the borrower still has a reason to pick up.
Three questions worth taking to the book this week:
- Which bucket has a standard script and the most accounts? That is the first call to give to a voice agent. It is usually the bounce, not the negotiation.
- Of the promises logged last month, how many cleared? The gap is the call to rewrite, before anyone adds attempts.
- Which calls produced the complaints? Define safe endings for those cases first, then layer on the payment ask.
Before the first bounce call, put the safe endings in place for cases like Imran's — already paid, wrong party, stop calling. Then run the bounce for a month and track roll rate and promises kept, alongside calls per hour. Once those rules are stable, add the tight-month call, with a clear part-payment window in the script. If you want the build itself, the setup steps live in the collections cookbook.
Most accounts that roll to the next bucket are still only one EMI behind. Reaching the borrower early is often the simplest way to keep them there.
Curious what this would sound like on your book? Talk to us →
