Collections
AI voice agents for NBFC collections
An NBFC’s early-bucket book is mostly reminders — borrowers who intend to pay and need the due date restated or the payment link resent. Doing that with people means a roster, a script nobody delivers the same way twice, and a compliance surface that grows with headcount. Telenow makes those calls in the borrower’s own language, inside the calling window, against your suppression list, with every conversation recorded — and hands the ones that need judgement to your team.
Last updated 2026-09-06
Which buckets this is actually for
A voice agent is very good at the part of the book where the conversation is the same every time, and poor at the part where it is not. Being straight about that line is the difference between a pilot that survives and one your collections head quietly kills in month two.
Pre-due and due-date reminders
D-3 and D-0 nudges at scale: the amount and date read from your own data, the payment link sent on the call, no queue and no engaged tone even when the whole batch falls on the same morning.
0–30 DPD
Where the agent pays for itself. A soft reminder, a resent link, a promise-to-pay date captured as a field, and a clean disposition on every number you dialled — including the ones nobody would have got to.
31–60 DPD
Useful for reach and intent capture: find out whether the borrower is contactable, willing and disputing, and get that in front of a human with the transcript attached instead of a callback note.
60+ DPD, settlements and disputes
Not this. Hard-bucket negotiation, hardship and settlement conversations are a person’s job — legally, commercially and ethically. The agent’s value there is finding the right person and warm-transferring them with context.
The gates in front of a collections dial
Four gates, each stated with where it sits rather than folded into one sentence about “every dial”. The precision is the point: a control that sits under unattended dialling and a control that sits under a person clicking call are different controls, and your auditor will ask which is which.
- 1Suppression check
Your organisation’s Do-Not-Call list is consulted before the number is dialled, on the paths a collections team places calls from: a campaign, a follow-up the agent scheduled itself, an operator clicking dial in the softphone or calling the dial endpoint, and a transfer out of a live call. It is not a per-campaign setting anyone can leave off.
- 2Calling-window floor
On the scheduled paths — campaigns and agent-booked follow-ups — a +91 number is held outside 09:00–21:00 IST by a destination-country floor covering mobile and landline. It intersects with your own campaign window rather than replacing it, so a stricter internal policy still wins. Held numbers stay in the queue for the next legal slot instead of being dropped — worth knowing, because a large batch loaded at 20:45 will look stalled rather than finished. It does not sit under a manual dial: an operator clicking call at 22:00 is checked against your suppression list and your consent policy, but not against the clock.
- 3Consent check, if you have switched it on
With consent enforcement enabled, a contact with no valid consent record is not dialled at all. It ships switched off on purpose: turning it on blind can silence most of a book overnight, so there is a coverage report that tells you how much of yours would go dark before you flip it.
- 4Record, transcribe, dispose
The call is recorded and transcribed, a structured disposition is written — promise to pay, already paid, wrong number, dispute, requests removal, needs a human — and anything the agent should not decide is escalated rather than improvised.
What the agent does on the call
It reads from your data, not from its imagination
Outstanding amount, due date, instalment number and reference come in as call variables from your system. The agent restates them; it does not compute them, round them or fill a gap with something plausible.
It sends the payment link mid-call
WhatsApp or SMS while the borrower is still on the line, then confirms it landed. The gap between “I’ll pay” and having the link in hand is where a large share of early-bucket promises die.
It captures a promise to pay as a field
A date and an amount written back as structured data, not a free-text note — so your next batch can be built from what borrowers actually committed to rather than from a summary someone typed later.
It has a defined path for the four hard replies
“I already paid”, “this isn’t my number”, “call me later” and “stop calling me” each end the call a specific way: flagged for reconciliation, flagged as a wrong number, rescheduled inside the window, or written to the suppression list with the recording attached.
It does not argue
A borrower who disputes the amount gets the dispute recorded and a human, not a rebuttal. You script where that line sits; the agent stays on the safe side of it.
It hands over warm
Ask for a person and the call transfers to your collections desk with the conversation so far, so the borrower does not start again from their name.
Language is a collection rate, not a feature
A reminder in the borrower’s own language gets answered differently from one in English, and in a semi-urban book that gap is most of the result. The engine speaks English plus ten Indian languages — Hindi, Tamil, Telugu, Malayalam, Kannada, Bengali, Marathi, Punjabi, Gujarati and Odia — and handles Hinglish and mid-sentence switching rather than breaking on it. One piece of hard-won advice: pin the language per agent instead of trusting auto-detection. On a live box we watched detection walk hi, hi, ur, en, ur across ten turns of a single call with one speaker. Pinning is a real control, and we would rather tell you to use it than let you find that out on your own book.
The evidence you are left holding
Telenow handles 30,000+ calls a month across our customers, and the reason a regulated lender can run volume like that on an agent is that each call leaves the same paperwork behind whether anyone looks at it or not.
- A recording of every call, stored with AES-256 server-side encryption on a retention window you set per organisation.
- A transcript with a structured outcome, so a bucket can be reported on without anyone listening to audio.
- An opt-out row carrying a link to the exact call it was asked for on — and a withdrawal flow that is one at a time, audit-logged, and offers to play the call first.
- A consent ledger recording source, timestamp, disclosure text and version, whose revocation writes suppression in the same transaction.
- Hash-chained audit checkpoints over platform changes, verifiable and exportable to CSV — tamper-evident, so later alteration is detectable.
- A legal hold that freezes a tenant or a single call against every deletion path, checked in SQL.
- Payment card numbers stripped from stored transcripts by default, so card data never enters your call archive.
What we are not claiming
- No RBI or TRAI approval — no regulator has reviewed this platform — and no compliance certification of any kind. The obligations are yours; what we do is make them enforceable in the dial path.
- No integration with the national NDNC or DND registry. The suppression list is your organisation’s own.
- No recovery-rate or contactability promise. Anyone quoting you a lift figure for a book they have never seen is guessing.
- No claim that this replaces your collections team. It removes the reminder calls so the team can spend the day on the accounts that need a person.
Read the mechanics, not the marketing
Everything on this page is a product behaviour with documentation behind it. Before a pilot, the two pages worth an hour of your compliance team’s time are how the Do-Not-Call list is enforced and how campaign windows and the India calling floor interact — including the E.164 detail that decides whether a number gets a floor at all.
Frequently asked questions
Will the borrower know they are talking to a machine?+
If they ask, yes — the agent answers honestly, and it will not pretend to be a named human. Disclosure today is reactive by default, which means a borrower who never asks is never told, so most collections teams script a proactive line into the opener instead. We would recommend you do the same: it costs two seconds, it removes an argument later, and it is the sort of thing a regulator asks to hear on the recording.
What stops it calling someone at 10pm?+
On the paths that dial without anyone watching — campaigns, and the follow-ups an agent books itself — a destination-country floor holds +91 numbers outside 09:00–21:00 IST, covering mobile and landline, and it intersects with your own campaign window so you can only make it stricter. Two practical details before you upload a list. The floor keys on the E.164 form of the number, so store numbers with the +91 country code; a bare ten-digit number is not recognised as Indian and gets no floor. And it is a floor under automated dialling rather than a lock on the phone system: an operator who dials one number by hand from the softphone at 22:00 is checked against your suppression list and your consent policy, but not against the clock. If you need that path closed too, say so — it is a change we would make, not a control we would pretend already ships.
A borrower asked to be removed mid-call. What actually happens?+
The agent writes an opt-out entry against your organisation’s suppression list with a link to that recording, and every subsequent dial is blocked by it — campaigns, agent-scheduled follow-ups, manual calls and live transfers alike. Those entries cannot be removed in bulk. Withdrawal is one at a time, audit-logged, and the screen offers to play you the call before you do it.
Can it actually collect the money?+
It sends the payment link on the call and confirms it arrived, and it captures a promise-to-pay date as structured data. It does not read card numbers over the phone: we hold no PCI DSS attestation, and card numbers are stripped from stored transcripts by default so that card data never reaches your call archive. Reconciliation stays with your payment stack, where it belongs.
What does it do when a borrower disputes the amount?+
It stops collecting and starts recording. The dispute is captured as a disposition, the agent does not argue the balance or make a commitment on your behalf, and the account is routed to a human — with a warm transfer there and then if the borrower wants one. You define where that line sits when you build the agent; the default is to escalate early rather than to hold the line.
How do you handle wrong numbers and third parties?+
You script identity confirmation before anything about an account is said — that is a design decision the flow enforces, not a hope. “Not my number” is a first-class disposition: the call ends, the record is flagged for your team, and the number stops being dialled on that account. Discussing a debt with someone who is not the borrower is a real risk in Indian collections, and the shape of the call should make it hard to do by accident.
Does this replace our tele-calling team?+
No, and a vendor promising that has not run a collections floor. It takes the reminder layer — pre-due, D-0, early-bucket, link resends, confirmations — which is where most of the dial volume and almost none of the judgement lives. Your team keeps the accounts that need negotiation, and gets to them with a transcript and a disposition already attached instead of a dialler queue.
How quickly can we pilot it?+
Most teams start from a ready-made template, connect a number, upload a small slice of one bucket and place a live test call to themselves the same day. Start narrow — one bucket, one language, a few hundred accounts — and read the transcripts before you widen it. We would rather you judge it on your own recordings than on a demo of ours.
Can we run this without our borrower audio reaching any third-party AI vendor?+
The audio, yes: recognition and synthesis both run on models we host ourselves, so borrower audio never reaches a third-party speech vendor. Be clear about the rest of it, because your security review will be. The language model on the hosted platform is not one of ours — it is a vendor you choose, and every turn goes to it as text: the borrower’s name, the outstanding amount, the wording of a dispute. That vendor is a subprocessor and belongs in your record as one. Two ways out if that is unacceptable: point the agent at your own OpenAI-compatible model endpoint, which keeps the reasoning on infrastructure you run, or take the appliance, where the language model is a quantised Qwen3-8B on your own GPU and nothing about the call leaves the building.
Start with one bucket, one language, and read the transcripts
Sign up free and get $1.01 in credit — no card required. Connect your number, pick a template, and go live in minutes.