Most missed payments are not refusals. The customer forgot, the salary came late, the auto-debit failed because the account was short by a few hundred rupees. A reminder at the right moment fixes most of them — and a lender, NBFC, school or subscription business has thousands of them to make every month.
That is a job AI voice agents are well suited to, and one where getting the tone or the rules wrong costs more than in almost any other. Here is how to run AI payment reminder calls that recover money without turning into harassment.
Reminder, not recovery
The single most useful line to draw is between a reminder — before or just after the due date, to someone who probably intends to pay — and recovery, for accounts that are seriously overdue and disputed or in hardship.
An AI agent is excellent at the first: polite, consistent, never tired at call number four hundred, and exactly on script. The second needs judgement, negotiation and empathy, and for regulated lenders it comes with its own conduct rules. Automate the reminder; hand recovery to people.
The rules to check first
If you are a bank or NBFC regulated by the RBI, its guidelines on recovery conduct apply to calls made on your behalf, including automated ones. Among them: calls only between 8am and 7pm, no threatening or abusive language, no calls to friends or family to shame the borrower, and respect for privacy. A badly written AI script can break those rules at scale.
Separately, telecom rules on commercial and service calls, the DND registry and your consent records apply to everyone. Service and transactional calls for financial institutions are moving to designated number series; confirm with your carrier and compliance team what applies to your calls. Consent, DND and calling hours.
Not a lender — a school collecting fees, a gym, a SaaS renewal? The RBI rules do not bind you, but they are a sensible standard to borrow.
Designing the sequence
- Three days before due — WhatsApp
- A utility template: amount, date, payment link. Most people pay here, and it costs a fraction of a call. Taking payment inside WhatsApp.
- On the due date — AI call to those who have not paid
- Short and friendly. Confirms the amount, offers to send the link again, asks when they expect to pay.
- Two to three days after — WhatsApp, then a second call
- Same tone, now with the promise they gave last time.
- After that — a person
- Anyone still unpaid, disputing, or in difficulty goes to your team with every attempt on the record.
Each step stops the moment the person pays. Nothing damages trust faster than a reminder for a payment made yesterday. How many attempts is enough.
Writing the call
Say who you are and that it is automated
"Hello, this is the automated assistant from Sharma Finance, calling about your EMI." Hiding that it is a bot saves nothing and costs trust when they work it out — and they will. Why people hang up on AI calls.
Verify before you disclose
The person who answers may not be the borrower. Confirm identity — a name and one detail — before mentioning an amount. Disclosing a debt to a family member is a privacy failure and, for regulated lenders, a conduct one.
Ask for a date, not a promise
"When do you expect to make the payment?" gets a date you can record and follow up on. "Will you pay?" gets "yes" and nothing else.
Know the exits
Write handover conditions for: a dispute ("I already paid", "that amount is wrong"), hardship ("I lost my job", a medical emergency), a request for a person, distress or anger, and anyone who says not to call again. None of these are the AI's to handle. Designing the handover.
Never improvise on money
No waivers, no settlements, no late-fee promises the agent has not been given. Every figure it says comes from your system through a lookup, or it does not say it. Stopping an agent inventing things.
What to record, and what to measure
Every call should leave an outcome on the record — reached, no answer, callback requested, promised date, dispute, do not call — with the transcript and recording. That is your audit trail if a customer complains, and the input to the next step.
Measure:
- Paid within 3 days of the call — the number that matters.
- Promise-kept rate — how many of the dates customers gave were honoured.
- Handover rate and reasons — rising disputes point at a billing problem, not a calling one.
- Complaints — should be close to zero. If not, read the transcripts today.
Call metrics worth a dashboard.
Running it in Convarza
Build the audience in Contacts from your own fields — due date, amount, status — with the filter builder, and run a campaign written in plain sentences: "Send the WhatsApp reminder. On the due date, call. If no answer, call again in four hours…". It stops per person the moment they connect, and never calls outside the calling hours set on your calling channel — a campaign can narrow those hours but never widen them.
The AI voice agent speaks English or Hindi on your own Twilio, Maqsam or SIP numbers. Every call lands on the contact with outcome, transcript and recording, and outcomes do what they say: Callback requested creates a task at the time set, and Do not call removes the person from calling lists. If you use HubSpot, Zoho, Pipedrive or Freshworks, calls are logged there too. Convarza for fintech teams.