AI Loan Payment Reminder Calls for Fintech: Live EMI, RBI Hours

AI loan payment reminder calls for fintech lenders ring each borrower a few days before the EMI date. As the borrower answers, the agent fetches the current amount, due date and paid status from the lender's system, verifies the borrower, then reads them. Calls stay inside RBI's 8 a.m. to 7 p.m. window, in the borrower's language.

Key Takeaways

• A campaign file is a snapshot, so fetch the EMI amount as the borrower answers.
• Keep pre-due reminders inside RBI's 8 a.m. to 7 p.m. recovery-call window.
• Greet in two languages, then let speech recognition follow the borrower.

This post is part of our guide to AI voice agents for BNPL lenders, where every call reads the right amount for each borrower. Here we apply it to Indian EMIs, from the live lookup to RBI's calling hours, in Dograh, the open-source voice agent platform we build.

Why a reminder list goes stale

Most EMI reminder systems start the same way. Each morning, the loan management system pulls the loans due in the next few days and hands the list to a dialer. The trouble starts after the list is built.

In a Dograh campaign, every column in the uploaded file becomes a variable the agent can read, such as first_name, emi_amount and due_date. Those values are fixed at upload. A borrower who pays through UPI at 11 a.m., part-pays, or has a loan restructured that afternoon still carries the morning's numbers in the file.

Reading a stale amount to a borrower who has already paid is the fastest way to lose their trust, and a wrong EMI figure on a recorded call is a complaint waiting to happen. The file is still the right place for who to call. The amount should come from somewhere fresher.

Reading the EMI when the call connects

Pre-Call Data Fetch on the Start Call node asks your system for current values on every call. Set it to Outbound calls only and give it your endpoint URL and, optionally, a credential.

For an outbound call, the request goes out after the borrower answers. Dograh posts the call's run ID, the agent ID and both phone numbers to your endpoint, and the borrower hears ringing while it waits. Your endpoint looks up the borrower by the number being called and returns values such as emi_amount, due_date and paid under initial_context. Those values replace the file's for this call and fill the greeting and the prompt.

Plan for a slow day. The request times out after 10 seconds, and if it fails or times out, the call carries on with the values from the file. Tell the agent in the prompt to treat a missing paid value as unknown and say a team member will confirm the amount, rather than reading a figure it cannot trust. Each request also carries an Idempotency-Key header, so your endpoint can ignore a repeat.

The amount still waits for the identity check. The prompt asks for the lender's approved check, such as a date of birth or the last digits of the loan account, and reads {{emi_amount}} and {{due_date}} only after it passes. A borrower whose lookup says paid hears thanks and a short goodbye.

Open Source Alternative to Vapi / Retell

Self-hosted voice agent platform — no per-minute fees

dograh-hq/dograh

Star on GitHub

Calling hours and the number you call from

India's rules on this are specific. RBI's directions on recovery practices bar regulated lenders and their agents from calling borrowers before 8 a.m. or after 7 p.m. to recover overdue loans, and from calling persistently. A reminder before the due date is generally not a recovery call, but keeping reminders inside that window is the safe default, and any call after a missed EMI falls squarely under it. Some vendors advertise reminders until 9 p.m., so check the hours before you copy a ladder. Confirm your calling hours and number setup with counsel before the first campaign.

The number matters too. RBI's circular on calling numbers tells regulated entities to place service and transactional calls from 1600-series numbers and promotional calls from 140-series numbers, and not from ordinary 10-digit mobile numbers, including through partners. An EMI reminder is a service call, so it goes out on the lender's 1600-series line once you have one.

A campaign runs on one time zone, which suits an all-India list on IST. Set the Call Schedule to start after 8 a.m. and finish well before 7 p.m., so the last retry still lands inside the window. Calls after a missed EMI, and the Hindi stack that suits them, are covered in Hindi AI voice agents for collections.

Reminders in the borrower's language

A borrower who opened the loan in Tamil should hear the reminder in Tamil. The campaign file can carry a greeting for each row: add greeting_override_type set to text and a greeting_override_text column in that borrower's language, or point to a recorded clip by its ID. Rows left empty use the agent's default greeting.

Sarvam voices cover Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu, and Bulbul v4 Flash is the newest of them. Sarvam's Saarika speech model can detect which of these the borrower is speaking. The catch is order: the greeting plays before the borrower says anything, so the agent cannot know the language yet. Where the file has no language, open in two, such as Hindi then English, and let the reply set the language for the rest of the call.

Fixed lines such as the greeting and the goodbye can be pre-recorded clips in a real person's voice, with text-to-speech reading the amount and date. Your lookup can also return a model_overrides block that sets the voice for that one call, so a Marathi borrower hears a Marathi voice from the same agent.

The same lookup for US lenders

US lenders running monthly instalment loans use the same lookup with different rules. There, an AI voice counts as an artificial voice under the Telephone Consumer Protection Act, and the FCC's calling rules expect prior express consent for a reminder to a mobile. Timing and a short script for pay-in-four plans are covered in BNPL payment reminder calls with AI.

Checking that the right amount was read

The one mistake that matters most on a reminder is the wrong figure. Every call keeps a full trace, so you can open any call and see the values the lookup returned next to what the agent said.

For a regular check, set QA to sample a share of calls and replace its System Prompt with checks such as whether the agent read an amount before the identity check passed, or read one that did not match the fetched value. A borrower marked paid should never be spoken to as if they owed money, so add that check too.

Self-hosted, the recordings, transcripts and traces stay on your own servers, and Sarvam can be swapped for open-source speech and voice models running there too.