AI collections for NBFCs: voice, WhatsApp and compliance

AI collections for an NBFC is a narrower thing than the phrase suggests, and it is better for being narrow. The AI handles the high-volume, low-judgement part of the cycle: pre-due reminders, early-bucket follow-ups, capturing a promise to pay, sending a payment link, confirming receipt and answering the ordinary questions that come with each of those. It works across voice calls and WhatsApp in the customer's language, within calling hours and contact frequency rules enforced by the system rather than by memory, with consent recorded and every contact logged. Anything that needs judgement, from a hardship case to a dispute, is handed to a human with the full history. This article sets out what that looks like, how the compliance constraints shape the design, and what to measure.

Why debt collection automation in India is a conduct problem first

Collections is the most regulated conversation an NBFC has with a customer. The Reserve Bank of India's fair practices and outsourcing guidance sets expectations on the conduct of recovery agents, including hours of contact, courtesy, and the lender's responsibility for agents acting on its behalf; the digital lending guidelines add requirements on disclosure and grievance handling. An AI agent is an agent acting on your behalf, and the design has to make compliant behaviour the only behaviour available. That means calling windows enforced in code, a cap on attempts per day and per week, no contact with anyone but the borrower and permitted references, scripts that identify the lender and the purpose, a clear route to a human, and a recording and transcript for every interaction. Get this right and the automation is defensible; get it wrong and it is a regulatory finding at scale.

Voice: what a collections voice agent does on a call

A collections voice AI identifies the lender, confirms it is speaking with the borrower, states the purpose, and then does a small number of things well: states the outstanding amount and due date, asks whether payment has been made or when it will be, records a promise to pay with a date, sends a payment link by SMS or WhatsApp during the call, and confirms receipt when the payment lands. It handles interruptions, works in Hindi, Kannada, Tamil, Telugu and other Indian languages, and keeps to a tone that a compliance officer has reviewed. If the borrower raises hardship, a dispute, a complaint or anything the script does not cover, the agent stops collecting, acknowledges, and schedules or transfers a human. The mechanics of the stack are in AI voice agents for collections and payment reminders; the per-minute cost model is in how much does an AI voice agent cost per minute.

WhatsApp: reminders, links and two-way conversation

WhatsApp is cheaper than a call and less intrusive, which makes it the first channel for pre-due and due-day reminders. A template message carries the amount, the date and a payment link; the borrower can reply, and the agent answers questions about the amount, the breakdown, how to pay and what happens next. The constraints are Meta's business messaging policies, which require opt-in and approved templates, and the lender's own consent record. The design rule we apply is that WhatsApp never carries anything that would embarrass the borrower if seen by someone else on the phone: no threats, no third-party references, no language beyond a neutral reminder. The channel's capabilities are covered in WhatsApp AI chatbot for business.

Consent, contact rules and the record

Three records make the system auditable. Consent: when and how the borrower agreed to be contacted on each channel, captured at onboarding and re-confirmed if the number changes. Contact policy: calling windows, attempts per day and per week, cool-off after a promise to pay, and a do-not-call flag honoured across channels, enforced by the orchestration layer so no prompt or model can override them. Interaction log: every message and call with transcript, recording, outcome, promise details and any hand-off, retained per policy and exportable per borrower. The DPDP Act's purpose limitation applies to all of it, and telephony rules from TRAI govern outbound calling; see voice AI compliance in India for the detail.

Where the line to a human sits

The line is drawn by bucket and by conversation content. By bucket: the agent works pre-due through the early delinquency window; later buckets are human-led, with the agent doing scheduling and reminders for the collections team rather than talking to the borrower. By content: any mention of hardship, job loss, medical circumstances, death, dispute over the amount, a complaint, a threat or a request to stop is a hand-off. Hand-offs arrive with the full history so the human does not restart the conversation. The list is reviewed monthly with compliance, and the escalation reasons are the roadmap for what to add and what to keep human. The control pattern is the same policy-gated design described in how to build an AI agent that is safe to run unattended.

Measuring collections AI honestly

The number the business wants is cash collected per bucket compared with the previous process. To make it honest, run the agent on a share of accounts and keep the rest on the existing process for a period, with matched buckets, so the comparison is real rather than seasonal. Alongside it: contact rate, promise-to-pay rate and kept-promise rate, payment-link conversion, hand-off rate by reason, complaint rate, and cost per account contacted. Compliance metrics are reported in the same pack: contacts outside window (target zero), attempts over cap (zero), do-not-call breaches (zero) and a sample of transcripts reviewed for tone each month.

A worked example

An NBFC with a large book of small-ticket loans was running collections through a call-centre partner whose agents were expensive per contact and inconsistent in conduct, which had produced customer complaints. We built a pre-due and early-bucket agent on WhatsApp and voice in three languages, with calling windows, attempt caps and a do-not-call register enforced in the orchestration layer, payment links issued in-conversation, and hand-off to the NBFC's own team for hardship, disputes and later buckets. The agent ran in shadow mode listening to recorded calls first, then on a slice of accounts with the rest on the old process. The kept-promise rate on the agent's slice was tracked against the control group, complaint volume was monitored, and a transcript sample was reviewed weekly by compliance for the first quarter. The partner's role narrowed to the buckets where a human conversation is the right tool. The voice agents page describes the stack; a comparable multilingual build is the voice agent for a hospital network.

Team and timeline

A collections agent is typically a Launch 6 build (six weeks, fixed price, $26,500–45,500 or from ₹17,60,000) covering the WhatsApp and voice channels, contact policy engine, payment-link integration, hand-off and the reporting pack, after a Sprint Zero discovery (ten working days, $3,250, credited) that maps buckets, scripts, consent records and integrations with the loan management system. Voice agents as a service start at $17,500 or ₹11.2L plus usage of $0.05–0.15 per minute; see the pricing page. The team is a lead engineer, a voice and conversation engineer, an integration engineer, and on your side a collections head who owns the scripts and a compliance owner who signs off the contact policy. Sector context is on the fintech page.

Before you start: a checklist

• Consent records for each channel, and a plan to capture them where missing
• Contact policy written down: windows, caps, cool-off, do-not-call handling
• Scripts by bucket and language, reviewed by compliance
• Hand-off rules and the team that receives them, with capacity
• Payment gateway or UPI link integration and receipt confirmation
• Loan management system integration for balances, due dates and outcomes
• A control group design for measuring collections lift
• Retention rules for recordings and transcripts, and the grievance route

Questions clients ask

• Can the agent negotiate settlements? No. It can capture a request and hand off. Settlement authority stays with your team under your policy.
• Will borrowers know it is an AI? Yes. The agent identifies itself and the lender, and offers a human on request.
• What languages are supported? Major Indian languages including Hindi, Kannada, Tamil and Telugu, with the script tested per language rather than translated once.
• How are calling hours enforced? In the orchestration layer, not in the prompt, so no model behaviour can override them.
• Does this replace the collections team? It replaces the repetitive early-bucket contact and frees the team for the conversations that need judgement.

Related reading

Outbound AI calling: use cases that work and ones that don't, AI in lending and AI voice agents for Indian languages cover the adjacent decisions.

Automate the reminders, enforce the rules in code, hand the hard conversations to people, and keep a record of all of it.