First-bucket collections (DPD 1–30) is one of the most operationally painful parts of an Indian lender's collections pipeline: high volume, low average recovery per call, expensive human capacity, and a regulatory environment that punishes aggressive practices. It is also where AI voice agents are most likely to pay off, because the conversations are repetitive and the risk per call is low.
Why first-bucket is the right entry point
- Conversation is repetitive. A handful of conversation types — reminder, promise to pay, partial payment, dispute, wrong number — happen thousands of times. AI handles repetition well.
- Outcomes are clear. Promise to pay, partial payment, dispute, no contact: finite and trackable.
- Risk per call is low. A wrong tone in first-bucket is recoverable. The same agent on legal-stage collections is not.
- Right-party contact is the bottleneck. AI can dial across more permitted windows and retry intelligently, which is often the biggest lever before conversation quality matters at all.
Architecture in plain terms
Inbound and outbound calls flow through a telephony gateway (Twilio, Plivo, Exotel, Gupshup). Audio streams to an Indic speech-recognition layer (Bhashini, Sarvam or a custom model). The transcript flows into a domain-tuned LLM with retrieval over the borrower's case file. The response is generated as text, converted to speech in the borrower's preferred language, and streamed back. Every call is logged, transcribed, sentiment-tagged and stored in the lender's own cloud environment.
Two design decisions matter more than model choice. Keep end-to-end response time low enough that borrowers do not talk over the agent. And make sure case data — amount due, due date, payment-link status — always comes from your systems through tools, never from the model's memory.
RBI fair-practices compliance — what to engineer for
- Calling windows: 8am–7pm by default, with state- and customer-specific overrides.
- Identification: the agent must identify itself, the lender and the purpose of the call. Hardcode this; it should never be an LLM responsibility.
- Escalation: any aggressive language, threat or off-policy phrase triggers an immediate handoff to a human and a logged incident. Classify in real time.
- Recording disclosure: explicit consent at the start of the call, captured in the audio.
- Frequency and privacy: cap attempts per borrower per day and week, and never discuss the debt with anyone other than the borrower.
- Audit trail: every interaction reproducible end to end, so the regulator or internal audit can request and review it.
Language, tone and code-switching
Borrowers rarely speak one clean language. Hindi-English code-switching is normal, and many portfolios also need Tamil, Telugu, Marathi, Bengali or Kannada. Test speech recognition on your own recorded calls in each language, especially numbers, dates and amounts, which is where transcription errors cost the most. Keep the tone firm but polite, with approved phrasings for reminders and consequences written by the collections and compliance teams rather than generated on the fly. Read amounts and dates back to the borrower before logging a promise to pay.
The hybrid model that actually works
Design three lanes from day one. First, calls the AI completes end to end: reminders, promises to pay, payment-link requests. Second, calls escalated mid-conversation to a human agent who receives the transcript and case context, so the borrower never repeats themselves. Third, explicit fall-throughs that never go to the AI at all: legal-stage accounts, sensitive customer flags and vulnerable-borrower indicators.
How traffic splits across the lanes depends on your portfolio, so measure it rather than assuming it. As the first lane grows, the human team's capacity moves toward higher-DPD work, where it produces more recovery per hour.
What to measure
| Metric | Why it matters | Compare against |
|---|---|---|
| Right-party contact rate | The first bottleneck | Same segment, human dialling, same period |
| Promise-to-pay and kept-promise rates | A promise that is not kept is not recovery | Human agents on matched accounts |
| Roll rate to the next bucket | The outcome the business cares about | A held-out control group |
| Cost per contacted account | Unit economics | Fully loaded human cost |
| Escalation and complaint rates | Compliance and customer experience | Human baseline |
| Script adherence in QA audits | Regulatory risk | Human QA sample |
Run a controlled test: randomly assign comparable accounts to AI-first and human-first handling for a few billing cycles, and judge on roll rate and kept promises rather than on call counts.
Common mistakes
- Letting the model improvise mandatory disclosures.
- Measuring connected calls instead of kept promises.
- Launching in one language when borrowers speak several.
- No vulnerable-customer detection: bereavement, medical emergencies and distress must route to a person immediately.
- Treating the pilot as a technology test only. Collections managers and compliance should own the success criteria.
FAQ
Can the agent take payments on the call?
It is usually safer to send a secure payment link by SMS or WhatsApp during the call than to collect card or bank details by voice.
How long should a pilot run?
Long enough to see roll rates, which means at least one or two full billing cycles on a controlled sample of accounts, not a two-week test of call volumes.
What about borrowers who ask for a human?
Transfer them, every time, with context. A borrower who has to fight to reach a person is a complaint waiting to happen.
How we approach this at Velura Labs
Our AI Voice Call Center service follows this pattern: Hindi-first, with the compliance controls above designed in from the start, and billed per resolved call. Read our unit-economics piece to check whether your call types are candidates, and pair it with our Agentic Systems service for the back-office workflows around collections. Talk to us if your collections team is at full utilisation and the cost to serve keeps climbing.
Our clients for this span US tech hubs (San Francisco, Seattle, Austin, New York), European markets (Paris, Milan, Rome), the Middle East (Dubai, Riyadh, Abu Dhabi) and India. Start a conversation from anywhere.