Sending a WhatsApp message to a borrower who ignores WhatsApp and always answers voice calls is not a soft reminder — it is a missed contact. Calling a borrower at 10 AM who has never answered before noon is not diligent collections — it is wasted attempts. The Early Bucket Caller AI selects channel, timing, and frequency per borrower from the first contact — and learns from every response to refine those selections through the DPD cycle.
Why channel choice is a collections variable, not an IT preference
The instinct to standardise channel — "we send WhatsApp to everyone" — treats collections as a broadcast operation. It is not. It is a communication problem. The question is not which channel reaches the most borrowers in aggregate — it is which channel reaches each borrower reliably, in a mode they are comfortable engaging with, at a time when they are able to respond.
A salaried professional in Mumbai who commutes for 90 minutes is reachable by WhatsApp during the commute and unreachable by voice during that same window. A first-generation borrower in a Tier 2 city who uses WhatsApp primarily for family messaging may not engage with a lending institution on that channel but will answer a voice call. A borrower who has set their number to DND between 9 PM and 9 AM will never receive an IVR that goes out at 8:30 PM regardless of the message quality.
The Early Bucket Caller AI builds a channel preference model for every borrower in the DPD 0–30 book — drawing from historical engagement data, borrower profile signals, and real-time response behaviour — and routes each contact through the highest-probability channel at the optimal time.
The four channels — and what each is optimised for
| Channel | Optimal Borrower Profile | Optimal Timing | Response Window | Best For | Limitations |
|---|---|---|---|---|---|
| Urban, smartphone-native, WhatsApp Business opt-in, prior WA engagement history | 8–10 AM or 7–9 PM — commute / evening leisure window | Average 34 minutes to read · 2.8 hours to reply | DPD 1–7 first contact, PTP confirmation, payment links, pre-reminder | Requires opt-in · ineffective for non-smartphone users · no reply if WA not used for finance | |
| Outbound Voice | Salaried borrowers with regular pick-up patterns, Tier 2 / Tier 3 borrowers, prior voice engagement, DPD 7+ | 11 AM–1 PM or 4–7 PM — post-commute, pre-evening windows | Immediate if answered · 3 attempts before fallback | DPD 7+ first voice contact, PTP capture, broken PTP follow-up, hardship escalation | DND restrictions · voicemail drop-off · agent availability for human escalation |
| IVR | Borrowers with high IVR historical engagement, simple reminder confirmations, PTP reminder at committed amount/date | 10 AM–12 PM · Avoid evenings (low IVR engagement) and Monday mornings | Immediate response during call · no async engagement | Pre-committed reminder, account status check self-service, simple PTP confirmation | Low engagement rate vs voice for new contacts · no nuance · no hardship detection |
| SMS | Borrowers without WhatsApp opt-in, feature phone users, pure notification use cases, DND-cleared short codes | 9–11 AM or 5–7 PM | High open rate · near-zero reply rate · one-directional | Payment link delivery, PTP confirmation fallback, regulatory disclosure | No engagement or conversation capability · click-through rate low · regulatory character limits |
The channel selection model: how the AI decides per borrower
The Early Bucket Caller AI's channel selection model draws from five data sources simultaneously to determine the highest-probability channel for each borrower at each contact moment.
Prior channel engagement history
Has this borrower responded to WA, voice, IVR, or SMS in prior contacts — at this institution or via consortium data?
The single strongest predictor of future channel engagement is past channel engagement. A borrower who has answered three voice calls in prior interactions will answer a fourth. A borrower who has clicked every WhatsApp payment link without ever answering a voice call will not suddenly become a voice call responder.
Borrower profile — app, channel, and device
Did this borrower apply via mobile app, WhatsApp, or DSA? What device type? Regional language preference?
App-native borrowers on Android or iOS are almost always WhatsApp-reachable. DSA-acquired borrowers in Tier 2 cities are more likely to be voice-responsive. A borrower who interacted in Hindi throughout onboarding is less likely to engage with English-language WA messages but more likely to engage with a voice call in Hindi.
DND status and time-of-day patterns
Is the number on DND? What time zone? What are the registered available calling hours?
DND-registered numbers cannot receive promotional voice calls but can receive transactional calls. The AI checks TRAI DND status before every voice attempt. Time-of-day analysis from prior call answer patterns identifies each borrower's personal peak-engagement window — some borrowers answer consistently only between 12 and 2 PM; others respond to evening WA messages only.
Real-time session signals
Has the WA message been delivered? Read? Did the payment link get clicked without payment following?
WhatsApp Business API provides message delivery and read receipts. A message that has been read but not acted on within 4 hours triggers an escalation to voice — the borrower has seen the message and has not responded, which is a different situation from a borrower who has not received any contact. Link click without payment is a high-intent signal that a voice call would address a friction point in the payment path.
Segment-level channel benchmarks
For borrowers with no prior engagement history, what channel performs best for this borrower profile cluster?
For first-time delinquent borrowers with no prior engagement data to draw from, the AI falls back to segment-level benchmarks: what is the best-performing channel for borrowers with this product, geography, onboarding channel, and borrower tenure combination? These benchmarks are updated monthly from portfolio-level channel performance data.
Channel performance by borrower segment — this month's data
| Borrower Segment | WA Engagement Rate | Voice Engagement Rate | IVR Engagement Rate | AI Primary Channel | PTP Capture Rate (primary channel) |
|---|---|---|---|---|---|
| Urban salaried · Metro · App-acquired | 68% | 44% | 22% | 61% | |
| Salaried · Tier 2 · DSA-acquired | 34% | 72% | 18% | Voice | 74% |
| Self-employed · Business loan · Any city | 41% | 66% | 14% | Voice (business hours) | 68% |
| Prior WA engagement — clicked links | 79% | 38% | 20% | 66% | |
| Prior voice answer — call history | 28% | 80% | 26% | Voice | 79% |
| No prior engagement · First-time delinquent | 44% | 52% | 18% | WA first, voice if unread | 54% |
| WA read but no response — link clicked | Read: 100% | Escalate to voice | — | Voice (friction resolution) | 71% |
The right channel is not the cheapest channel — it is the one the borrower will respond to
WhatsApp messages cost a fraction of outbound voice calls. An institution that sends every early bucket contact by WhatsApp because it is cheaper is optimising for cost and producing a lower-quality collections outcome. The voice call that captures a promise-to-pay that WhatsApp would never have produced is not an expensive contact — it is a revenue recovery action that the cheaper channel would not have completed. The Early Bucket Caller AI optimises for outcome, not cost-per-contact. The cost-per-recovered-rupee is always lower when the right channel is used, even when that channel costs more per attempt.