TL;DR
A voice agent playbook for delinquency management is a structured set of rules, scripts, escalation logic, and compliance guardrails that govern how an AI voice agent interacts with borrowers at each stage of loan delinquency. Playbooks are mapped to DPD (Days Past Due) buckets, from pre-due reminders through 90+ day collections, and they define everything from tone and language to when the AI should hand off to a human. This glossary covers every key term, metric, and concept that collections teams at banks and NBFCs need to understand before building or evaluating one.
What Is Delinquency Management?
Delinquency management refers to the strategies and processes lenders use to address overdue borrowers and prevent defaults. It covers everything from a simple payment reminder sent two days before an EMI due date to legal recovery proceedings on accounts that are months past due.
The scope is wide. At one end, you have automated SMS nudges. At the other, field agents knocking on doors. In between sits the bulk of the work: identifying which borrowers are at risk, reaching them through the right channel at the right time, offering flexible repayment options when warranted, and documenting every interaction for compliance.
Why does it matter? Because the consequences of doing it poorly are severe. As of September 2025, the GNPA ratio of Indian NBFCs stands at 2.3% according to the RBI Financial Stability Report, with microfinance and unsecured retail showing stress in the 31 to 180 DPD window. In FY2024-25, the Reserve Bank of India imposed over ₹48 crore in penalties on NBFCs and banks for violations in their collection practices. Bad delinquency management doesn’t just cost you money in write-offs. It costs you money in regulatory fines.
For a broader look at how voice AI fits into the banking picture, see our strategic guide to voice AI in Indian BFSI.
Key Sub-Terms
Days Past Due (DPD): The number of days a borrower’s payment is overdue. DPD is the fundamental unit of measurement in delinquency management. An account at 0 DPD is current. At 1 DPD, it’s one day late. RBI classifies loans as non-performing assets (NPAs) when they cross 90 DPD.
Roll Rate: The percentage of accounts that move from one DPD bucket to the next in a given period. If 15% of your 1 to 30 DPD accounts roll into 31 to 60 DPD next month, your roll rate for that bucket is 15%. High roll rates signal that your early-stage interventions are failing.
Cure Rate: The percentage of delinquent accounts that return to current status after intervention. This is the single most important metric for evaluating whether your voice agent playbooks for delinquency management are working.
NPA (Non-Performing Asset): A loan where principal or interest remains overdue for more than 90 days. Once an account hits NPA, provisioning requirements kick in and recovery costs multiply.
Write-off: When a lender removes a loan from its books after determining it’s uncollectible. This is the worst-case outcome that effective delinquency management exists to prevent.
What Is a Voice Agent Playbook?
A voice agent playbook is a structured framework that governs how an AI-powered voice agent conducts collection calls. It defines the call flow logic, compliance rules, tone of conversation, escalation triggers, and outcome-capture mechanisms for each stage of delinquency.
Think of it as the operating system for your AI collector. A human agent walks into the office with training, intuition, and a laminated script. A voice agent walks into every call with a playbook that encodes all of that, minus the intuition but plus perfect consistency.
The concept draws from how platforms like Kore.ai define playbooks in their agent-assist documentation: structured guidance that streamlines task management, where supervisors define stages and steps, choose whether progression is sequential or flexible, and set adherence notifications. In the collections context, the playbook becomes more specialized, incorporating product-aware call flows, policy-gated compliance, and bucket-specific behavior.
A voice agent playbook is not the same as a legacy IVR script. IVR scripts are rigid, menu-driven, and push-button. They don’t listen, adapt, or capture nuanced outcomes. A voice agent playbook powers a conversational AI that can understand natural speech, respond in the borrower’s preferred language (including mixed-language patterns like Hinglish), and make real-time decisions about what to say next.
It’s also not a copy of your human call script. Practitioners in the Indian NBFC space have flagged this as a common mistake. As one practitioner-focused playbook from Caller Digital puts it: “Human scripts are written for human constraints. Voice AI can use a cleaner, more conversational flow and will perform worse if you force-fit the legacy script.”
Book a demo with Awaaz AI to see how DPD-specific playbooks work in practice.
Components of a Voice Agent Playbook
Every effective playbook for delinquency management contains these building blocks:
Call Flow Logic
The sequence of conversational steps the agent follows. This includes the opening greeting, identity verification, statement of purpose, payment discussion, objection handling, and closing. The flow branches based on borrower responses. If the borrower says they’ve already paid, the agent follows a different path than if they say they can’t pay.
Identity Verification (Right-Party Contact)
Before discussing any debt details, the agent must confirm it’s speaking to the actual borrower. RBI compliance requires identity disclosure within 30 seconds: the recovery agent (human or synthetic) must state its name, the lender’s name, and the purpose of the call. RBI inspectors test this on randomly pulled call recordings.
Compliance Guardrails
These are hard-coded rules that the playbook enforces without exception. RBI’s Fair Practices Code restricts collection calls to 8 a.m. to 7 p.m. A voice AI that calls at 7:45 p.m. because the dialer queue ran long has just generated a regulatory event. Guardrails also cover required disclosure language, prohibitions on abusive language, and data handling rules under the DPDP Act. For a deeper look at the regulatory requirements, our guide on automated calls, laws, and AI covers the specifics.
Tone and Language Rules
The playbook specifies the emotional register of the conversation, which shifts by DPD bucket. Pre-due calls are friendly reminders. Bucket 1 calls carry a note of urgency about CIBIL impact. Bucket 2+ calls become firmer. Language rules also define which languages the agent can converse in and how it handles code-switching between languages.
Escalation Triggers
Conditions that route the call from AI to a human agent. Common triggers include: the borrower mentions financial hardship, disputes the debt, becomes hostile, requests to speak with a supervisor, or raises a legal concern. The playbook defines exactly which keywords or sentiment signals activate each trigger.
Outcome Capture
After every call, the agent logs a structured outcome. This typically includes disposition codes (paid, promise-to-pay, refused, hardship, dispute, no-answer), PTP details with dates and amounts, and any follow-up actions required. These outcomes write directly to the LMS and CRM.
QA and Adherence Tracking
The playbook includes checkpoints that supervisors can audit. Did the agent disclose identity within 30 seconds? Did it stay within calling hours? Did it capture a valid PTP? Adherence tracking turns every call into auditable data.
DPD Buckets and Playbook Mapping
Voice agent playbooks for delinquency management are not one-size-fits-all. The most effective implementations treat DPD as a first-class field, with each bucket getting its own playbook variant that differs in tone, urgency, actions, and escalation depth.
Here’s how playbooks should map to each bucket:
Pre-Due (Before EMI Date)
What the playbook does: Sends a friendly reminder about the upcoming payment. The voice agent informs the borrower of the due date, confirms the auto-debit mandate is active, and offers to help set up NACH if the borrower is still paying manually.
AI autonomy: Fully autonomous. No human involvement needed.
Key metrics: Pickup rate, auto-debit conversion rate.
For more on building effective reminder flows, see our guide on automated payment reminder software.
Bucket 0 (EMI Due Date)
What the playbook does: On the last day of payment, the voice agent reminds the borrower to ensure sufficient balance for auto-debit. It can also assist in converting manual payment to auto-debit, reducing future delinquency risk.
AI autonomy: Fully autonomous.
Key metrics: Same-day payment rate, NACH setup conversions.
Bucket 1 (1 to 30 DPD)
This is the highest-ROI window for voice AI in delinquency management.
What the playbook does: The agent contacts the borrower with moderate urgency. It explains that the EMI is overdue, notes that continued non-payment will impact the borrower’s CIBIL score, and captures a promise-to-pay with a specific date and amount. It can send a payment link via SMS or WhatsApp during the call.
AI autonomy: Largely autonomous. Escalation only for disputes or hardship declarations.
Key metrics: RPC rate, PTP conversion rate, cure rate.
The math here is straightforward. Indian retail credit stress is concentrated in early DPD buckets, 1 to 30 and 31 to 60 days past due, where the vast majority of accounts will self-cure if they get the right reminder in the right language at the right time. Miss that window and the account slides into harder buckets where recovery cost multiplies and write-off risk starts to matter.
Financial services companies using AI calling for soft collections typically see a 20 to 35% improvement in recovery rates in the 30 to 60 DPD bucket, with 40 to 60% reduction in cost per rupee recovered compared to human-only teams.
Bucket 2 (31 to 60 DPD)
What the playbook does: Multi-channel escalation kicks in. The playbook orchestrates outbound voice calls alongside SMS and WhatsApp follow-ups. NACH retry attempts are scheduled against the borrower’s salary credit pattern, not an arbitrary calendar date. A telecalling case is auto-assigned based on loan value and borrower risk segment.
AI autonomy: Partial. AI handles initial contact and PTP capture, but complex negotiations and restructuring discussions route to human agents.
Key metrics: Roll rate (are accounts moving to Bucket 3?), PTP fulfillment rate, cost per recovered rupee.
Bucket 3+ (61 to 90+ DPD)
What the playbook does: At this stage, conversations require more contextual depth and negotiation skill. The voice agent may still initiate contact, but the playbook is configured for rapid escalation to trained human collectors who can discuss settlement options, restructuring, or legal next steps.
AI autonomy: Limited. AI handles initial outreach and data gathering, but substantive conversations are human-led.
Key metrics: Recovery rate, settlement acceptance rate, cost per case.
Practitioners consistently recommend against starting voice AI pilots in this bucket. As one practitioner-oriented guide puts it: “Do not pilot on 60+ DPD. Pilot on pre-due. Prove the engine, then move down the funnel.”
Core Terms Inside a Voice Agent Playbook
This mini-glossary covers the terms you’ll encounter when building, evaluating, or operating voice agent playbooks for delinquency management.
Right-Party Contact (RPC)
Confirming that the person on the line is the actual borrower before discussing any debt details. RPC rate, the percentage of calls where you successfully reach and verify the borrower, is a foundational metric. Industry data shows that only 20% of manual collection calls lead to a successful outcome, and low RPC rates are a major reason.
Promise-to-Pay (PTP)
A verbal or digital commitment from the borrower to pay a specific amount by a specific date. Voice agents capture PTPs in structured format and write them to the LMS. PTP conversion rate, the percentage of connected calls that result in a payment commitment, is a primary playbook performance metric. NBFCs using voice AI report 22 to 37% higher PTP rates during early-stage collections.
Disposition Code
A standardized label applied to each call outcome. Common codes include: Paid, PTP (with date), Refused, Hardship, Dispute, No Answer, Wrong Number, and Callback Requested. Consistent disposition coding is what makes portfolio-level analytics possible.
Escalation Trigger
A predefined condition that transfers the call from the AI agent to a human. Triggers can be keyword-based (borrower says “lawyer,” “harassment,” or “I lost my job”), sentiment-based (rising anger or distress detected), or rule-based (account value exceeds a threshold, or borrower has filed a previous complaint).
NACH Retry Logic
NACH (National Automated Clearing House) is India’s auto-debit system. Retry logic determines when and how many times the system re-attempts a failed auto-debit. Smart playbooks schedule retries based on salary credit patterns rather than fixed calendar dates, significantly improving success rates.
Call Cadence
The frequency and timing rules governing how often an account is contacted. This includes minimum gaps between call attempts, maximum attempts per day or week, and time-of-day preferences based on historical pickup data. RBI’s Fair Practices Code places hard limits on cadence, and playbooks must enforce them automatically.
Sentiment Detection
Real-time analysis of the borrower’s vocal tone, word choice, and speech patterns to gauge emotional state. Sentiment detection feeds into escalation logic: if the system detects distress, anger, or confusion beyond a threshold, it triggers a human handoff. This is one area where voice agent playbooks differ fundamentally from text-based collections.
Compliance Score
The percentage of calls in a given period that pass all QA checkpoints. A compliance score of 98% means 2% of calls had at least one violation, whether it was a late call, a missed disclosure, or an improper statement. For voice AI, this metric should approach 100% by design, since compliance guardrails are hard-coded into the playbook rather than left to agent memory.
How Voice Agent Playbooks Work in Practice
Understanding the theory is one thing. Here’s how voice agent playbooks for delinquency management operate in a real integration stack.
The Integration Stack
A production deployment connects several systems:
- Loan Management System (LMS): The source of truth for account data, including DPD status, outstanding amount, payment history, and borrower contact details.
- Voice AI Platform: Executes the playbook, makes calls, handles conversation, and captures outcomes.
- CRM: Stores interaction history and manages follow-up workflows.
- Payment Gateway: Enables real-time payment link generation and delivery during calls.
- QA/Analytics Layer: Records calls, scores compliance adherence, and surfaces performance metrics.
For details on how automated reminder calls integrate with compliance requirements, see our dedicated guide.
A Walkthrough: Day 1 Through Day 30
Day -2 (Pre-due): The LMS flags the account as approaching its EMI date. The voice agent calls with a friendly reminder, confirms the auto-debit mandate, and wishes the borrower well.
Day 0 (Due date): If payment hasn’t cleared by midday, the agent calls to remind the borrower to maintain sufficient balance for the evening’s NACH sweep.
Day 1 (1 DPD): Payment missed. The playbook shifts tone from friendly to informative. The agent calls, verifies identity, informs the borrower that the EMI is overdue, and asks when they can pay. If the borrower commits, the agent captures a PTP, sends a payment link via SMS, and sets a follow-up timer.
Day 5 (5 DPD): If the PTP was not fulfilled, the agent calls again. Tone is slightly firmer. It references the previous commitment and asks for an updated payment date. It mentions that continued non-payment may affect the borrower’s credit score.
Day 15 (15 DPD): If still unpaid, the playbook triggers multi-channel outreach. The voice agent calls, a WhatsApp message is sent with the payment link, and the system schedules a NACH retry aligned to the borrower’s expected salary credit date.
Day 25 (25 DPD): Final AI-led attempt before the account approaches Bucket 2. The agent’s tone conveys urgency. If the borrower expresses genuine hardship, the system escalates to a human agent who can discuss restructuring options.
The Compliance Layer
Every step above is governed by RBI’s Fair Practices Code and the Digital Personal Data Protection Act. The playbook enforces:
- Calling hours restricted to 8 a.m. to 7 p.m.
- Identity disclosure within 30 seconds
- No threats, abusive language, or misleading statements
- Call recording retention for a minimum of 90 days (often 180 days under internal policy, sometimes 7 years for litigation-track loans)
- Data handling rules under DPDP, including consent management and data residency requirements
For a comprehensive overview of AI debt collection calls including recovery strategies and compliance, our detailed guide covers the full picture.
Voice Agent Playbooks vs. Human Call Scripts
The distinction matters because many teams approach their first voice AI deployment by copying their existing human scripts into the AI system. This is a mistake.
Where Voice Agent Playbooks Win
Consistency. A human agent has good days and bad days. They skip disclosures when rushed, lose patience with difficult borrowers, and forget to log dispositions accurately. Voice agent playbooks execute identically on the first call of the day and the ten-thousandth.
Scale. Consider the math: an NBFC with a 2-lakh borrower portfolio and 15% delinquency rate has 30,000 accounts to chase monthly. A 50-person team handles roughly 600 calls each. That’s 30,000 calls, each a potential compliance violation. A voice AI handles this volume without hiring delays or attrition.
Compliance. You can train human agents, monitor their calls, run quality audits. But at scale, human inconsistency is structural, not fixable. A playbook’s compliance guardrails are architectural. They can’t be skipped because an agent is having a bad shift.
Borrower psychology. This is counterintuitive but well-documented by practitioners. As one collections-focused AI provider notes: “Debt is embarrassing. When you call someone to ask for money, their heart rate goes up. They feel judged, which leads to avoidance behavior.” When using AI, borrowers get zero judgment and privacy. They can negotiate at 11 p.m. on a Sunday from their couch, without anyone overhearing. This actually increases right-party contact rates and PTP percentages.
Where Human Scripts Still Win
Complex negotiation. When a borrower needs a restructured payment plan, is dealing with a genuine financial crisis, or is considering legal action, human agents bring empathy and judgment that AI can’t replicate yet.
The practical answer is a hybrid model. AI handles Bucket 0 and Bucket 1 volume, where the conversations are relatively structured and the ROI is highest. Humans focus on Bucket 2+ and dispute resolution, where contextual understanding matters most.
For more on how conversational AI transforms contact centers, including the human-AI handoff dynamic, see our complete guide.
Common Mistakes When Building a Delinquency Playbook
Practitioners who have deployed voice agent playbooks for delinquency management across Indian NBFCs consistently flag the same errors.
Starting With the Hardest Bucket
The instinct is to throw AI at your biggest problem, which is usually the 60+ DPD accounts that are burning the most money. This is backwards. Those accounts need nuanced, human-intensive conversations. Start with pre-due or 5 to 15 DPD reminder calls. They’re impactful, carry bounded risk, and are easy to measure. Prove the engine works, then expand down the funnel.
Copying Human Scripts Verbatim
Human scripts are written for human constraints: they account for agents who need to read from a screen, who might forget steps, who need memory aids. Voice AI can use cleaner, more conversational flows. Force-fitting a legacy script into an AI system produces awkward, unnatural interactions that perform worse than a purpose-built conversational design.
Optimizing Per-Minute Cost Instead of Cost Per Recovered Rupee
A cheaper bot that sounds robotic is more expensive than a slightly pricier bot that actually collects. The metric that matters is cost per recovered rupee, not cost per minute of talk time. Skit.ai reports 49% of total collection value recovery per campaign and debt recovery from 78% of delinquent accounts without human assistance. A bot that achieves those numbers at ₹3 per minute is far cheaper than one at ₹1.50 per minute that recovers half as much.
Ignoring Language and Code-Switching Needs
India’s borrower base doesn’t speak in clean, single-language sentences. A borrower in Maharashtra might start in Marathi, switch to Hindi mid-sentence, and throw in English financial terms. A playbook that only handles standard Hindi will miss context, frustrate borrowers, and tank pickup rates.
Treating Compliance as a Feature Checkbox
Compliance isn’t a feature you toggle on. It’s an architectural requirement. The playbook must enforce calling hours, disclosure rules, and data handling at the infrastructure level, not through post-hoc QA that catches violations after they’ve already happened. With RBI having imposed over ₹48 crore in penalties in a single year, this isn’t a theoretical concern.
Metrics That Matter
When evaluating whether your voice agent playbooks for delinquency management are performing, track these numbers:
Recovery Rate by DPD Band: What percentage of delinquent accounts in each bucket return to current status? This is the ultimate measure of playbook effectiveness. AI-led collections can cut operating costs by around 40% while improving recoveries, according to a McKinsey-linked analysis.
RPC Rate: What percentage of calls reach the actual borrower? Low RPC rates indicate problems with contact data, calling times, or caller ID trust.
PTP Conversion Rate: Of calls where you reach the borrower, what percentage result in a payment commitment? This measures the playbook’s persuasive effectiveness.
PTP Fulfillment Rate: Of borrowers who made a promise-to-pay, what percentage actually paid? Low fulfillment rates may indicate the AI is capturing weak commitments or not following up effectively.
Cost Per Recovered Rupee: Total collection operation cost divided by total amount recovered. This is the efficiency metric that matters more than per-minute call costs.
Compliance Score: Percentage of calls passing all QA checkpoints. For AI-powered playbooks, this should be near 100%.
Roll Rate Reduction: How much has the percentage of accounts rolling from one bucket to the next decreased since playbook deployment? This measures whether you’re actually bending the delinquency curve.
The market for these solutions is growing fast. The AI debt collection market is projected to reach $15.9 billion by 2034, growing at 17% CAGR from $3.3 billion in 2024. That growth reflects the measurable impact these systems deliver.
Learn more about how Awaaz AI’s multilingual voice agents help NBFCs reduce delinquency across DPD buckets.
Pilot Strategy for Voice AI in Collections
For teams building their first voice agent playbook for delinquency management, practitioners recommend a structured approach.
Define your thin slice. Start with broken PTP follow-ups or 5 to 15 DPD reminder calls. These are impactful, carry limited risk, and produce measurable results quickly.
Map policy to dialog. Disclosure text, hardship triggers, inconvenient-time checks, and voicemail content must be explicit in the playbook before the first call goes out.
Run a bounded pilot. Pick one DPD band or one queue. Instrument the KPIs you care about. Run a 6 to 10 week pilot with daily QA review, then expand what works.
Compare against your baseline. Measure AI performance against your existing human team on the same account segment. Without a controlled comparison, you can’t distinguish AI impact from seasonal variation or portfolio mix changes.
For a step-by-step deployment plan, our voice AI pilot checklist for NBFCs walks through the 30/60/90 day timeline.
Frequently Asked Questions
What exactly is a voice agent playbook for delinquency management?
It’s a structured framework that defines how an AI voice agent conducts collection calls at each stage of loan delinquency. It includes call flow logic, compliance guardrails, escalation triggers, tone rules, and outcome-capture mechanisms, all mapped to specific DPD buckets.
How is a voice agent playbook different from a regular IVR script?
IVR scripts are rigid, menu-driven, and rely on keypad input. Voice agent playbooks power conversational AI that understands natural speech, adapts responses in real time, handles multiple languages including code-switching, and captures structured outcomes. The difference is between a phone tree and an actual conversation.
Which DPD bucket should we start with when deploying voice AI?
Pre-due or Bucket 1 (1 to 30 DPD). These accounts have the highest self-cure probability, the conversations are relatively structured, and the compliance risk is lowest. Practitioners consistently warn against piloting on 60+ DPD accounts.
Can voice AI handle the full collections process without humans?
No. Voice AI is highly effective at Bucket 0 and Bucket 1 with no human assistance. Later buckets require more contextual conversation, negotiation, and dunning from human collectors. The best model is hybrid: AI handles volume in early buckets, humans focus on complex cases.
What compliance rules must a voice agent playbook enforce in India?
At minimum: calling hours restricted to 8 a.m. to 7 p.m. (RBI FPC), identity disclosure within 30 seconds, no abusive or misleading language, call recording retention for at least 90 days, and data handling per the DPDP Act. These must be hard-coded into the playbook, not left to post-hoc QA.
What metrics should we track to measure playbook performance?
The most important are: cure rate by DPD band, PTP conversion rate, PTP fulfillment rate, cost per recovered rupee, RPC rate, compliance score, and roll rate reduction. Cost per minute is a secondary metric, not a primary one.
How long does a voice AI pilot for collections typically take?
Most practitioners recommend 6 to 10 weeks for a properly instrumented pilot on a single DPD band or queue, with daily QA review. This gives enough data to measure impact while keeping risk bounded.
Does the borrower know they’re speaking to an AI?
Regulatory requirements vary, but transparency is generally recommended and increasingly required. The playbook should include a disclosure that the call is AI-assisted. Interestingly, this doesn’t hurt performance. Borrowers often prefer the privacy and non-judgmental nature of AI interactions, which can actually improve RPC and PTP rates.