The 90-Day Window: How Banks and NBFCs Can Stop Delinquency Before It Becomes Default

By Oli AI · 2026-06-06 · 13 min

The most valuable collections work happens before an account crosses 90 days past due. Voice AI helps lenders contact borrowers earlier, identify genuine hardship, and route each case toward the right resolution path.

Once a term loan remains overdue for more than 90 days, it is generally classified as a non-performing asset under the applicable RBI prudential framework. By then, recovery is harder, the borrower relationship is more strained, and the operational cost of resolution is significantly higher.

The decisive collections work therefore happens earlier: from the first missed instalment through the 1-90 DPD window. What a bank or NBFC does during this period determines whether a temporary payment problem is resolved, converted into a workable assistance case, or allowed to roll into deeper delinquency.

This window is not only about making more calls. It is about recognising the difference between a forgotten payment, a short cash-flow gap, a disputed amount, and genuine financial hardship, then applying the right response with the right level of human authority.

Voice AI can make that operating model possible at portfolio scale. It can contact every eligible account consistently, communicate in the borrower's preferred language, capture the reason for non-payment, record a promise to pay, and escalate cases that require judgement. It should not independently approve restructuring or make promises outside the lender's policy.

1-90 DPD The intervention window where early contact, accurate classification, and timely human action can prevent avoidable roll-forward

The 1-90 DPD Window Is Not One Collections Bucket

RBI's Special Mention Account framework separates early stress into stages commonly aligned with up to 30 days overdue, more than 30 and up to 60 days, and more than 60 and up to 90 days. Applicable reporting and classification requirements depend on the regulated entity, exposure, product, and current RBI directions, but the operational lesson is universal: each stage requires a different conversation.

DPD 1-30: Resolve Friction Before It Becomes Behaviour

Many borrowers in the first month have experienced an auto-debit failure, forgotten due date, salary delay, or temporary liquidity gap. The objective is to establish right-party contact, explain the overdue accurately, provide authorised payment options, and capture a realistic payment date.

The tone should be service-led. Most accounts at this stage do not require negotiation; they require clarity and a simple next action.

DPD 31-60: Separate Avoidance, Dispute, and Hardship

By this stage, repeated reminders alone are less useful. The conversation must identify why payment has not happened. A disputed debit, medical event, income loss, business receivable delay, or deliberate avoidance should not enter the same workflow.

Voice AI can collect structured facts and create the appropriate case, while trained collections or customer-assistance staff handle decisions that involve revised terms, concessions, or formal resolution.

DPD 61-90: Prioritise Specialist Resolution

As the account approaches NPA classification, speed and accuracy matter more than call volume. AI can maintain compliant contact attempts, confirm current circumstances, and identify borrowers willing to engage.

Complex cases should move quickly to authorised specialists with the transcript, account history, stated hardship reason, and previous commitments already visible. The borrower should not have to restart the story from the beginning.

Why Human-Only Telecalling Struggles at Portfolio Scale

Human collectors are essential where judgement, negotiation, documentation review, or an emotionally sensitive decision is required. The problem is using those skilled people for every unanswered call, routine reminder, language switch, and CRM update.

At scale, queues grow faster than teams can work them. Contact timing becomes inconsistent, regional-language coverage depends on staffing, and experienced agents spend much of the day on cases that could have been resolved through a straightforward payment reminder.

• Coverage gaps — Accounts may wait days for a first meaningful conversation because human capacity is allocated to the loudest or oldest cases rather than the earliest preventable risk.
• Inconsistent classification — The same phrase, such as "salary has not come," may be recorded as a callback, a refusal, or hardship depending on the agent and the pressure of the queue.
• Language mismatch — A borrower may understand the reminder but avoid discussing a sensitive financial problem unless the conversation happens naturally in their preferred language.
• Limited quality review — Manual QA usually samples only a fraction of calls, making it difficult to identify missed disclosures, inappropriate pressure, or incorrect commitments across an entire campaign.

Detecting Genuine Financial Hardship Without Overstepping

The most important distinction in early collections is between a borrower facing a solvable payment disruption and a borrower who is unwilling to engage. That distinction cannot be made from one keyword or tone score alone.

A responsible Voice AI workflow combines what the borrower says with available account context. It can recognise phrases associated with job loss, hospitalisation, salary delay, business disruption, crop loss, bereavement, or a disputed payment. It can then ask a small number of neutral questions and create a structured hardship case.

Acoustic or sentiment indicators may help prioritise review, but they should not be treated as proof of hardship or used alone to determine eligibility. The lender's policy, verified information, and authorised decision process remain decisive.

• Confirm the borrower's identity using the lender's approved verification flow before discussing account details
• State the lender's identity, purpose of the call, overdue amount, and due date accurately
• Ask an open, non-judgmental question about what prevented payment
• Capture the stated hardship reason and expected duration without demanding unnecessary sensitive information
• Check whether the borrower can make a full payment, an affordable part-payment, or needs specialist assistance
• Create a case and schedule a callback or warm transfer when policy decisions are required
• Avoid promising approval, waiver, moratorium, credit-score protection, or revised terms before authorisation

A More Realistic NBFC Hardship Conversation

The Seven-Stage Hardship Resolution Architecture

Restructuring is not a generic product feature that should be improvised during a collections call. Any revised repayment schedule, deferment, settlement, concession, or tenure change must follow the lender's board-approved policy, applicable regulation, product terms, delegated authority, and documentation requirements.

Voice AI is most valuable when it shortens the distance between the first hardship signal and an informed decision by the right NBFC team.

Different Borrowers Need Different Resolution Playbooks

Retail and Personal Loans

Common triggers include job loss, salary delay, medical expense, and family disruption. The workflow should establish the likely duration of the disruption, current payment capacity, and whether a specialist assessment is needed.

MSME and Business Loans

Cash-flow issues may relate to delayed receivables, seasonality, inventory, or a disrupted customer contract. These cases should reach a business-credit or restructuring specialist earlier because a simple retail hardship script will not capture the operating context.

Gold Loans

Borrowers are often anxious about the pledged asset. The AI should communicate only approved account and process information, avoid speculative statements about auction timing, and escalate questions involving extension, valuation, notices, or enforcement.

Microfinance

The interaction requires careful language, privacy, and sensitivity to household or group dynamics. The AI should never disclose debt to third parties or use social pressure. Multi-lender exposure and livelihood disruption should trigger specialist review.

Compliance Must Be Built Into the Workflow

For regulated lenders, a successful collections conversation is not only one that obtains payment. It must also protect borrower dignity, privacy, data rights, and access to grievance redressal.

RBI instructions require lenders and their recovery agents to avoid intimidation, harassment, misleading representations, and intrusion into the privacy of borrowers, family members, referees, or friends. Regulated entities must also control calling hours, train agents, protect customer information, and maintain accountability for outsourced service providers.

• Contact controls: Enforce the lender's approved calling window and campaign suppression rules automatically
• Identity and disclosure: Identify the lender and purpose before discussing account information
• Privacy: Complete approved verification and never disclose dues to an unauthorised third party
• No misleading promises: Do not guarantee waiver, restructuring approval, bureau protection, or settlement terms
• Consent and data use: Apply purpose limitation, access controls, retention rules, and applicable DPDP obligations
• Grievance route: Provide a clear path to an authorised officer when the borrower disputes the account or raises a complaint
• Auditability: Retain transcript, recording status, disclosures, outcome, handoff, and case history according to policy

What a BFSI-Grade Voice AI Platform Must Deliver

A 90-Day Deployment Roadmap

1. Days 1-30: Define policy and connect data — Map DPD stages, products, languages, contact rules, hardship reason codes, grievance paths, authorised messages, and LMS/CRM integration. Decide exactly what the AI may say, record, and trigger.
2. Days 31-60: Pilot on controlled cohorts — Begin with selected early-stage accounts and a limited language set. Compare right-party contact, payment completion, promise fulfilment, complaint rate, and roll-forward against a matched human-led cohort.
3. Days 61-90: Add hardship triage and scale — Introduce validated hardship flows, specialist handoffs, QA review, and management dashboards. Expand only after policy, language, integration, and escalation behaviour meet agreed thresholds.

How Oli AI Supports the 1-90 DPD Window

Oli AI helps Indian banks and NBFCs build a controlled early-delinquency operating layer across routine reminders, payment-intent capture, hardship identification, and specialist escalation.

The platform is designed to extend the reach of collections teams without transferring credit judgement or policy authority to an unconstrained model.

• Multilingual borrower conversations across 10+ Indian languages
• DPD-specific conversation flows for early reminder, repeated delinquency, dispute, and hardship
• Real-time integration with lender systems for accurate account context and outcome updates
• Policy-controlled response logic and prohibited-promise guardrails
• Structured hardship case creation and human handoff with full context
• Call-level audit records, QA search, and compliance reporting

Frequently Asked Questions

What does 1-90 DPD mean in loan collections?

DPD means days past due. The 1-90 DPD window covers the period from the first day an instalment is overdue until the account approaches the applicable 90-day NPA threshold for many term loans. It is the key period for early intervention and resolution.

Can a Voice AI agent approve loan restructuring?

A Voice AI agent should not independently approve restructuring. It can identify possible hardship, collect borrower information, explain options already authorised for that account, record consent, and route the case to an authorised lender team for assessment and approval.

How should an NBFC handle a borrower who says they lost their job?

The lender should verify the borrower, acknowledge the situation without pressure, capture the stated hardship and likely payment capacity, avoid promising concessions, and arrange review by an authorised assistance or collections specialist under the lender's policy.

Can an NBFC promise that restructuring will not affect a borrower's credit score?

No blanket promise should be made. Credit reporting depends on the account status, the approved arrangement, applicable reporting rules, and the borrower's subsequent performance. Any impact should be explained accurately by the lender before the borrower accepts revised terms.

What should Voice AI do when a borrower disputes the overdue amount?

The AI should stop the standard payment-demand flow, record the dispute, provide a reference number, and route the case to the lender's authorised dispute or grievance process with the transcript and account context attached.

Which metrics should NBFCs track for early-stage AI collections?

Useful metrics include right-party contact, payment completion, kept promise-to-pay rate, cure rate, roll-rate reduction, specialist resolution, complaint rate, compliance exceptions, and cost per resolved account.

← Back to all articles