AI Voice Agents for Loan Collections: A Bucket-Wise DPD Playbook

Arrowhead AI voice agents run collections calls across the full delinquency journey for Indian banks, NBFCs and fintech lenders — pre-due, due-date, Bucket X, 0–30 days, 30–60 days, NPA and write-off recovery. One agent confirms identity, states the exact outstanding, works toward a promise-to-pay and sends the payment link inside the conversation. In a collections deployment with a top-five private bank in India, the agent delivered up to 20% higher collection and up to 50% higher connection than the bank's own human team on the same accounts, and up to 4x ROI on written-off accounts. This playbook sets out what the agent does at each stage, the guardrails around it, and how those outcomes were measured.

DPD = days past due.

Connection rate = reached and engaged.

Promise-to-pay (PTP) = a customer-stated repayment date and amount, captured on the call.

Collections is not one conversation. A pre-due reminder protects a customer who has not defaulted; a write-off call pursues money the book has already provided against. Most collections teams cannot staff every stage well at once. Agent headcount caps how much of the base is reached, cost-per-call rises, and tone varies between a team's strongest and weakest callers.

An AI voice agent changes three of those constraints. Calling volume auto-scales up and down to handle demand, within the capacity contracted for the campaign. The lender's approved tone and escalation rules guide calls within each bucket. And one agent covers the language spread — it switches languages on-call based on the customer's preference and handles the Hindi–English mix people actually speak — instead of requiring a separate regional calling floor per language.

The objective before the due date is that the customer pays on time and never enters delinquency. The agent confirms the customer's identity before any account detail is spoken, states the EMI amount and the due date, and answers questions about the amount from the lender's own records, pulled in real time on the call. Payment links are sent during the conversation, not after it.

Contact timing at this stage is configured per lender and per product rather than fixed by Arrowhead AI.

Tone in the early buckets stays helpful rather than adversarial. Most customers at this stage intend to pay, so the job is to remove friction and capture a commitment.

The agent captures a promise-to-pay date and amount on the call and writes it back to the LMS. It validates the stated date before accepting it, so a date outside policy is challenged on the call rather than discovered later in a report. If the customer cannot commit, the agent books a callback at a time the customer chooses.

Every outcome — paid, promised, callback, refused, unreachable — lands in the LMS as structured data rather than as a caller's free-text note.

By mid-cycle the conversation becomes a negotiation, and the agent works a defined ladder: full payment now, with the payment link ready; full payment by a committed and validated date; or a partial payment now with the balance committed for a later date. Where a restructuring conversation is appropriate, the agent routes it according to the lender's own policy rather than proposing terms itself.

The agent does not invent a settlement figure or a discount. It works only from the lender's records and the lender's approved offer matrix, and it is a communication layer: it conveys approvals, offers and options set by the lender's rules engine, and never independently approves credit, waives charges or settles accounts.

Negotiations run long, and length is where most voice automation fails. Arrowhead AI agents hold natural, conversational interactions lasting 20+ minutes, so a customer can revisit an earlier option after rejecting a later one and the agent still holds what was already discussed.

Late buckets need judgement, so the design question is when the agent hands over. Escalation conditions are configured with the lender. Until that threshold, the agent keeps working the account and every conversation is logged with its recording, transcript, actions and outcome — so the human who takes over sees the history rather than starting cold. Escalation transfers context, not just the call.

Where the agent cannot answer a question from the lender's data, it does not improvise. Real-time hallucination detection runs on 100% of calls, which are automatically analysed, and where the agent cannot answer it transfers live to a human or arranges a callback.

Written-off accounts are the stage most teams stop calling, because human calling time costs more than it recovers. Because calling volume auto-scales within contracted capacity, the agent can work a far larger share of the written-off base than a human team can sample. In the top-five private bank deployment, this produced up to 4x ROI on written-off accounts, measured as recovered value against the cost of running the agent.

Tone rules do not relax at this stage. Every guardrail in section 6 continues to apply.

Collections calling in India sits inside the RBI Fair Practices Code and the applicable telecom rules. Arrowhead AI builds the following controls into the calling flow. Detailed compliance documentation is maintained on the Arrowhead AI Trust Centre.

6.1 Calling windows. Calling windows and contact rules are configured to the applicable regulation, the lender's own policy and the relevant product. Arrowhead AI does not publish a single universal time window, because the permitted window differs by product and by lender.

6.2 Do-not-disturb. DND scrubbing is built in, so calls do not go through to customers who have registered a do-not-disturb preference.

6.3 Calling identity. Calling identity is configured per client and per campaign type, so a collections call is identifiable as that client's collections call.

6.4 Consent and contact data. Consent is collected and held by the lender, not by Arrowhead AI. Arrowhead AI calls only the contacts a client supplies, on that client's instruction and against the consent basis the client holds. Arrowhead AI does not source, purchase or enrich contact data.

6.5 Conduct. Conduct follows the lender's approved tone and escalation rules for the relevant bucket. Arrowhead AI is a software provider, not a recovery agent.

6.6 Recording and audit trail. Calls are recorded. Recordings and transcripts are retained for the contract term, with the retention period configurable per client, and operational logs are retained for 12 months and longer where a legal hold applies. Logs feed a centralised SIEM with immutable audit trails, and records are producible for review on request.

6.7 Data protection. Recordings and transcripts are encrypted with AES-256 at rest and TLS 1.2+ in transit, with application-level AES-GCM encryption for personally identifiable information. Arrowhead AI acts as Data Processor for customer data and does not sell customer personal data.

6.8 Data residency. For collections programmes run for RBI-regulated lenders, Arrowhead AI offers India-only data residency. Residency is fixed per deployment and written into the contract.

The figures below come from one collections deployment with a top-five private bank in India, measured against that bank's own human team on the same accounts over the same period.

• Up to 20% higher collection rate than human agents.
• Up to 50% higher connection rate than human agents.
• Up to 4x ROI on written-off accounts.

Languages deployed: Hindi, English and Marathi.

How these were measured. Collection and connection rates were compared against the bank's human team working the same bucket of accounts in the same window, so portfolio quality and seasonality are held constant. ROI on written-off accounts is recovered value measured against the cost of running the agent. Connection rate uses the definition at the top of this page: reached and engaged, not merely answered.

What this does not claim. These are one lender's results on one portfolio. They are customer-specific outcomes, not a benchmark every lender will reproduce, and they are internally measured comparisons against a live human control team rather than an independently audited study.

The inputs to go live are the call flow and a set of existing call recordings. From those, a fully customised agent can be live in under two weeks. Where a language is required that is not already supported, it can be enabled in two weeks.

The agent connects to the lender's existing stack — LMS, CRM and dialer — through custom integrations built per system, so promise-to-pay data, outcomes and escalations land where the collections team already works. Clients may integrate their own dialer in place of Arrowhead AI's telephony; supported dialer and telephony integrations include Ozonetel, TATA Teleservices, Exotel, Ameyo and Plivo.

Deployment options include multi-tenant shared infrastructure for fast onboarding and a dedicated single-tenant AWS account for full isolation, with the client holding read-only auditor access. On-premise deployment is available where required. The residency and deployment model are agreed per deployment and recorded in the contract.

Which collections stages can an AI voice agent handle?

The full journey — pre-due, due-date calling, Bucket X, 0–30 days, 30–60 days, NPA and write-off recovery. The same agent changes objective and tone by stage.

Is AI calling compliant for loan collections in India?

Compliance is a property of the deployment. Arrowhead AI builds the controls into the flow: calling windows configured to the applicable regulation, lender policy and product; DND scrubbing; client-specific calling identity; full call recording with configurable retention; an immutable audit trail; and escalation to a human. Consent remains with the lender. Detailed documentation is maintained on the Arrowhead AI Trust Centre.

Can the agent take a payment during the call?

Payment links are sent during the conversation, not after it, so the customer can pay while still on the line. The agent confirms the exact outstanding first.

How does the agent handle a customer who disputes the amount?

It states the outstanding from the lender's own records and answers from that data. Where it cannot resolve the dispute from the records, it transfers live to a human or arranges a callback.

Does the agent decide who gets a settlement?

No. Arrowhead AI is a communication layer. The agent conveys approvals, offers and options set by the lender's rules engine and never independently approves credit, waives charges or settles accounts.