AI Voice Agent for BFSI Collections with Live-Agent Handoff

Collections floors in India have converged on a two-tier model: the AI voice agent works buckets 1 to 15, sometimes 15 to 30 on softer accounts, and human collectors take over from 30+, plus every hardship, dispute, and negotiation that needs judgement. The moment the AI decides a bucket-30 borrower needs a senior collector is the moment most tools break — the AI hangs up, a queue system dials the collector separately, the borrower hears silence or a fresh ring, and the two halves of the conversation land as two separate call recordings in two separate storage buckets. For a business regulated by the RBI Fair Practices Code, that broken chain is not a UX problem. It is an audit problem.

This guide is written for collections heads, recovery-ops leaders, and the compliance officers who sign off their vendor stacks. It covers where the AI belongs in the DPD funnel, when it should hand off to a human recovery officer, what "single recording continuity" actually gets you in an RBI FPC audit, how to staff and route the human tier, and the cost math that determines whether the AI-plus-handoff model actually pays for itself.

Where the AI belongs in the collections funnel

Every collections floor we work with maps its calling strategy to DPD buckets. The AI voice agent slots in cleanly at the top of that funnel and gets progressively less useful as the account ages.

| DPD bucket | Primary handler | Why | |---|---|---| | 1 to 15 | AI voice agent | Reminder-style calls, high volume, low complexity, most PTPs are self-cure | | 15 to 30 | AI voice agent with escalation | Borrowers start negotiating; AI handles the routine, escalates the rest | | 30 to 60 | Human collector, AI as pre-call qualifier | Willingness-to-pay dropping, need judgement, RBI FPC scrutiny higher | | 60 to 90 | Senior human collector | Restructure conversations, hardship claims, legal-notice framing | | 90+ | Recovery team, legal desk | Field visit coordination, settlement negotiation, no AI |

The reason the split works is that the calls in bucket 1 to 15 are almost identical to each other. A polite Hindi or English reminder, a request for PTP, a confirmation of the amount and channel. An [AI voice agent](/ai-voice-agents) handles ten thousand of these in a shift without fatigue and without cost blowout. The moment a borrower says "I lost my job last month" or "the EMI amount is wrong", the AI is out of its depth — that call belongs with a human. What separates a competent collections floor from a bad one is how cleanly that transfer happens.

The transfer triggers that actually matter in collections

Escalation rules for the AI should be written by the collections head, not the vendor. The four triggers we see across NBFC and bank floors:

• Explicit request — the borrower says "let me speak to someone", "give me your manager", "insaan se baat karvao". No debate, transfer immediately.
• Dispute — borrower claims the amount is wrong, the EMI already got paid, or the loan account itself is disputed. Field-force verification and CRM lookup can wait; the borrower needs a human on the line now.
• Hardship claim — job loss, medical emergency, business failure. RBI FPC is unambiguous that recovery must be conducted "without harassment", and a hardship conversation with an AI is exactly the kind of interaction the ombudsman later cites.
• PTP negotiation above threshold — borrower offers to pay 40% now and the rest in three tranches. AI can accept a PTP inside pre-set bounds; anything outside those bounds is a settlement conversation, and that is a senior collector's call.

The trigger matters. What matters more is what happens in the two seconds after the trigger fires. This is where product choice becomes an audit choice.

Single recording continuity — what it means for RBI FPC audits

The RBI Fair Practices Code requires regulated entities to be able to reconstruct the full conversation with a borrower on demand — for internal audit, for the RBI banking ombudsman, for a court, or for the borrower themselves under DPDP Act 2023 access rights. Retention windows run five to seven years depending on product and internal policy.

If your AI voice agent hands off by dropping the call and having a queue re-dial the human collector, the auditor gets two files. File A is the AI portion. File B is the human portion. In the two seconds in between, nobody can say for certain what the borrower heard — dead air, a fresh ring, a hold tone, a click. If the borrower alleges that the collector used a threatening tone or misrepresented the loan terms, and the borrower's claim spans the join between the two files, the two-file evidence set becomes contested.

The dialque handoff model produces a single recording covering both the AI portion and the human portion of the same call. The borrower never gets dropped and re-dialled. The audit artefact is one file per case, timestamped end-to-end, matched to one CRM entry. Same principle we cover in the [seamless handoff deep dive](/blog/ai-voice-agent-human-transfer-single-recording-continuous-session) — for a collections desk, the compliance value is the point.

Bland, Retell, Vapi, and Synthflow typically require a separate outbound call to the human agent when the AI escalates, which produces two separate recordings and a break in the borrower's experience. That is a design choice they made; it works for many use cases; it does not fit BFSI recovery cleanly.

Setup pattern — how a collections floor actually deploys this

The floors that run this well share a common configuration.

Routing — escalations round-robin across a pool of senior collectors keyed by product (secured vs unsecured), language (Hindi, Tamil, Marathi, Telugu, Kannada, Bengali, Malayalam, Gujarati, Punjabi are the common regional ones), and DPD bucket. Bucket-30 escalations go to level-1 seniors; bucket-60 goes to a smaller pool of level-2s who handle restructure conversations.

Device mix — office collectors take the transferred call in a browser softphone from their desk. Field officers who spend the day on ground take the transferred call on their personal or company-issued mobile phone. The AI does not care which of the two the collector is on — same borrower, same session, same recording either way. This matters because collections is not a call-centre-only business; field recovery officers in tier-2 and tier-3 cities work off mobile phones by necessity.

Whisper brief — before the human collector says hello, the AI whispers a one-sentence summary into their ear: "Rakesh Sharma, personal loan account, 22 days past due, EMI ₹8,400, claims salary got delayed to end of month, asking for extension." The collector picks up already knowing the case. The borrower does not repeat their story. Handle-time drops by 30 to 40 seconds per call and — more importantly — the collector opens with something that acknowledges the borrower's situation, not a cold "how can I help".

Compliance guardrails at handoff — the AI script is DLT-registered and follows TRAI TCCCPR 2018. NDNC scrub happens pre-call. When the AI transfers to the collector, the DLT template context and consent flags travel with the call so the human collector's post-call disposition writes back to the same regulated record. One case, one audit trail.

The cost math that determines whether this pays back

The reason collections floors move to AI in the first place is the delta between a human tele-caller's cost per meaningful conversation and an AI voice agent's cost per meaningful conversation.

| Channel | Cost per meaningful conversation | Notes | |---|---|---| | Human tele-caller (metro) | ₹22 to ₹32 | Fully loaded — seat, salary, telco, supervision, attrition | | Human tele-caller (tier-2 hub) | ₹18 to ₹25 | Same math with lower salary base | | AI voice agent (dialque Growth) | ₹3.50 to ₹7 | ₹2,000 agent/month + telco + LLM, amortised across bucket-1 calls per shift |

The AI economics look transformational until you stress-test the escalation rate. If your AI escalates 40% of bucket-1 calls to a human, you are paying the AI cost plus the human cost on nearly half the conversations, and the effective cost per meaningful conversation climbs back toward ₹15 to ₹20. The saving vanishes.

Well-tuned floors keep the escalation rate under 15%. The levers:

• Tighten the escalation triggers — don't escalate on ambiguity, escalate on named categories
• Give the AI a broader PTP-acceptance range within DLT-approved templates so more negotiations resolve without a human
• Train the AI on the top 20 objection patterns your best collectors handle in bucket 1 to 15
• Route only genuine escalations to the human tier; route routine callbacks back to the AI on the next dial

The [AI vs human cost breakdown](/blog/ai-calling-bot-vs-human-agent-cost-conversion) covers the underlying math; the collections-specific overlay is the escalation-rate control. That number is where the business case lives or dies.

Case pattern — a ₹500 crore-book NBFC with 12% delinquency

A mid-book unsecured-lending NBFC we work with runs roughly 60,000 loans, average ticket ₹85,000, and a rolling delinquency of 12% in bucket 1 to 30. That is around 7,200 accounts to dial each cycle, most in bucket 1 to 15.

Before AI, the floor of 42 tele-callers made about 5,100 meaningful conversations a week at a fully loaded cost of roughly ₹1.28 lakh per week (₹25 average). Bucket-1 self-cure rate was 61%.

After moving bucket 1 to 15 to the [virtual recovery agent](/virtual-recovery-agent) and keeping the human floor for 15+ and escalations, the same volume ran on 18 human collectors plus the AI. Escalation rate settled at 11% after two months of prompt tuning. Cost per meaningful conversation dropped to ₹6.90 blended. Bucket-1 self-cure rate held at 60% — statistically flat — and the collections head redeployed 24 seats to bucket-30+ senior work, where each collector's daily recoveries went up because they were handling higher-value cases. Full breakdown in the [BFSI collections deep dive](/blog/virtual-recovery-agent-bfsi-collections).

FAQ

Does the RBI FPC prohibit AI voice agents in collections?

No. The RBI FPC governs conduct — no harassment, no intimidation, no calling outside permitted hours, no misrepresentation. Nothing in the code prevents a regulated entity from using an AI voice agent to conduct routine reminder calls, as long as the conduct rules are met and the audit trail is intact. In practice AI-run calls are more consistently compliant than human calls because the script is DLT-registered and the tone is uniform.

How does dialque handle Hindi and regional languages in collections?

Native. Hindi and English are the defaults; Tamil, Marathi, Telugu, Kannada, Bengali, Malayalam, Gujarati, and Punjabi run in production. Borrowers can code-switch mid-call — the AI stays with them. When escalation fires, routing sends the call to a human collector in the same language pool.

What are the DLT and NDNC obligations for AI-conducted collections calls?

Same as human-conducted calls. Templates registered on DLT, NDNC scrub before every dial, calling window respected (typically 9am to 6pm for collections, per your internal policy and RBI guidance). The AI enforces these automatically because they are wired into the dialler layer.

When should we not use AI in collections?

Bucket 60+, hardship conversations, settlement negotiations, disputed accounts, legal-notice-served accounts, and any case where the borrower has previously complained about tone or conduct. These belong with senior humans from the first ring.

How large a senior-collector pool do we need to sit behind the AI?

Depends on volume and escalation rate. Rule of thumb: for every 1,000 bucket-1-to-15 calls per shift at a 12% escalation rate, you need 3 to 4 senior collectors on the browser softphone or on rotation via mobile phone. Round-robin routing with a fallback pool handles peaks.

Which dialque tier fits a mid-book NBFC collections floor?

[Growth at ₹2,000 per agent per month](/pricing) covers most collections use cases — enough AI agents to run parallel dialling, the handoff feature, multilingual, DLT-integrated. Enterprise at ₹2,500 makes sense when you need dedicated compliance reporting, custom retention policies, or a private tenancy for RBI examination readiness.

---

If you run a BFSI collections floor and want to see the handoff, the single-recording audit artefact, and the whisper-brief working on your own DLT templates, [book a working demo](/contact?source=demo&topic=ai-voice-agent-bfsi-collections-live-agent-transfer) and bring three real bucket-1 and three real bucket-30 scenarios. We will run them live.