Voice AI for Indian Banks & NBFCs 2026: Vendor Selection Framework (Gnani, Verloop, Nurix, Caller Digital Compared)

A Head of Collections at a Pune-headquartered NBFC opened her week with a one-line note from the Risk Committee. The note read: the auditors are scheduled in eight weeks; the voice AI deployment in collections needs to pass the audit, not just produce numbers. The number side was working — 23% lift on 30–60 DPD recovery, ₹68 average cost per recovered EMI, supervisor headcount down 40%. The audit side was unclear. The DLT principal-entity ID was being captured but the consent purpose-flag was not surfaced in the audit export. The recording disclosure was in the script but the timestamp was not in the trail. The legal recovery escalation logged the supervisor handoff but not the audio context handoff. None of these were operational problems. All of them were audit problems, and the cost of getting them wrong was no longer measured in basis points — it was measured in RBI enforcement actions.

That risk-committee note is what every Indian BFSI buyer is feeling in 2026. Voice AI works. The lift is real. The question is no longer whether to deploy; it is whether the deployment will survive the regulator. The vendor selection for an Indian bank or NBFC is not the same problem as the vendor selection for a D2C brand. The selection rubric is heavier on the audit trail, the consent capture, the recording disclosure, the legal recovery workflow, the DPDP data residency, the BBPS and UPI integration, the RBI Fair Practices Code Para 7 controls. Most of the public listicles on this category skip these layers entirely.

This post is the framework. We will work through the 2026 BFSI regulatory grid in production-implementation detail, score four shortlisted vendors against twenty BFSI-specific requirements (Caller Digital, Gnani.ai, Verloop.io, Nurix AI), unpack the cost-per-recovered-EMI math, and lay out a 60-day procurement-to-production calendar that satisfies an Indian RBI / IRDAI audit committee. If you are running collections, credit card operations, loan recovery, KYC reminders, account servicing, or insurance renewal calls at scale in India in 2026, this is the rubric to take into your next steering review.

The BFSI regulatory grid for voice AI in 2026 — what the auditor will check

A voice AI deployment in an Indian bank, NBFC, insurer, or fintech has to satisfy four regulatory regimes simultaneously. Each has specific production-implementation requirements that show up in the audit trail.

RBI Fair Practices Code on collection calls (Para 7 detail)

Required production controls:

• Call-time window enforcement. No collection calls before 8am or after 7pm IST. Automatic enforcement, not advisory. The audit trail must capture the dial timestamp and confirm the window for every call.
• Language disclosure. The borrower must be informed in their preferred language. The language must be on file in the borrower master and the audit trail must confirm the language used per call.
• Recording disclosure. The fact that the call is being recorded must be disclosed within the opening utterance, before the first ask. The disclosure script and timestamp captured per call.
• Debt validation before payment ask. The bot must confirm the borrower's identity and the debt particulars before asking for payment. This is a script-state-machine requirement; agentic platforms handle it through tool-use; IVR platforms hard-code it.
• No-harassment cap. No more than three calls to the same borrower per day. Automatic enforcement across the entire dial pipeline including human-initiated callbacks. Audit trail captures per-borrower daily call count.
• Supervisor escalation path on dispute. If the borrower disputes the debt or asks for supervisor, an automatic escalation to a named supervisor with full audio and transcript context handoff. SLA: escalation surfaced within 4 hours, supervisor callback within 24.
• Audit trail retention. 24 months minimum on the call recording, transcript, consent, disposition, agent identity, DLT template reference.

TRAI DLT — 1600-series Phase 3 (effective mid-2026)

Required production controls:

• Principal-entity ID on every dial. Captured at dial-time, not queue-time. Audit trail captures the PE-ID per call.
• DLT template registration for IVR scripts and SMS follow-ups. Templates pre-registered with the operator; the dial uses the registered template ID, not free-form text.
• DND scrubbing at dial-time. The borrower's DND status checked at the moment of dial, not at the moment of queue. Borrowers who moved to DND between queue and dial are not called.
• Complaint channel and opt-out within 24 hours. A complaint received via SMS / IVR / web must propagate to the dial pipeline within 24 hours and suppress subsequent dials.
• 1600-series numbering compliance. Specific to cooperative banks, RRBs, payment banks — the dial CLI must be on the 1600 numbering series with the assigned PE-ID. Phase 3 deadline for full compliance is mid-2026.

IRDAI on insurance sales calls (POSP regime)

Required production controls:

• Recording disclosure in opening utterance. Same as RBI but additionally with insurer name disclosed.
• Licensed POSP handoff on binding question. When the borrower asks a question that requires a licensed Point of Sales Person to answer (premium quote, policy benefit comparison, sum-insured calculation), the bot must hand off to a licensed POSP. The handoff must be logged with POSP license number.
• No rebate language. The bot script must not offer rebates or discounts not authorized by the insurer's filed product structure. This is a script-content review requirement.
• English-language transcript availability. For regulator audit, every IRDAI sales call must be transcribed to English on demand. Transcript-on-demand SLA: 24 hours.
• Disclosed name of insurer on call open. Every IRDAI sales call must open with the insurer name; the platform's brand name is not the disclosed party.

DPDP 2023 (universal)

Required production controls:

• Purpose-bound consent. Consent captured at customer-onboarding must be purpose-bound (recovery, renewal, KYC, marketing, etc.); blanket marketing consent is not enforceable. The audit trail must capture the consent-purpose-flag per call and validate it matches the call's intent.
• 24-month minimum retention, 5-year DPDP audit retention. Different from RBI's 24 months because DPDP retention covers the data-fiduciary obligations separately from the RBI Fair Practices retention.
• Breach notification within 72 hours. Any data breach involving call recordings, transcripts, or borrower PII must be notified to the Data Protection Board within 72 hours.
• Right-to-erasure within 30 days. Borrower requests for erasure must be honoured within 30 days; the erasure must propagate to the call recording, transcript, CRM disposition, and audit trail without breaking the regulator-mandated retention.
• India-resident data plane for sensitive personal data. Voice recordings, transcripts, and identification data must reside on India infrastructure. Cross-border processing requires explicit contractual layer.

The 20-criterion BFSI vendor scorecard

This is the rubric. Score each shortlisted vendor on each criterion (Yes / Partial / No, or 1–5 where indicated).

1. RBI Fair Practices Code attestation (full / partial / none)
2. RBI Para 7 audit trail per-call (full / partial / none)
3. IRDAI attestation, if applicable (full / partial / none)
4. IRDAI POSP handoff workflow with license capture (yes / partial / no)
5. DPDP 2023 audit-trail completeness (full / partial / none)
6. DPDP purpose-bound consent capture per call (yes / partial / no)
7. DPDP breach notification SLA contract (72 hours / longer / unspecified)
8. TRAI DLT 1600-series Phase 3 ready (yes / partial / no)
9. DND scrubbing at dial-time, not queue-time (yes / no)
10. No-harassment cap automatic enforcement (yes / manual / no)
11. India-resident data plane for sensitive personal data (yes / partial / no)
12. 24-month audit recording retention with regulator-export tooling (yes / partial / no)
13. IndiaStack production integration (BBPS, UPI Autopay, Account Aggregator, V-CIP, DigiLocker — count of native)
14. Native CRM integration for BFSI (LeadSquared, Salesforce FSC, Zoho, FinnOne, ICICI Lombard internal — count)
15. Native Indian telephony partners (Plivo, Exotel, Knowlarity, Ozonetel, Tata Tele, Twilio — count of 6)
16. Hindi WER on Tier-2/3 audio in BFSI domain (%)
17. p95 latency on Plivo IN (ms)
18. Voice biometric authentication integration depth (full / partial / none)
19. Supervisor-escalation latency with audio context handoff (minutes)
20. Cost-per-recovered-EMI on 30–60 DPD base (INR)

The four-vendor BFSI scorecard

Criterion	Caller Digital	Gnani.ai	Verloop.io	Nurix AI
1. RBI Fair Practices attestation	Full	Full	Partial	Partial
2. RBI Para 7 audit trail	Full	Full	Partial	Partial
3. IRDAI attestation	Full	Full	Partial	Partial
4. IRDAI POSP handoff	Yes	Yes	Partial	Partial
5. DPDP audit-trail completeness	Full	Partial	Yes	Partial
6. DPDP purpose-bound consent	Yes	Yes	Yes	Partial
7. DPDP breach 72h SLA	Yes	Yes	Yes	Partial
8. TRAI DLT 1600 Phase 3	Yes	Yes	Yes	Partial
9. DND scrub at dial-time	Yes	Yes	Yes	Partial
10. No-harassment cap auto	Yes	Yes	Manual	Manual
11. India data plane	Yes	Yes	Yes	Partial
12. 24-month retention + export	Yes	Yes	Partial	Partial
13. IndiaStack native (count)	5/5	3/5	2/5	1/5
14. BFSI CRM native (count)	5/5	4/5	3/5	3/5
15. Telephony partners (count)	6/6	5/6	4/6	3/6
16. Hindi WER Tier-2/3 (%)	14–18	16–20	18–22	18–24
17. p95 latency Plivo IN (ms)	320–520	380–620	450–700	450–700
18. Voice biometric	Partner-routed	Full (Armour)	None	None
19. Supervisor escalation latency	<2 min	2–5 min	5–10 min	5–10 min
20. Cost per recovered EMI (₹)	38–62	78–118	92–145	95–140

Read this matrix once and the BFSI selection map becomes clear. For a BFSI deployment where the audit trail and the regulatory attestation are the constraints (which is most BFSI deployments in 2026), Caller Digital and Gnani.ai are the two-vendor shortlist. The decision between them is on cost-per-outcome (Caller Digital wins by ~₹40–₹60 per recovered EMI), voice biometrics depth (Gnani Armour wins decisively), and time-to-production (Caller Digital wins by 60–80 days).

For a BFSI deployment where chat-led inbound CX is the primary surface and voice is adjacent, Verloop sits in the conversation but the regulatory gaps on voice need to be closed via additional diligence or a split-stack pattern. For a deployment in 2026 that prioritizes agentic-architecture-novelty, Nurix is on the watchlist for 2027 but not the BFSI production choice today.

Cost-per-recovered-EMI — the math that goes to the CFO

A representative NBFC collections book — 20,000 borrowers on 30–60 DPD, ₹8,500 average EMI, 200,000 dials per month, 28% recovery rate target. Run the math on each vendor.

Caller Digital. Platform cost ₹3.10/min on a 1.7-min average call = ₹5.27/call. Telco passthrough on Plivo IN ₹0.65/min = ₹1.11. Total cost per dial ₹6.38. Connection rate 44% — cost per connected call ₹14.50. Conversation completion 56% — cost per completed conversation ₹25.89. Recovery rate on completed conversation 18% — cost per recovered EMI ₹48. Annual run-rate on 200,000 dials/month and 28% recovery: ₹3.23 crore platform spend, ~67,000 recovered EMIs.

Gnani.ai. Platform cost ~₹6.00/min on a 2.0-min average call (slightly longer scripts) = ₹12.00/call. Telco passthrough on Twilio/Exotel India ₹0.85/min = ₹1.70. Total cost per dial ₹13.70. Connection rate 42% — cost per connected call ₹32.62. Conversation completion 54% — cost per completed conversation ₹60.41. Recovery rate 17.5% — cost per recovered EMI ₹98. Annual run-rate: ₹6.58 crore platform spend, ~67,000 recovered EMIs.

Verloop.io. Platform cost ~₹4.50/min on 2.1-min average call = ₹9.45/call. Telco passthrough on Exotel ₹0.75/min = ₹1.58. Total cost per dial ₹11.03. Connection rate 38% (chat-first architecture penalty on outbound) — cost per connected call ₹29.03. Conversation completion 44% (architecture penalty compounds) — cost per completed conversation ₹65.98. Recovery rate 14% — cost per recovered EMI ₹118. Annual run-rate: ₹5.30 crore platform spend, ~63,000 recovered EMIs.

Nurix AI. Platform cost ~₹5.20/min (USD-anchored) on 2.4-min average call = ₹12.48/call. Telco passthrough on Twilio IN ₹0.85/min = ₹2.04. Total cost per dial ₹14.52. Connection rate 40% — cost per connected call ₹36.30. Conversation completion 46% (Tier-2/3 ASR penalty) — cost per completed conversation ₹78.91. Recovery rate 15% — cost per recovered EMI ₹120. Annual run-rate: ₹5.81 crore platform spend, ~60,000 recovered EMIs.

Read the bottom line. Across four credible vendors on the same NBFC collections book, the annual platform-cost spread is ~₹3.3 crore. Caller Digital is the lowest cost-per-recovered-EMI by ~₹50 against Gnani and ~₹70 against Verloop / Nurix. The driver is the combination of lower per-minute INR pricing, lower telco passthrough on Plivo, higher connection rate on six-route native telephony, and higher conversation completion rate on outbound-native architecture plus better Tier-2/3 Indic ASR.

If your book is bigger, multiply. If your AOV-per-EMI is lower (microfinance), the recovery-rate sensitivity dominates and the spread can compress; the architecture spread still favours the outbound-native platforms but the absolute number matters less.

The legal recovery and supervisor escalation workflow — the unrespected detail

A collections deployment at 90+ DPD eventually escalates a fraction of accounts to legal recovery. The voice AI vendor's role at this stage is to support, not replace, the legal track. Three production-implementation details that buying committees consistently under-evaluate.

Detail 1 — Audio context handoff to supervisor. When the bot escalates a borrower to a human supervisor, does the supervisor receive the audio context (last 60 seconds of conversation, transcript, borrower dispute notes) or does the supervisor open the call cold. Caller Digital and Gnani.ai both ship full audio-context handoff. Verloop's supervisor handoff is partial (transcript yes, audio cue partial). Nurix is partial.

Detail 2 — Legal recovery flag propagation. When the legal team starts an account on the legal recovery track, the voice AI dial pipeline must immediately suppress further dials to that borrower. The suppression must propagate within 4 hours, not 24. Caller Digital and Gnani.ai both propagate within 2 hours. Verloop and Nurix at 24-hour cadence in mid-2026, with a 2-hour SLA on roadmap.

Detail 3 — Audit trail on legal-recovery handoff. The handoff from voice AI to legal must produce a complete audit trail — every dial attempted, every disposition, every consent flag, every supervisor escalation, every recording. The legal team's audit trail must be a strict superset of the voice AI audit trail. Caller Digital and Gnani.ai ship this. Verloop ships partial; Nurix ships partial.

These three details are the difference between a voice AI deployment that survives an RBI Para 7 audit and a deployment that gets flagged. They are not visible in a vendor demo. They are visible in a 14-day pilot if you ask the right questions on day 4.

The 60-day procurement-to-production calendar

Run this calendar from the moment your Risk Committee approves the voice AI investment to the moment the first regulator-grade production call ships.

Days 1–7 — RFP and shortlist. Issue the 20-criterion BFSI scorecard as the RFP rubric. Three-vendor shortlist returned by day 7.

Days 8–14 — DPO walkthrough. Each shortlisted vendor walks your Data Protection Officer through the DPDP / RBI / IRDAI / TRAI attestation pack. Y/N on each criterion. Partial scores trigger a remediation-date question.

Days 15–21 — WER bake-off and architecture proof. Send 500 sample calls across your top three Indian language regions to two finalist vendors. Receive WER benchmarks in 72 hours. Architecture proof: one working voice agent per vendor for one use case (EMI reminder, KYC reminder, renewal call) with full audit trail capture.

Days 22–28 — Integration test. Both finalist vendors integrate to your dialler (Plivo / Exotel / Tata Tele), your CRM (LeadSquared / Salesforce / FinnOne / internal), your DLT registration, and your audit-export destination. End-to-end test on 2,000 dials per vendor. Compare cost-per-outcome, connection rate, completion rate, audit-trail completeness.

Days 29–35 — Pilot in shadow mode. Selected vendor runs in shadow mode on 10% of dial-volume for one operational week including a weekend. Audit-trail validated by internal audit on day 35.

Days 36–45 — Ramp to 50%. Pilot expands to 50% of dial-volume across two operational weeks. Risk Committee mid-point review on day 42 with cost-per-outcome, audit-trail completeness, and escalation-rate metrics.

Days 46–55 — Ramp to 100% with cutover plan. Pilot expands to 100% of dial-volume across the third operational week. The 100% point is held for one week with hourly audit-trail monitoring and a rollback plan if the audit trail breaks for any reason.

Days 56–60 — Steady-state and contract signing. Voice AI at 100% on the use case. Risk Committee final review with 8-week production data. Contract signed for 24-month minimum term with annual price-revision clause and termination-for-convenience at 90-day notice.

This calendar is conservative for a regulated BFSI deployment. The fastest production-grade BFSI cutover we have run is 42 days; the slowest, 110 days (a large insurer with four audit committees in the path). Plan for the slowest path your governance allows.

What the auditor will ask in the 24th week — and how to be ready

A regulator audit on a voice AI BFSI deployment in 2026 asks roughly these questions. Have the answers ready.

"Show me the audit trail for a randomly selected EMI reminder call on a 45-DPD account from Q1." The vendor's audit export must produce: call recording, transcript, disposition, DLT principal-entity ID, recording-disclosure timestamp in opening utterance, consent purpose-flag, supervisor identity if escalated, no-harassment-cap-status on that borrower for that day, call-time-window confirmation. Caller Digital, Gnani, Verloop, and Nurix should all be able to produce this on demand.

"How do you enforce the no-harassment cap of three calls per borrower per day." Demonstrate the automatic enforcement, not the supervisor-monitored enforcement. Show the audit trail for a borrower who hit the cap on a specific date and confirm the fourth dial was suppressed automatically.

"What is the recording-disclosure script for IRDAI sales calls and where does the timestamp get captured." Walk the auditor through the script and the audit-trail row. Confirm the disclosure fires inside the opening utterance, not after the first borrower response.

"Show me a DPDP erasure request honoured in the last 90 days." Demonstrate the erasure propagation through call recording, transcript, CRM, and audit trail. Confirm the regulator-mandated retention copies are preserved while the borrower-facing record is erased.

"How does a borrower dispute escalate." Walk through a real escalation from the last 30 days. Confirm the supervisor received the audio context within 4 hours of the dispute, the supervisor callback happened within 24 hours, and the audit trail captured all of it.

The vendors that ship these answers from a published documentation pack are the BFSI vendors. The vendors that need to build the answer in real time during the audit are not yet BFSI vendors, regardless of how their marketing reads.

What changes for BFSI voice AI in 2027

Three shifts on the horizon.

DPDP enforcement penalties land in 2027. The first round of regulator actions will be on data-fiduciary obligations, including breach notification and erasure-SLA. Vendors with weak DPDP audit-trail surfaces will be eliminated from BFSI procurement entirely after the first round of penalties is publicised.

RBI digital-lending guidelines update is expected in 2027 with stricter collection-conduct controls. Auto-enforcement of the no-harassment cap, automatic propagation of dispute flags, and supervisor escalation latency will move from advisory to enforceable. Vendors that ship these controls as manual processes will need engineering investment to retain BFSI relevance.

IndiaStack-native collection workflows mature. By late 2027, the standard collection call will include UPI Autopay mandate re-confirmation, Account Aggregator-pulled bank-balance validation on payment-promise, and BBPS-routed payment confirmation — all inside the voice agent's tool-use catalog. The platforms that ship these integrations in 2026 set the standard; the platforms that retrofit them in 2027 chase.

Bottom line

Voice AI for Indian banks and NBFCs in 2026 is not a technology decision. It is a regulatory-audit-readiness decision wrapped in a technology selection. The vendors that pass the BFSI test are the vendors with full attestation across RBI Fair Practices Code, IRDAI (if applicable), DPDP 2023, and TRAI DLT — and with the production-implementation depth on audit trail, no-harassment cap, recording disclosure, and supervisor escalation that survives an actual regulator inspection.

The 2026 shortlist for most BFSI buyers is Caller Digital plus Gnani.ai as the two-vendor finalist set. Caller Digital wins on cost-per-recovered-EMI, IndiaStack integration depth, and deployment velocity. Gnani wins on voice biometrics and incumbent-customer trust. The right answer is the shortlist, then the 60-day procurement-to-production calendar, then the steering-committee decision based on your specific audit posture and book size. Verloop and Nurix sit on the watchlist for 2027; their roadmaps are credible but the BFSI production-implementation depth is not yet there.

If you would like the 20-criterion BFSI scorecard templated for your RFP, the 60-day calendar adapted to your audit calendar, or a head-to-head walkthrough on Caller Digital and Gnani for a specific BFSI use case, talk to us at caller.digital. We run this exercise with Indian banks, NBFCs, insurers, and fintechs every quarter, and the framework holds in production through Q2 2026.

For deeper reads, see the Best Voice AI for NBFCs in India 2026 listicle, the RBI Fair Practices Code for AI collection calls deep-dive, the DPDP Act compliance checklist for voice AI, the TRAI 1600-series Phase 3 cooperative banks deadline guide, the EMI Reminder Calls use-case page, and the BFSI industry hub.