Our Voice AI agent handles the call, the WhatsApp follow-up and the payment link. It only dials inside the 08:00–19:00 window, states who it is and who it is calling for, logs every word to your CRM, and hands the difficult conversation to a human before it becomes a complaint.
30 minutes with a solution architect who has shipped this into regulated environments. We will map one workflow — usually bucket 0–30 — against your LMS and show you the call.
Most lenders solved them with process notes and agent training. Process notes do not scale to four lakh calls a month, and a training deck is not an audit trail. Rules that live inside the dialler do not get forgotten at 7:04pm on a quarter-end Friday.
Service and transactional voice calls must originate from a TRAI-allocated 1600-series number. Phased by entity: commercial banks first, NBFCs above ₹5,000 crore in assets from 01 Feb, the rest by 01 Mar 2026. A customer who does not recognise your number does not answer — and the whole point of the series is that they now can.
TRAI direction under TCCCPR 2018; RBI circular dated 17 Jan 2025.No contact before 08:00 or after 19:00. No intimidation, no persistent calling, no approaching family or referees, no misleading representation. The exposure is not theoretical — it is the single most common route to an Ombudsman complaint against a lender's collections function.
RBI/2022-23/108, 12 Aug 2022 — responsibilities of REs employing recovery agents.Notified November 2025 with a phased runway. Consent has to be specific, recorded, and withdrawable — which means the consent state has to be readable by whatever places the call, at the moment it places it. Retrofitting that into a dialler later is the expensive version.
Digital Personal Data Protection Rules, 2025, notified by MeitY, Nov 2025.NBFC credit grew 14.4% year-on-year in June 2026, with retail lending up 20.3%. Headcount in collections and servicing did not grow 20%. Something has been quietly giving way, and it is usually the same three things.
Pre-due and 0–30 DPD is where a reminder still works and a human agent is least likely to be assigned. Teams triage upward, towards the harder buckets, because that is where the manager's number is. The soft bucket silently rolls forward.
Month-end targets and a hard 19:00 cut-off pull in opposite directions. One agent making one call at 19:20 is a Fair Practices Code breach, and it will be the one call the borrower records. Enforcement belongs in the system, not in a supervisor's memory.
A pan-India book needs Hindi, Marathi, Gujarati, Tamil, Telugu, Kannada, Bengali and Punjabi on any given shift. Staffing that in one contact centre is a rostering problem with no clean answer, so the calls go out in the wrong language and contact rate drops.
Collections is where the ROI is obvious, so it is usually first. It is rarely the only one worth doing — the same agent, the same integration, the same consent record.
Every call placed inside the permitted window, opening with a disclosure of who is calling and on whose behalf, and stopping the moment a dispute is raised.
What the call actually sounds like
Configured, not coded: permitted-hours window, per-state holiday calendar, maximum attempts per day and per week, cooling-off after a PTP, and a hard stop on any number in your DNC or litigation list.
The high-volume, low-complexity inbound traffic that fills your queue and inflates average handle time without generating a rupee.
What the call actually sounds like
The point of this tab: deflecting routine servicing is what frees the human bench for the collections and dispute calls that actually need a person.
The stretch between application and first EMI is where drop-off is highest and where a call costs least to make.
What the call actually sounds like
Worth noting: a first-EMI bounce caused by an unregistered mandate is an onboarding failure that gets recorded as a credit event. This tab prevents the ones that were never about ability to pay.
Speed to lead decides who converts. An agent that answers in seconds, at 11pm, in Tamil, is not competing with your sales team — it is competing with the lender who called first.
What the call actually sounds like
Consent, explicitly: promotional outreach runs against your DLT-registered consent state and DNC scrubbing. If consent is absent or withdrawn, the number is not dialled — there is no override in the interface.
The calls where seconds matter and volume is unpredictable — exactly the shape of demand a human roster handles worst.
What the call actually sounds like
Escalation is a first-class path, not a failure state. Sentiment, keyword and intent triggers route the call to a human with the full transcript already on their screen.
A reminder call that ends without a payment link has done half a job. GirikVoice moves between channels inside a single interaction and keeps one thread against one customer record.
Inbound and outbound, on your 1600-series numbering, with barge-in and interruption handling.
Payment links, statements and document uploads pushed mid-call and logged to the same record.
DLT-registered templates for the confirmations and reminders that do not need a conversation.
The same agent on your site and app, so a customer does not restart when they switch device.
Escalation with full context, so the agent who picks up is not asking the customer to repeat themselves.
We are not going to put someone else's recovery percentage on this page. What we will do is be precise about the mechanism, so you can run the arithmetic against your own book before the call.
Falls because authentication, balance lookup and status queries stop consuming agent minutes. The residual human calls get longer, not shorter — they are the complex ones now.
Rises where the agent can complete the action — raise the request, send the link, book the slot — instead of promising a callback that becomes a second contact.
Rises on three levers at once: calling in the right language, calling at the hour that customer historically answers, and calling from a number the 1600 series makes recognisable.
Falls in bucket 0–30 and in routine servicing, where volume is high and each interaction is nearly identical. It does not fall in dispute handling, and you should not model as if it will.
For context on what the sector is publishing: girikon.ai's own banking practice reports a 45-second average hold time, a 38% increase in first-call resolution and a 22% reduction in call-centre staffing cost within two quarters. Your numbers will depend on your product mix, your book and where you start.
These are configurable platform capabilities, set per deployment. They are stated as capabilities, not as certifications — the certifications we hold are named separately below.
08:00–19:00 by default, with per-state holiday calendars and per-campaign overrides that require a named approver.
Every outbound call opens by stating that it is automated, who is calling and on whose behalf, before any account matter is raised.
The agent reads consent and DNC status at the moment of dialling, not from an overnight extract. Withdrawal takes effect on the next call, not the next batch.
Full recording and transcript against the customer record, with retention tiers set to your policy and audit-ready export.
Account numbers, identifiers and other personal data masked in transcripts and logs, with role-based permission to unmask.
In-region hosting available for Indian deployments. Confirm the specific region and sub-processor list with our architect — it is a contractual term, not a marketing line.
Encryption in transit and at rest, role-based access, immutable audit logging and a 99.99% uptime SLA.
SOC 2 Type II audited. GDPR-aligned workflows. HIPAA-compliant deployments where relevant. Girikon is a Salesforce Consulting Partner.
Independently audited and certified
GirikVoice sits over the stack you already run rather than asking you to change it. If you happen to be on Salesforce we go deeper than most — Girikon has been a Salesforce partner since 2015 — but nothing on this page assumes it, and no capability is gated behind it.
Salesforce, Dynamics, Zoho, HubSpot — or the in-house LOS your team built in 2016 and nobody wants to touch. REST and webhooks both directions, so outcomes land where your reporting already looks.
Bring your own carrier, SIP trunk and 1600-series numbering. We are not a telecom reseller, so there is no commercial reason for us to push you off what you already have.
Campaigns fire from a presentment failure, a DPD bucket change or an application stage — from Salesforce Flow or from your own webhook. Not from a list somebody exports on Monday morning.
Most lenders we speak to have already run an AI pilot that never left the sandbox — usually because it was scoped as a technology evaluation rather than as one workflow with a number attached. Pick the workflow with the clearest arithmetic, put it against a controlled segment with a human control group, and let the result decide whether it widens.
We map one workflow end to end with your collections and compliance leads: triggers, permitted windows, escalation rules, what the agent may and may not say.
Connect to your LMS or CRM, wire the consent and DNC read, build the call flow in your two or three highest-volume languages.
Run against a ring-fenced segment with a human control group alongside. You measure contact rate, PTP capture and complaints — not us.
Widen the segment, add languages, then take the next workflow. Each one after the first is configuration, not a new project.
It is the right worry, and it is exactly what the 1600-series migration is meant to fix — a recognisable, regulator-allocated number for service and transactional calls. On top of that, the agent discloses in the opening line that the call is automated, names your institution and states what it is calling about, before it asks for anything. It never asks for an OTP, a PIN, a password or a card number. That last rule is not configurable, because the moment a legitimate lender's automated agent asks for a credential, it has trained the customer to give credentials to whoever asks next.
The 08:00–19:00 constraint sits in the dialling layer, not in an agent script or a supervisor's judgement. Calls outside the window are not placed — they queue for the next permitted slot. Per-state holiday calendars, maximum attempts per day and per week, and cooling-off after a captured promise-to-pay are configured the same way. Overrides exist for genuinely non-recovery communication but require a named approver and are logged as such. The reference is RBI/2022-23/108 dated 12 August 2022.
GirikVoice supports over 100 languages across deployments. For an Indian lending book the working set configured is typically English, Hindi, Marathi, Gujarati, Tamil, Telugu, Kannada, Bengali and Punjabi, including code-mixed speech — which is how people actually talk about money on the phone. The language for each customer is selected from your record, and the register is set to the formal BFSI tone rather than a marketing voice. Confirm your specific language list with the architect on the call; it is a scoping item, not a licence tier.
The agent stops collecting. A dispute, a hardship claim, a legal reference or a rise in distress signals routes the call to a human with the transcript already on screen, or books a callback if no one is available. This is deliberate design, not a limitation — an automated system that argues about a disputed amount is precisely how a collections function ends up in front of the Ombudsman. Negotiation stays with your people; the volume work does not.
You remain the data fiduciary; we operate as a processor under your instructions. In-region hosting is available for Indian deployments, and the specific region, sub-processor list, retention tiers and breach-notification terms are contractual — we will put them in writing rather than leave them on a web page. Encryption in transit and at rest, role-based access, PII masking in transcripts and immutable audit logging are standard. Bring your infosec team to the second call; that is the one they should be on.
No — it usually simplifies them. GirikVoice places calls over your telephony, so it inherits whatever numbering you have provisioned. If the migration is still in progress, the agent runs on your existing numbers and moves across when the series is live, with no rebuild of the call flows. What does need a decision early is the split between service and transactional traffic on the 1600 series and promotional traffic on the 140 series, because that determines which campaigns can run where.
Not in the deployments that work. The pattern that holds is: the agent takes bucket 0–30 and the routine servicing volume, and the human bench moves up to the buckets and conversations where a person genuinely changes the outcome. Most lenders find the constraint was never headcount in the abstract — it was that trained collectors were spending their day on reminder calls. If your business case depends on a headcount reduction, say so on the first call and we will model it honestly rather than let it surface at procurement.
Around four weeks to a controlled pilot on one workflow, assuming your integration contact is available and the compliance sign-off is not sequential. That timeline is a plan, not a guarantee — the variables are how clean the LMS event is, how many languages are in the first release, and how long your risk review takes. We will give you a dated plan after the scoping call rather than a number on a landing page.
Thirty minutes with a solution architect, not a sales development rep. Come with a real workflow — bucket 0–30 reminders, application status calls, NACH failure follow-up — and we will map it against your stack on the call.
Tell us the workflow you want to look at and we will map it against your stack on the call. No sequence, no drip — we reply within one business day.
Go to the booking formTakes about 40 seconds. Seven fields.