Agent-Assist and Voice AI at an Indian Private Bank Contact Centre — 38% AHT Reduction Across Hindi, English and Tamil
Sentiment Analyzer + ChatNext + Voice Bot deployed across 4,200 contact-centre agents — agent-assist plus voice-AI deflection in a single platform.
The challenge
The bank — a major Indian private-sector retail bank with a contact-centre operation of approximately 4,200 seats across three centres — was facing a chronic agent-productivity and customer-experience challenge. Average handle time across the contact centre had crept upward as the product portfolio had expanded and the regulatory-disclosure scripting had thickened. Agent attrition was running at industry-typical but still painful 38% annualised. The bank's customer-experience scores on the contact-centre journey had been stagnant for three years.
The bank's previous chatbot — a rules-based engine on the website and the bank's mobile app — handled approximately 12% of customer interactions but with low customer satisfaction and frequent escalation to live agents. Voice-AI had been tried twice previously: a global vendor's voice IVR (replaced after eighteen months due to poor Indian-English accent handling) and an in-house Hindi voice prototype that never reached production.
The constraints were specific. RBI requirements meant all customer-conversation processing had to happen inside India's borders, with seven-year retention on conversation records. The bank's customer base required support in Hindi, English and Tamil at minimum, with frequent code-switching within single conversations. The bank's existing contact-centre platform (a customised Avaya / Verint deployment) could not be replaced — the bank's CCO had been explicit that the new AI capability had to integrate with the existing platform, not require its replacement.
The approach
MindMap deployed a unified agent-assist-and-voice-AI platform composed of Sentiment Analyzer (Sa) for real-time call analysis, ChatNext (Cn) as the agent-assist conversational engine, Voice Bot (Vb) for the front-line voice-AI deflection, NLP Router (Nr) for the intent-routing layer and Knowledge Base Builder (Kb) for the agent knowledge base.
Phase one was the agent-assist layer. While the agent is on a live call, the platform performs real-time speech-to-text on the call, identifies the customer intent, retrieves the relevant knowledge-base articles and policy excerpts, and presents the agent with a structured response suggestion. For multi-system tasks (account look-up, transaction status, dispute filing) the agent-assist surfaces a one-click action that pre-fills the required fields in the agent's workstation.
Phase two was the voice-AI deflection layer. For calls in the bank's defined automatable-intent set (balance enquiry, mini-statement, card-block, EMI confirmation, branch and ATM locator, deposit-rate enquiry) the Voice Bot handles the entire conversation in Hindi, English or Tamil based on the caller's language preference detected at the start of the call. Calls that fall outside the automatable set are routed to the appropriate live agent with the conversation context already captured.
Phase three was the post-call layer. After every call (both AI-handled and agent-handled) the platform generates a structured call summary, the customer-relevant follow-up action items, the call-disposition categorisation, and the compliance-script-adherence assessment. The previous post-call work — which had absorbed roughly 4 minutes per call for the agent — is now generated automatically and presented to the agent for one-click confirmation.
The pre-built building blocks
Rather than commission a ground-up build, the engagement leaned on MindMap's pre-built accelerator library — production-tested components that compress what would otherwise be a six-to-nine-month build into weeks.
ChatNext
Agent-assist conversational engine
Voice Bot
Front-line voice-AI deflection in Hindi, English and Tamil
Sentiment Analyzer
Real-time call sentiment and intent detection
NLP Router
Intent routing across voice-AI and live-agent paths
Knowledge Base Builder
Bank policy and FAQ knowledge base with daily refresh
The architecture
The platform runs entirely on the bank's private cloud inside its primary data centre in Mumbai, with active-active failover to a secondary site in Chennai. The voice processing — speech-to-text, the LLM inference, text-to-speech — happens inside the bank's perimeter; no audio leaves the bank's network.
Speech-to-text uses a fine-tuned Whisper Large v3 model with custom adapters for Indian-English, Hindi and Tamil. The fine-tuning corpus is approximately 4,500 hours of in-domain audio drawn from the bank's own call recordings (PII-redacted before training use). Latency is critical: the speech-to-text round trip stays below 280ms to preserve conversational flow on the voice-AI calls.
The reasoning layer uses Llama 3.1 70B for the agent-assist conversational engine and a fine-tuned Mistral 7B for the voice-AI conversation graph (chosen for lower latency on the voice path). Both are served via vLLM on the bank's tenant H100 fleet. The voice-AI's conversation graph is constrained — the model cannot generate commitments outside the bank's defined response set — but is free to vary phrasing, handle interruptions and manage objections.
The knowledge-base layer is built on Knowledge Base Builder, with the bank's full policy library, product documentation, FAQ corpus, and historical resolved-case repository ingested and indexed. Knowledge-base content is refreshed nightly from the source-of-truth systems, with the agent-assist's response suggestions grounded in the current knowledge-base content with explicit document citations.
Integration with the existing Avaya / Verint contact-centre platform is via the platform's standard CTI integration for the voice-routing layer and via the conversation-analytics export for the post-call processing. The bank's existing contact-centre routing logic was not replaced.
Every call (voice-AI and agent-handled) is recorded in full, transcribed, scored against the bank's compliance criteria and stored with the RBI-required seven-year retention.
The numbers behind the story
On agent-assisted calls, average handle time has dropped by 38%. The reduction is driven by the agent-assist's pre-emptive knowledge-retrieval (agents no longer search for policy answers mid-call), the one-click multi-system actions (agents no longer navigate across applications during the call) and the automated post-call processing (agents no longer manually type call notes).
Voice-AI deflection sits at 31% of total inbound contact-centre volume. The deflected calls have a customer-satisfaction score that is, on the bank's measurement, statistically indistinguishable from comparable live-agent calls — the calls in the automatable-intent set are the calls where customers most want speed and predictability rather than human empathy.
Compliance-adherence has improved. The voice-AI calls are 100% script-compliant by construction; the agent-assist's real-time prompts have raised the live-agent compliance score from 89% to 96% on the bank's quality-assurance sampling. This was a meaningful factor in the bank's RBI inspection outcome.
Agent experience has improved. The bank's agent satisfaction survey, conducted six months post-rollout, recorded the highest agent-engagement scores in the contact centre's history. Agent attrition has dropped from 38% annualised to 27%, with the agent population's working theory being that the platform handles the repetitive intent classes and leaves the agent with the more interesting and varied work.
An unexpected outcome: the bank's contact-centre training time for new agents has dropped by roughly 40%, because the platform's agent-assist provides the in-context knowledge support that previously required extended pre-deployment training to internalise.
“We had two failed voice-AI attempts and an understandably sceptical contact-centre leadership team. MindMap's Indian-bank reference broke that scepticism. Six months in, we are deflecting thirty-one per cent of contact volume to voice-AI, agent-assisted calls are thirty-eight per cent shorter, and our agents are happier than they have ever been. The RBI inspection called out the compliance improvement specifically.”— Chief Customer Officer· Indian Private Bank
Why MindMap was chosen
The bank had two prior voice-AI failures behind it and an internal team that was understandably sceptical of yet another attempt. MindMap's ChatNext deployment at another Indian bank (with the same language mix and a comparable contact-centre scale) was the reference that broke that scepticism — the bank's contact-centre leadership did a site visit and spoke directly to their peer team about the deployment's actual operational reality.
The willingness to deploy entirely on-premises within the bank's RBI-compliant infrastructure — including the fine-tuning, which involved sensitive customer-conversation data — was the regulatory differentiator against the global vendors who required some component to run in cloud regions outside India.
Our embedded contact-centre operations expertise — two former contact-centre operations heads from peer Indian banks — was the third factor. The bank's CCO felt that the team understood the operational realities of running an Indian contact centre, not just the AI technology.
Related deployments
Sovereign WhatsApp Banking
ChatNext-powered WhatsApp bot deployed inside the bank's air-gapped data centre, handling balance, transfers, statements and loan applications in English and Swahili.
Cheque OCR at 99.1% Accuracy
DocuMage replaced a legacy template OCR for cheque clearing, processing 10,000 cheques per day at 99.1% field accuracy with 94% straight-through processing.
UK Challenger Bank KYC
OnboardX rebuilt the KYC pipeline with liveness, sanctions and PEP enrichment in a single STP flow, collapsing onboarding from 5 days to 4 hours.
Want an outcome like this?
Start with a 2-week AI Readiness Sprint. We deliver a prioritised use-case backlog and business case grounded in what's actually buildable with our accelerator library.