Reversing port-outs at scale with a Hinglish AI retention voicebot
An AI voicebot intercepted Airtel Mobile Number Portability port-out requests in the Bihar circle and, after rapid prompt-and-flow optimization, lifted verified in-call retention from ~9% to a 48% peak — comfortably past the ~10% target.
About the client
Airtel is one of India's largest mobile network operators. Within the Mobile Number Portability (MNP) window, a subscriber who files a port-out request can be won back only in a narrow span before the port completes — a high-volume, time-critical retention problem. Airtel engaged ORI to run an AI-voicebot proof-of-concept in the Bihar circle to intercept these requests, diagnose the reason, resolve it, and persuade subscribers to stay — starting at 4,000 contacts per day from a 24 June 2026 go-live, with expansion to Madhya Pradesh, Gujarat and UP-West to follow.
The challenge
Retaining porting-out subscribers inside a short MNP window demanded personalized, high-volume, compliant outreach that manual tele-calling could not deliver at scale.
Manual tele-calling couldn't scale to thousands of daily port-outs without heavy agent cost.
Generic scripted outreach converted poorly against subscribers already decided to leave.
The retention window is short — once a port completes it cannot be reversed.
Reasons for leaving varied widely — network, price, competitor offers, validity.
Every interaction had to stay strictly on approved plans and prices.
Multilingual subscriber base — Hinglish for Bihar/MP/UP, Gujarati for Gujarat.
Beyond an off-the-shelf bot: the AI engineering we ran
Retention in the MNP window is not a scripting problem — it is an AI-engineering problem. Seven concrete workstreams carried the lift.
- Hallucination control on live data. A hard pricing lock — the bot pitches only from a fixed catalog and never invents any price, plan, discount or benefit. Unknown plans are deflected to the Airtel Thanks app rather than guessed.
- Reliability engineering on the dialing layer. Retry and redial logic at ~2.2 attempts per unique contact, pushing connected reach to ~56% of the port-out base even in a non-integrated, campaign-mode setup running seven days a week.
- Real-time conversation engineering. One-idea-per-turn discipline, no repeated sentences, barge-in handling, sub-25-word turns, and TTS-safe formatting keep conversations natural and low-latency.
- Observability and reporting. Connectivity, dial attempts, disposition, retention and reason-for-porting instrumented and reported daily in consolidated dashboards per campaign.
- Data engineering for attribution. Retention counted only when the subscriber sends CANCEL to 1900 on the call — not on a verbal "yes". Measured retention is deliberately conservative; true business retention runs higher.
- Closed-loop automation. Callback capture, network-complaint registration and human-agent handoff scheduling all happen inside the conversation, with no manual handoff.
- Continuous experimentation. A/B testing of opening script, retry rules and call timing, combined with a prompt/flow rewrite, drove the step-change from ~9% to a 48% peak retention rate.
Results, measured end-to-end
| Funnel metric | Result | Measured against |
|---|---|---|
| Connectivity (unique contacts reached) | 56% | 44,821 unique port-out contacts |
| Connect rate per dial attempt | ~26% | 99,385 dial attempts (~2.2/contact) |
| Blended retention | 14% | 46,390 connected calls |
| Peak daily retention | 48% 4.2× uplift | 2,003 connected calls · 12 Jul |
Retention is counted only when a subscriber confirms the CANCEL-to-1900 SMS on the call; verbal-only agreements are excluded, so measured retention is deliberately conservative and true business retention runs higher.
The solution we designed
End-to-end journey
A port-out triggers an outbound Hinglish call. It opens with a single warm probe into why the subscriber wants to leave, acknowledges the reason once, and routes into the matching objection branch — network, price, competitor offer, validity, hardship, ownership transfer. It resolves the concern, pitches the single best-fit plan from the approved catalog with a now-anchored close, then guides the subscriber to send the cancellation SMS, confirms it was sent, and closes — all in one seamless conversation.
Brand-tuned, multilingual experience
The bot speaks natural Hinglish exactly like an experienced North-Indian tele-caller — Devanagari for Hindi, Roman for English, in one sentence — with dialect and accent handling and strict gender-neutral address. Hinglish covers Bihar, MP and UP-West; Gujarati is used for the Gujarat circle. Mid-conversation language guardrails keep the interaction on a supported language and hand off gracefully when it isn't.
Integrated systems
The POC runs in a non-integrated, campaign-mode setup that sits alongside Airtel's existing stack without disruption. Integration with Airtel's in-house dialer and SMS gateway is in the pipeline, which will further tighten reach, latency and the closed-loop cancellation step.
Measurement built in
A daily consolidated reporting layer tracks unique contacts, dial attempts, connectivity, disposition, reason-for-porting and retention across the funnel, with retention attributed to the confirmed cancellation event so outcomes can be proven rather than asserted.
Retention within our Mobile Number Portability window is one of the most time-critical challenges we face. Working with Oriserve, we lifted retention from roughly 9% to a peak of 48% — going from winning back less than one in ten customers to nearly half. That result has given us real confidence to expand this approach beyond our initial circle.
ABusiness Stakeholder · Bharti Airtel
Why a generic solution would fail here
Hallucination. An off-the-shelf bot would invent plans, prices or offers under pressure. The pricing lock and catalog grounding make invented figures impossible.
Poor reachability at scale. Without retry engineering, a generic bot wastes the short MNP window on unanswered calls. Redial logic at ~2.2 attempts per contact lifted connected reach to ~56%.
Attribution blindness. A generic bot counts a verbal "yes" as a win. Attribution to the confirmed CANCEL-to-1900 event ties every retention to a real, auditable action.
No iteration loop. A static script plateaus — here at roughly 9%. Continuous A/B testing and prompt optimization drove retention to a 48% peak, past the ~10% target.
The takeaway
The retention lift did not come from a bot; it came from engineering — grounding on live data, reliable reach, honest attribution and relentless experimentation. That is what carried a Bihar proof-of-concept from roughly 9% to a 48% peak retention rate, and it is what makes the expansion to Madhya Pradesh, Gujarat and UP-West a scale-up rather than a restart.
Have a retention or renewals workload like this?
We'll pull the closest live production analog and walk you through how the numbers were reached.