Ringg’s AI agents resolve up to 65% of customer calls with OpenAI

Customer operations usually scale by adding people, which adds cost and complexity per interaction. Ringg’s story is that high-efficiency models plus serious orchestration can clear a majority of routine inquiries across voice, chat, WhatsApp, and web.

For consumer businesses in India and similar markets — insurance, health appointments, investing — that is a direct P&L and response-time lever, not a pilot slide.

We’re covering this as a production agent-ops case, with the usual caveat that it’s a vendor-and-customer narrative.

Ringg built an enterprise agent platform after watching large consumer businesses struggle both to answer volume and to complete tasks across fragmented systems. Migrating suitable real-time workloads from GPT-4.1 to GPT-5.6 cut model costs by about 90% while meeting quality and latency needs, per the company. Headline metrics: more than 7 million connected calls a month, up to 65% request resolution via agents, average CSAT 4.8.

Architecture matters more than the logo on the model. Agents interpret requests, select tools, and walk multi-step workflows; an orchestration layer hits CRMs, ticketing, payments, scheduling, and internal APIs, escalating to humans with a summary when needed. Knowledge mixes structured filters and semantic retrieval over PDFs, CSVs, and docs. Specialized subagents handle qualification, support, verification, scheduling, and escalation across channels.

Model routing is deliberate: GPT-4.1 still carries much real-time voice and chat; GPT-5.6 Luna for certain price-performance real-time work; Terra for post-call summaries and sentiment; Sol for evals, prompt improvement, and model-as-judge. Long conversations get a structured summary near ~80,000 tokens so history isn’t resent forever. Offline evals, canary production traffic, latency-aware routing, and versioned deployments are how they claim to ship without breaking voice.

Customer vignettes are specific. Policybazaar: 67% of calls without humans, response time from 8–12 minutes to under 60 seconds (~88% improvement) across 57,000+ requests. Practo: 85% first-call resolution, sub-3-second responses, 70% lower operating cost versus prior human-led workflow, 1,000+ bookings a day. Groww: 72% of IPO/F&O inbound queries self-serve at ~2 minutes average handle time. They’re also pushing computer-use browser agents for KYC, claims, and IT troubleshooting.

Useful AI in support looks like resolved intents and preserved human escalation — not chatbots that trap people. Cost cuts that follow real model migration are healthier than headcount fantasies. The downside: automation rates can hide frustrated customers who never reach a person; CSAT averages can mask skewed samples; and regulated domains (insurance, health, finance) demand audit trails when an agent books, sells, or advises. Multilingual code-switching is a strength in their markets and a failure mode if evals are thin.

I’m watching whether the 65% resolution band holds as browser agents expand into KYC and claims — higher-stakes flows where a wrong tool call is not a mild inconvenience.

OpenAI customer story, Sep 23, 2026, on Ringg (Asia-Pacific startup) using GPT-5.6 and related models via the API for multilingual voice/chat agents. Includes Policybazaar, Practo, and Groww results as reported by OpenAI/Ringg.