Ringg, which runs voice and chat agents for large consumer businesses in India, has built its platform around OpenAI's GPT-5.6 model family and now handles more than 7 million connected calls a month. Ringg says its agents resolve up to 65% of routine customer inquiries without a human taking over, with customers reporting an average CSAT of 4.8.
The setup routes work across several model variants: GPT-4.1 carries most real-time voice and chat traffic, GPT-5.6 Luna takes select real-time tasks, GPT-5.6 Terra handles post-call analysis and sentiment classification, and GPT-5.6 Sol supports evaluation and prompt-improvement workflows. Ringg says moving suitable real-time workloads from GPT-4.1 to GPT-5.6 Luna cut model costs by roughly 90% while holding the quality and latency the workloads require.
How it works
- Multi-model routing: An orchestration layer picks among GPT-4.1, GPT-5.6 Luna, Terra, and Sol based on the task — real-time voice and chat, post-call analysis, or evaluation.
- Reported cost cut: About 90% lower model cost on the real-time workloads moved from GPT-4.1 to GPT-5.6 Luna, according to Ringg.
- Scale: More than 7 million connected calls a month across voice, chat, WhatsApp, and web.
- Customer deployments: Policybazaar handles 67% of calls without a human and cut response time from 8–12 minutes to under 60 seconds; Practo reports 85% first-call resolution, response times below three seconds, and 70% lower operating costs; Groww resolves 72% of IPO, futures, and options queries through self-service at a 2-minute average handling time.
- Still in progress: Structured conversation summaries when threads approach ~80,000 tokens, plus browser agents on OpenAI's computer-use capabilities for KYC, IT troubleshooting, on-call incident support, and claims.
Why it matters
Ringg's deployment shows a maturing pattern for enterprise voice agents: rather than leaning on one model, teams build routing layers that match each task to a variant tuned for latency, cost, or reasoning depth. If the reported 90% cost drop versus GPT-4.1 holds up outside vendor reporting, the economics of running real-time voice at high volume look considerably better than they did a year ago. And carrying context from voice to WhatsApp to browser without making customers repeat themselves addresses a long-standing friction point in support.
Our take
The reported 65% automation rate matters because insurance, healthcare, and investing are regulated, high-stakes domains where mistakes carry compliance risk. Ringg's evaluation loop — offline testing against historical conversations, simulated flows, and model-as-judge runs with GPT-5.6 Sol — is the kind of discipline many voice-agent rollouts skip. Extending into browser agents with computer-use could push automation beyond conversation into screen-guided workflows, but that surface brings prompt-injection and permission risks that will need explicit guardrails.
