Ringg’s AI Agents Resolve 65% of Customer Calls Using OpenAI

Overview

Ringg, an enterprise agent platform specializing in voice and chat solutions, has reported that its AI agents now resolve up to 65% of customer service calls without human intervention. By leveraging OpenAI’s latest language models, the company has achieved a significant reduction in operational costs while maintaining high customer satisfaction scores across Asia-Pacific and Oceania markets.

The platform now processes over 7 million connected calls each month, utilizing a hybrid architecture that routes tasks to specialized models based on real-time performance and economic efficiency. According to the company, migrating suitable real-time workloads from previous generations to the newer GPT-5.6 models has reduced model costs by approximately 90% for selected workloads.

This shift represents a broader trend in the technology sector toward autonomous, multilingual customer support systems that can handle complex, multi-step workflows across multiple channels, including voice, web chat, and WhatsApp.

Operational Scale and Customer Impact

Ringg developed its platform after observing firsthand the challenges faced by large consumer businesses in India. These enterprises were struggling not only with surging call volumes but also with fragmented, manual systems that hindered their agents’ ability to complete tasks such as booking appointments or processing insurance purchases. The resulting inefficiency drove up the cost and complexity of every customer interaction.

To address these pain points, Ringg built an enterprise-grade agent platform centered on high-efficiency models. The system is designed to span multiple communication channels, providing a unified interface for customers regardless of whether they prefer voice calls, text-based chat, or messaging apps like WhatsApp.

“Model quality is only part of the equation. We also need low latency, reliable tool use, strong instruction following, and economics that work at scale. OpenAI gave us the balance we needed.”

— Siddharth Tripathi, Co-founder, Ringg

The results of this architectural choice have been measurable. Ringg’s customers currently report an average customer satisfaction (CSAT) score of 4.8 out of 5.0. This high satisfaction rating coincides with the platform’s ability to handle a substantial volume of interactions, with more than 7 million connected calls processed monthly.

Architectural Approach and Model Routing

Building an agent platform around customer outcomes requires more than just answering questions; it involves executing actions. For a customer request to be fully resolved, an agent may need to check a policy, retrieve an account record, schedule an appointment, update a CRM, or transfer the conversation to a human specialist with full context.

Ringg’s orchestration layer handles these multi-step workflows by interpreting customer requests, selecting appropriate tools, and guiding users through complex processes. The system executes actions across various integrations, including CRMs, ticketing platforms, payment systems, and internal APIs. When a case exceeds the agent’s capabilities, it escalates to a human representative accompanied by a comprehensive conversation summary.

A key component of this architecture is its dynamic model routing system. Rather than relying on a single foundation model for all tasks, Ringg evaluates workloads and directs them to the most suitable OpenAI model. This strategy allows the company to optimize for a balance of conversational quality, latency, instruction following, tool-calling reliability, multilingual performance, and cost.

The routing logic operates as follows:

GPT-4.1 handles the majority of real-time voice and chat traffic, providing a stable and cost-effective baseline for standard interactions.
GPT-5.6 Luna is utilized in the production stack when its specific performance, latency, or price-performance profile better suits a particular request.
GPT-5.6 Terra is dedicated to post-call analysis, including generating summaries and classifying sentiment.
GPT-5.6 Sol supports internal evaluation workflows, prompt improvement, and model-as-judge tasks to ensure continuous quality control.

For longer interactions, the system employs a sophisticated context management strategy. When the conversation context approaches approximately 80,000 tokens, the orchestration layer creates a structured summary. This allows the conversation to continue seamlessly, preserving critical information without repeatedly transmitting the entire history, which helps manage costs and latency.

Production Evaluations and Reliability

Ringg places a heavy emphasis on rigorous testing before deploying any new model or configuration into production. The company uses a dedicated evaluation platform that tests models against historical conversations and simulated customer flows. These evaluations help identify weaknesses and recommend prompt improvements, creating a continuous feedback loop for agent performance.

In one significant evaluation, Ringg compared GPT-5.6 Terra against alternatives like Gemini 2.5 Flash for its post-call analysis workflow. Terra emerged as the superior option, maintaining high accuracy for summaries and sentiment analysis while materially improving unit economics. Consequently, Ringg migrated its summary and sentiment classification tasks to Terra.

Multilingual performance was another critical factor in Ringg’s model selection. The platform serves diverse markets where conversations frequently switch between English and local languages. Ringg tested models across various language and regional variations, including code-switching scenarios common in its operational regions. GPT-5.6 Terra demonstrated up to 97% accuracy on common regional languages, outperforming competitors and proving to be a better fit for agents serving Ringg’s multicultural customer base.

Infrastructure and Continuous Improvement

Once models pass offline testing, they are introduced to a small share of production traffic. This staged rollout allows Ringg to monitor real-world performance before expanding the user base. In production, the router continuously monitors latency and endpoint health across different regions. If an endpoint becomes unavailable or crosses a predefined latency threshold, the system automatically shifts traffic to healthier nodes.

This infrastructure includes specialized monitoring nodes, automated alerts, and versioned deployments. These features help isolate problems and limit their impact, ensuring that issues with one model or region do not cascade across the entire platform.

Ringg’s agents now handle more than 7 million connected calls each month, and its customers have an average customer satisfaction (CSAT) score of 4.8.

By combining advanced AI models with a robust orchestration layer and rigorous evaluation protocols, Ringg has created a system that scales efficiently. The company’s approach demonstrates that balancing model capability, latency, and cost is achievable at enterprise scale, setting a benchmark for AI-driven customer service operations in the Asia-Pacific region.