当呼叫量上升时，客户服务运营通常通过增加人手来扩展，但这会增加每次交互的成本和复杂性。Ringg（在新窗口中打开）是一个语音和聊天代理平台，在与印度大型消费企业合作时亲身遇到了这个问题。
这些公司不仅难以接听更多电话，其客户服务代理还在与碎片化的手动系统作斗争，以帮助客户完成从购买保险到预约等各种任务。
这段经历促使 Ringg 构建了一个企业代理平台，以 GPT‑5.6 等高效模型为核心，覆盖语音、聊天、WhatsApp 和网页。将合适的实时工作负载从 GPT‑4.1 迁移到 GPT‑5.6，使模型成本降低了约 90%，同时提供了所需的质量和延迟。
“模型质量只是等式的一部分。我们还需要低延迟、可靠的工具使用、强大的指令遵循能力，以及能够规模化运作的经济性。OpenAI 给了我们所需的平衡。”
——Siddharth Tripathi，Ringg 联合创始人
Ringg 的代理现在每月处理超过 700 万通已连接通话，其客户的平均客户满意度（CSAT）得分为 4.8。
围绕客户成果构建代理平台
代理要完成客户请求，可能需要检查保单、检索账户记录、安排预约、更新 CRM，或将对话连同相关上下文完整地转接给专家。
对于实时交互，Ringg 使用 GPT‑5.6 Luna 和其他模型来解释客户请求、选择工具，并引导客户完成多步骤工作流。Ringg 的编排层跨 CRM、工单平台、支付系统、调度工具和内部 API 执行操作，并在需要时连同对话摘要升级给人工处理。
Ringg 的知识系统结合了结构化过滤与跨数据集、PDF、CSV 和业务文档的语义检索，帮助代理使用相关企业信息回答问题。
Ringg 还可以在专门子代理之间分配工作，用于资格认定、支持、验证、调度和升级。该平台协调这些步骤，同时在语音、聊天、WhatsApp 和网页上保持与客户的一致对话。
将每项任务路由到合适的 OpenAI 模型
在从对话质量、延迟、指令遵循、工具调用、多语言性能、可靠性和成本等方面评估 OpenAI 与替代方案后，Ringg 发现 OpenAI 为生产工作负载提供了最强的整体平衡。
GPT‑5.6 和其他 OpenAI 模型在遵循复杂指令、执行工具调用、保持对话上下文以及在企业规模下可靠运行方面也表现得更加一致。
Ringg 根据任务需求将工作路由到特定的 OpenAI 模型。GPT‑4.1 处理大多数实时语音和聊天流量。GPT‑5.6 Luna 仍保留在生产栈中，并在其性能、延迟或性价比更适合某个请求时使用。GPT‑5.6 Terra 处理通话后分析，包括摘要和情感分类。GPT‑5.6 Sol 支持评估、提示改进和模型作为评判者的工作流。
当客户在语音通话中说话或发送聊天时，Ringg 的编排系统会将该输入与代理指令、对话历史、客户特定数据、相关知识库上下文和可用工具结合起来。Ringg 的路由层随后选择合适的模型和配置。模型的输出会通过编排系统传递，并通过客户选择的渠道交付。
对于较长的交互，当上下文接近约 80,000 个 token 时，系统会创建结构化摘要。随后对话可以在保留重要信息的情况下继续，而无需反复发送整个历史记录。
使用生产评估来提高质量和降低成本
Ringg 在生产部署前使用历史对话和模拟客户流程测试模型。其评估平台利用这些结果识别弱点并推荐提示改进，为其代理和配置创建持续改进循环。
其中一项评估涉及 Ringg 的通话后分析工作流。该公司将 GPT‑5.6 Terra 与 Gemini 2.5 Flash 等替代方案进行了测试，Terra 胜出。Ringg 将摘要和情感分类迁移到 GPT‑5.6 Terra。Terra 在摘要和情感分析方面保持了高准确率，同时显著改善了单位经济性。
该公司还跨语言和地区变体测试模型，包括在语言之间切换或将英语与本地语言短语结合的对话，这在其服务的市场中是常见行为。GPT‑5.6 Terra 优于 Gemini 2.5 Flash，在常见地区语言上准确率高达 97%，使 GPT‑5.6 更适合为 Ringg 客户提供服务的代理。
通过离线测试的模型会在推广扩展前先引入一小部分生产流量。在生产环境中，Ringg 的路由器会监控各地区的延迟和端点健康状况，并在某个端点不可用或超过延迟阈值时转移流量。专用节点、警报和版本化部署有助于隔离问题并限制其影响。
这种方法也帮助 Ringg 改善单位经济性。将某些实时工作负载从 GPT‑4.1 迁移到 GPT‑5.6 Luna，使模型成本降低了约 90%。
“当我们需要模型迁移支持时，OpenAI 一直非常响应迅速。专用的 Slack 支持和对核心工程团队的访问，帮助我们以更少的工程不确定性快速推进、更快交付，并扩展我们的代理能够完成的任务。”
——Siddharth Tripathi，Ringg 联合创始人
跨行业交付客户成果
对其客户而言，Ringg 的代理可在无需人工代理参与的情况下解决多达 65% 的常规客户咨询。
Policybazaar 是印度最大的在线保险平台之一，使用 Ringg 连接了超过 57,000 个客户请求，其中 67% 的通话无需人工干预即可处理。Policybazaar 的平均响应时间从 8–12 分钟降至 60 秒以内，改善了约 88%。
在全球医疗平台 Practo，Ringg 的代理帮助客户预约医疗保健服务并完成服务请求。部署后，Practo 实现了 85% 的首次呼叫解决率和低于三秒的响应时间。与之前以人工为主导的工作流相比，运营成本下降了 70%，而 Ringg 现在每天完成超过 1,000 次预约。
在线投资平台Groww利用Ringg，通过自助服务完全解决了72%与IPO、期货和期权相关的入境查询，平均处理时间为两分钟。
将代理扩展至浏览器
借助OpenAI的计算机使用能力，Ringg正在开发用于平台注册、了解您的客户（KYC）流程、IT故障排除、待命事件支持和索赔处理的浏览器代理。
该公司还在开发一个上下文层，以便跨渠道和交互保留信息。客户可能通过语音开始请求，在WhatsApp上继续，并在浏览器中完成，而无需每次重复相同的细节。
“OpenAI的计算机使用能力加速了我们的浏览器代理路线图。它们让我们能够将用户屏幕上发生的事情与对话上下文结合起来，使代理能够实时引导人们完成复杂的工作流程。”
—Siddharth Tripathi，Ringg联合创始人
对于Ringg而言，下一代客户运营将以完成的业务成果和自动化深度来衡量，而不是通话量或员工人数。
When call volume rises, customer service operations typically scale by adding people, but this increases the cost and complexity of every interaction. Ringg(opens in a new window), a voice and chat agent platform, saw this problem firsthand while working with large consumer businesses in India.
These companies were not only struggling to answer more calls, their customer service agents were also battling fragmented, manual systems to help customers complete tasks, from buying insurance to booking appointments.
That experience led Ringg to build an enterprise agent platform with high-efficiency models like GPT‑5.6 at its core, spanning voice, chat, WhatsApp, and the web. Migrating suitable real-time workloads from GPT‑4.1 to GPT‑5.6 reduced model costs by approximately 90% while delivering the required quality and latency.
“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
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.
Building an agent platform around customer outcomes
For agents to complete a customer request, it may require checking a policy, retrieving an account record, scheduling an appointment, updating a CRM, or transferring the conversation to a specialist with the relevant context intact.
For real-time interactions, Ringg uses GPT‑5.6 Luna and other models to interpret customer requests, select tools, and guide customers through multi-step workflows. Ringg’s orchestration layer executes actions across CRMs, ticketing platforms, payment systems, scheduling tools, and internal APIs, escalating cases to a human with a conversation summary when needed.
Ringg’s knowledge system combines structured filtering with semantic retrieval across datasets, PDFs, CSVs, and business documents, helping agents answer questions using relevant enterprise information.
Ringg can also divide work among specialized subagents for qualification, support, verification, scheduling, and escalation. The platform coordinates those steps while maintaining a consistent conversation with the customer across voice, chat, WhatsApp, and the web.
Routing each task to the right OpenAI model
After evaluating OpenAI alongside alternatives across conversational quality, latency, instruction following, tool calling, multilingual performance, reliability, and cost, Ringg found that OpenAI delivered the strongest overall balance for production workloads.
GPT‑5.6 and other OpenAI models also performed more consistently in following complex instructions, executing tool calls, maintaining conversational context, and operating reliably at enterprise scale.
Ringg routes work to a specific OpenAI model based on the needs of the task. GPT‑4.1 handles most real-time voice and chat traffic. GPT‑5.6 Luna remains in the production stack and is used when its performance, latency, or price-performance profile better suits a request. GPT‑5.6 Terra handles post-call analysis, including summaries and sentiment classification. And GPT‑5.6 Sol supports evaluation, prompt improvement, and model-as-judge workflows.
When a customer speaks during a voice call or sends a chat, Ringg’s orchestration system combines that input with the agent’s instructions, conversation history, customer-specific data, relevant knowledge-base context, and available tools. Ringg’s routing layer then selects the appropriate model and configuration. The model’s output is passed through the orchestration system and delivered through the customer’s chosen channel.
For longer interactions, the system creates a structured summary when the context approaches approximately 80,000 tokens. The conversation can then continue with the important information preserved, without repeatedly sending the entire history.
Using production evals to improve quality and cost
Ringg tests models using historical conversations and simulated customer flows before production deployment. Its evaluation platform uses those results to identify weaknesses and recommend prompt improvements, creating a continuous improvement loop for its agents and configurations.
One evaluation involved Ringg’s post-call analysis workflow. The company tested GPT‑5.6 Terra against alternatives like Gemini 2.5 Flash, and Terra came out on top. Ringg moved summaries and sentiment classification to GPT‑5.6 Terra. Terra maintained high accuracy for summaries and sentiment analysis while materially improving unit economics.
The company also tests models across language and regional variations, including conversations that switch between languages or combine English with local-language phrases, common behavior in the markets it serves. GPT‑5.6 Terra outperformed Gemini 2.5 Flash, with up to 97% accuracy on common regional languages, making GPT‑5.6 a better fit for agents serving Ringg’s customers.
Models that pass offline testing are introduced to a small share of production traffic before rollout expands. In production, Ringg’s router monitors latency and endpoint health across regions and shifts traffic when an endpoint becomes unavailable or crosses a latency threshold. Specialized nodes, alerts, and versioned deployments help isolate problems and limit their impact.
This approach also helps Ringg improve unit economics. Migrating certain real-time workloads from GPT‑4.1 to GPT‑5.6 Luna reduced model costs by approximately 90%.
“OpenAI has been highly responsive when we’ve needed support with model migrations. The dedicated Slack support and access to the core engineering team help us move quickly with less engineering uncertainty, ship faster, and expand what our agents can accomplish.”
—Siddharth Tripathi, Co-founder, Ringg
Delivering customer results across industries
For its customers, Ringg’s agents resolve up to 65% of routine customer inquiries without any involvement from a human agent.
Policybazaar, one of India’s largest online insurance platforms, uses Ringg to connect more than 57,000 customer requests, and 67% of calls are handled without human intervention. Policybazaar’s average response time fell from 8–12 minutes to under 60 seconds, an improvement of approximately 88%.
At Practo, a global healthcare platform, Ringg’s agents help customers book healthcare appointments and complete service requests. Following deployment, Practo achieved an 85% first-call resolution rate and response times below three seconds. Operating costs declined by 70% compared with its previous human-led workflow, while Ringg now completes more than 1,000 appointment bookings each day.
Online investment platform Groww uses Ringg to resolve 72% of inbound queries related to IPOs, futures, and options entirely through self-service, with an average handling time of two minutes.
Extending agents into the browser
Using OpenAI’s computer-use capabilities, Ringg is developing browser agents for platform onboarding, Know Your Customer (KYC) processes, IT troubleshooting, on-call incident support, and claims processing.
The company is also developing a context layer that can preserve information across channels and interactions. A customer might begin a request over voice, continue it on WhatsApp, and finish it in a browser without having to repeat the same details each time.
“OpenAI’s computer-use capabilities accelerated our browser-agent roadmap. They let us combine what is happening on a user’s screen with conversational context, so agents can guide people through complex workflows in real time.”
—Siddharth Tripathi, Co-founder, Ringg
For Ringg, the next generation of customer operations will be measured by completed business outcomes and automation depth, rather than call volume or headcount.
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文