Improving support with every interaction at OpenAI

这是我们关于 OpenAI 如何在自身技术上构建解决方案系列的一部分。

不仅仅是工单，而是一种新的运营模式

传统上，支持意味着排队、工单和吞吐量。但在 OpenAI，这还不够。我们为数亿用户提供服务，每年处理数百万次请求，而且这种规模每年以几倍速增长。

很多组织能应对规模问题，但很少能同时应对规模与超高速增长。几乎没有组织在面临这两者的同时，还在构建可能改变这一局面的核心技术。这种组合让我们有独特条件，从零开始重新思考支持体系。

“支持从来不只是回复工单。关键是人们是否得到了所需的帮助，这些帮助是否真正对他们有用。”
Glen Worthington，用户运营主管

支持不是一个单纯的量化挑战，而是一个工程与运营设计挑战。因此我们打造了不同的东西：一种让每次交互都能改进下一次交互的运营模型。

把交互系统连成一体

Ops 团队不满足于仅用聊天机器人来转移支持问题。团队的愿景是：将支持重新想象为一种持续学习并不断改进的 AI 运营模式。

核心由三块构建模块组成：

• 界面（Surfaces）。用户与支持系统交互的地方——聊天、邮件、电话，且越来越多的是直接内嵌在产品中的帮助。
• 知识（Knowledge）。不仅是静态文档，而是从真实对话、政策和上下文中抽取、持续演进的活知识库。
• 评估与分类器（Evals and classifiers）。软件与人工共同构建的质量共识，以及用于衡量、改进和呈现反馈的工具。

这些模块不是孤立存在的。它们形成一个循环。企业对话中发现的模式可以成为开发者常见问题的来源。为某个场景编写的评估能增强模型在成千上万类似场景下的表现。而且因为相同的原语驱动着所有界面——聊天、邮件、语音——改进会自动跨渠道扩展。

支持人员成为系统思考者

支持人员的角色正在变化。我们的目标是把模型从主要关注处理事务性工作，转变为成为整体构建的一部分。他们被授权直接通过自下而上的变更交付影响架构，也通过日常工作自然推动改进。

支持人员会标注应成为测试用例的交互，发现新模式时提出并部署分类器，甚至能在几天内原型化轻量自动化以弥补流程缺口。培训也随之转变，不再只是学习政策，而是学会评估交互、识别结构性缺陷，并将改进反馈回系统。

新的方法力求让支持人员既是回应者，也是构建者。

“坐席不仅仅是在回复工单，他们是在为我们的知识库和政策提供信息。他们对一线情况的敏感度是我们所没有的。”
Shimul Sachdeva，工程经理

由此形成的支持组织不再以吞吐量定义，而以演进能力定义。每个人不仅在为用户服务，同时也在积极改进服务所有用户的机器。

从原语到生产化

以这种方式构建支持只有在我们基于 OpenAI 的技术栈时才成为可能。

• Agents SDK 为我们默认提供逐步跟踪和可观测性。我们可以重放执行、检查工具调用，并即时调试根因。
• Responses API 支持对语气、正确性和政策遵守性的分类器。
• Realtime API 使语音支持成为可能。
• OpenAI 的 Evals 仪表盘让质量可测并易于随时间可视化。

因为这些平台原语是现成的，我们花更少时间去拼接系统，而能把精力放在更重要的事情上：定义什么是优质、去衡量它并进行改进。

我们从一个简单的问答回答器开始并运行良好。借助 Agents SDK，我们迅速扩展到可执行的动态操作，例如退款、发票、事故查询等。随着模型在更大上下文窗口、更深入的研究（Deep Research）和更强的能动性方面持续改进，我们能够立即采用这些进步。

复利式的学习

Evals 将日常对话转化为生产测试。它们将“优秀”进行编码——不仅是解决问题，还要礼貌、清晰且一致。支持人员在其中发挥直接作用，标注正反例成为 evals，这些 evals 在生产环境中持续运行以引导模型行为。

“通常当你遇到问题时，你只想尽快得到帮助。通过使用我们的 AI 工具，我们能更快给出回复——同样重要的是，我们也知道何时模型不应该回答。”
Jay Patel，支持自动化软件工程师

学习并不止于问题解决。模式会反馈到知识库、自动化和产品设计中。系统呈现复利效果：用户获得更快的答案，构建者获得更紧密的反馈环，所有界面的质量门槛持续提高。

不仅 AI 在学习，组织也在同步学习。专家们会发现模型的短板、定义新的分类器并贡献用于微调的数据集。可观测性仪表盘让质量可被度量，展示性能随时间的提升。

面向支持未来的蓝图

最深远的变化不是工具，而是人以及组织如何衡量成功。支持专家的价值不再只是解决问题，而是精炼知识、改进模型并扩展系统本身。领导者开始寻找一种新型队友：将一线共情与设计直觉结合，把支持技巧与改进系统的好奇心融合。

“我们开始看到深厚的工艺专长与深厚的工程专长相结合。这就是部门运作的未来。”
Glen Worthington，用户运营主管

我们的愿景是：支持不再是你需要去到的一个终点，而是一种行动，编织进每一个产品界面。用户不再“提交工单”，他们在所在位置就得到所需帮助。

起初这是对规模问题的回应，如今已成为人类与 AI 如何协作的蓝图：协作、适应并持续改进。

准备好在你的业务中使用 ChatGPT 吗？联系我们 (https://openai.com/contact-sales/)

This is part of our series about how OpenAI is building its own solutions on our technology.

More than tickets, a new operating model

Support has historically meant queues, tickets, and throughput. But at OpenAI, that wasn’t enough. We serve hundreds of millions of users, handle millions of requests each year, and see that volume grow by multiples annually.

Lots of organizations deal with scale. Fewer deal with scale and hypergrowth. Almost none face both—while also building the very technology that could change the equation. That combination uniquely positioned us to rethink support from the ground up.

“Support has never really been about replying to just tickets. It’s about whether people get what they need, whether it actually serves them well.”

Support isn’t a volume challenge. It’s an engineering and operational design challenge. So we built something different: an operating model where every interaction improves the next.

Connecting a system of interactions

The Ops team wanted to go well beyond using a chatbot to deflect support questions. The team has a vision: reimagine support as an AI operating model that continuously learns and improves.

At the center are three building blocks:

• Surfaces. Where support systems are interacted with. Chat, email, and phone, but increasingly, help embedded directly inside the product.
• Knowledge. Not just static docs, but living and continuously improving guidance drawn from real conversations, policies, and context.
• Evals and classifiers. Shared definitions of quality built by software and humans in unison, plus tools to measure, improve, and highlight feedback.

These pieces don’t sit in isolation. They form a loop. A pattern spotted in an enterprise conversation can inform a developer FAQ. An eval written for one case strengthens the model for thousands more. And because the same primitives power every surface - chat, email, voice—improvements scale across channels automatically.

Support reps as systems thinkers 

The role of a support rep is changing. Our aim is to shift the model from primarily focusing on processing transactional work to being a part of the overall build. They’re empowered to contribute to the architecture itself, both directly through bottom up shipping of changes and indirectly through the natural motions of their day to day work.

Reps flag interactions that should become test cases, propose and ship classifiers when they see new patterns, and even prototype lightweight automations to close workflow gaps in days. Training shifts too, it’s not just about policies, but about evaluating interactions, identifying structural gaps, and feeding improvements back.

The new approach strives to ensure that support reps are builders as much as responders.

“Agents aren’t just responding to tickets. They’re informing our knowledge base and our policies. They have an ear to the ground that we don’t.”

The result is a support organization defined less by throughput and more by its capacity to evolve. Every person is not only serving users but also actively improving the machinery that serves all users.

From primitives to production

Building support this way is only possible because we’re built on OpenAI’s stack.

• Agents SDK gives us step-level traces and observability by default. We can replay runs, inspect tool calls, and debug root causes instantly.
• Responses API powers classifiers for tone, correctness, and policy adherence.
• Realtime API makes voice support possible.
• OpenAI’s Evals dashboard makes quality measurable and easy to visualize over time.

Because the platform primitives come ready-made, we spend less time stitching systems together and more time focusing on the work that matters: defining what good looks like, measuring it, and improving it.

We started with a simple Q&A answerer that worked well. With Agents SDK, we quickly expanded into dynamic actions for things like refunds, invoices, incident lookups. As the models continue to improve with larger context windows, Deep Research, and stronger agentic capabilities, we can adopt those advances immediately.

Learning that compounds

Evals turn everyday conversations into production tests. They codify what “great” means—not just solving the issue, but doing it politely, clearly, and consistently. Reps play a direct role here, flagging strong and weak examples that become evals, and those evals run continuously in production to steer model behavior.

“Usually when you have a problem, you just want help as quickly as possible. By using our AI tools, we’re able to get those responses much more quickly—and just as importantly, we know when the model shouldn’t answer,” says Jay Patel, Software Engineer, Support Automation.

Learning doesn’t stop at resolution. Patterns feed back into knowledge, automation, and product design. The system compounds: faster answers for users, tighter feedback loops for builders, and a consistently higher bar for quality across every surface.

And it’s not just the AI that learns. The organization learns alongside it. Specialists see where models fall short, shape new classifiers, and contribute datasets for fine-tuning. Observability dashboards make quality measurable, showing how performance improves over time.

A blueprint for the future of support

The most profound shift isn’t the tooling, it’s the people and how the organization measures success. Support specialists are recognized not just for solving problems, but for refining knowledge, improving models, and extending the system itself. Leaders look for a new kind of teammate: someone who pairs frontline empathy with design instincts, combining support craft with curiosity to improve the system.

“We’re starting to see this marriage between deep craft expertise and deep engineering expertise. That’s the future of how departments run.”

And our vision is support stops being a destination you go to. It becomes an action, woven into every product surface. Users don’t “open a ticket.” They simply get what they need, where they are.

What began as a response to scale has become a blueprint for how people and AI can work together: collaborative, adaptive, and continuously improving.

Ready to put ChatGPT to work in your business?

Generated by RSStT. The copyright belongs to the original author.

Source