Most companies treat support like a factory floor: tickets in, replies out, and success measured by throughput. OpenAI decided that model was broken. In a rare look under the hood, the company detailed how it rebuilt its entire support operation around a self-improving AI loop, turning millions of annual requests from hundreds of millions of users into fuel for a continuously learning system. The goal, as the Ops team put it, was to go well beyond a simple deflection chatbot and reimagine support as an engineering and operational design challenge. “Support has never really been about replying to just tickets. It’s about whether people get what they need,” the team noted.
At the core of the new model are three building blocks that form a relentless feedback loop. There are the surfaces where people ask for help—chat, email, phone, and increasingly, support embedded directly inside the product itself. Then there is the knowledge layer, which isn’t a dusty wiki but a living body of guidance drawn from real conversations and continuously updated. Finally, shared evals and classifiers define what a great interaction looks like, blending human judgment with software to steer model behavior across every channel automatically. A pattern caught in one enterprise chat can instantly strengthen a developer FAQ. An eval written for a tricky billing question improves the model for thousands of future cases.
This architecture completely changes the job of a human support rep. Instead of just processing transactional work, reps are now builders who actively improve the machinery. They flag interactions that should become test cases, propose new classifiers when they spot novel patterns, and even ship lightweight automations to close workflow gaps in just days. “Agents aren’t just responding to tickets. They’re informing our knowledge base and our policies,” a team member explained. Jay Patel, a Software Engineer on Support Automation, highlighted a critical safety net: “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.”
The whole system sits on OpenAI’s own stack, which the company says eliminates the usual plumbing headaches. The Agents SDK provides step-level traces so engineers can instantly replay and debug model runs. The Responses API powers classifiers for tone and policy adherence, while the Realtime API enables voice support. The internal Evals dashboard then turns everyday conversations into production tests, codifying what “great” means—solving the issue politely, clearly, and consistently. Because the platform primitives are ready-made, the team spends less time stitching tools together and more time defining and measuring quality. As models gain larger context windows and stronger agentic capabilities, those advances can be adopted immediately, creating a support organization defined not by how many tickets it closes, but by its capacity to evolve.
