Improving support with every interaction at OpenAI

I would argue support has never really been about replying to just tickets. So, we're responsible for providing support for all of our customers, all of our products globally. Usually, when you have a problem, you want to get help as quickly as possible. And by using our AI tools, we're able to get those responses much more quickly. Tens of thousands of people every single day, millions of people on our website. Check in and try to get help with ChatBT and OpenAI products. If you come into our help center, there's a chatbot there where you'll get connected directly with some of the tools that we build. We use the responses API to generate responses that go to customers and the agents SDK that helps us do that tactically. All of that is plumbed right through into our dashboard in the OpenAI platform. And our human agents and support specialists know exactly what great support looks like. And so we're building systems that allow them to one flag both good and bad examples that help us steer the AI to do more of that and two they're also participating a ton in the labeling process so that we can take one great example flag by a support specialist and scale that to hundreds and thousands of conversations in the future. — What an agent is truly interacting also with is a knowledge base of like information that they need to reference to be able to respond to a ticket and a set of policies. Increasingly, we're finding an opportunity where agents can actually be part of informing our knowledge base and our policies. — We want our people in the loop. We want them um not just consumers of these things, but participating in it cuz they're on the ground floor like seeing the work and likely have expertise to make things better. So how do we architect this solution so that if me as a tier three support rep I see something that isn't quite going right I can start experimenting with a new way of running an eval or I can propose a new classifier that could you know change how we work and not just propose it but potentially ship it and test it. When we have to think about scaling, we're not about just thinking about like scaling the support agent workforce, but also our own tech so that we can build the right foundations so that when a new product comes along, a new language comes along, we're able to quickly expand without needing to rebuild things from scratch. — There's plenty of signal both in terms of like hard objective metrics, but I think the more interesting part is um you know, it feels as though we're on the right path when it comes to how we're doing it as well.