OpenAI on OpenAI: Applying AI to Our Own Workflows

Hey, everyone. How's it going? How are we feeling? Do we like the keynote? Good. My name is Scotty. Super excited to be here today. You just heard on the main stage about new features that are going to make it easier to deploy agent-driven workflows. In the next 20 minutes, we're going to talk about how we put that technology to work inside OpenAI. I'm joining from the go-to-market innovation team. We are a team of developers that are responsible for deploying our technology inside our go-to-market organization. So a quick show of hands, and we got some people filtering in in the back. A quick show of hands here. Who here is building a product that they plan on selling to customers? That's sort of externally facing, right? You expect maybe like most of the hands in here. Today, we're going to actually focus on a slightly different surface where there's really never been a better opportunity to build, and that's internal applications. Internal applications are sort of like the glue that can make your company more effective. And today, in the world of internal app development, most teams tend to think about how can they streamline operations? And the question that people are asking in that space is, how do we use AI to drive efficiency? This is a question we hear all the time. And we think this is part of the right answer, or the right question. But there's a bigger question. And we've been trying a different frame internally to really capture this opportunity. So we've been asking, How do we use AI to amplify expertise? So let me explain what I mean. Every company has embedded expertise. You have your top salesperson, Sophie, who knows just how to connect with customers. You have your support specialist, Ken, who knows how to untangle even the most complex system. You have maybe your engineer, Alex, who can just crank out code and release features in days. The promise of AI is not only automation. It's capturing the craft of these practitioners and then finding a way to encode and distribute it across your entire organization so that every employee in your organization can operate like your best operator. This opportunity is driving what we call a golden age of internal building. And as a developer, it's also incredibly fun work. You get to pull up a chair next to your top operator, understand their work, iterate in fast loops, and drive real value in days. The right builds internally can 10x your company's agility and impact. And today, We're going to talk about three real examples that have hit that mark internally here at OpenAI across sales, people, and support. So each of the sort of internal products that we talk about today are only being used inside OpenAI. But you should think about them as illustrative examples of how you can deploy frontier AI to solve real problems. We're going to walk through what made each of these use cases successful and some patterns that you can take back to your companies. And we're going to start today with a use case that's very near and dear to my heart coming from the go-to-market organization, and that's the go-to-market assistant. So let's talk a little bit about sales at OpenAI. I bet everyone here can imagine being on the sales team at OpenAI can be really intense. You got new customers, new features, everything you saw live on stage today, bringing that to customers. And so last year, about this time, our teams were really hitting a breaking point. And what they needed was quicker customer research and technical answers from questions from customers, while also expanding their capacity and maintaining a great customer experience. These two things are very much intention, to do them both well. And so I started by sitting down with our very best rep. Her name's Sophie. Sophie is amazing, a real person also. Even during that period last year, Sophie just had this incredible way, even with extremely unscalable workflows, of finding a way to win with customers. And so I started just sitting down with Sophie, getting to know her. She had joined the company about the same time that I had. She came from a startup, so she kind of had this scrappy building spirit. And through our conversations, I learned that Sophie had this sort of special formula to scaling her craft. And she showed me exactly how she wins across prepping for customer meetings, generating product champions, thinking about product demos across our API and Chat2T, and then following up as quick as she possibly could to meetings with customers. So I really focused on this first workflow, which was prepping for customer meetings. And there was a really clear definition of success. And we built an agent that was specialized to just Sophie's version of excellence. And then we tried to repeat that for the rest of these skills. And what we found is we were able to build a system. And when you put that system to work inside of Sophie's daily work across systems she already uses, so ChatGPT and Slack primarily, we were able to create this great experience that you're going to see in just a moment. And the result was GoToMarket Assistant. GoToMarket Assistant is sort of this collection of workflows that enable every single one of our reps to bring their very best to every single customer engagement. And so let's get into the nuts and bolts. What you're going to see here is a familiar slide. We're going to come back to this slide three times today for each of our use cases. Also, headline for people who are joining us in the back. We're talking about the way they're deploying our products internally. And first, we're starting with sales. And so we're talking about the GoToMarket Assistant here. We have our kind of foundation. And really what it starts with is getting your data right. And so down here, we have customer data. Anybody who's worked on a go-to-market team before knows that understanding your customers is like priority number one. But doing that is very hard because the data is everywhere. It's spread out across systems. And so what was important here to start was to build a simplified data model and then to set up a semantic layer where GPT-5 could really understand what our customer is and how it can provide good data on that customers back to our reps. We vectorized a bunch of key documents here, including some internal resources to talk about our go-to-market strategy. And then you have this core kind of agent spine across Agents SDK, GPT-5, and the Responses API. And what you see inside of here are the skills that we started with. And we call these skills, because these are agents that are specified to this specific version of excellence, again, that came from those conversations with Sophie. We have meetings, product knowledge, custom demos, customer research. We started with four. We have about 10 live today. And finally, you have self-improvement on the side here, which is really driving that continual improvement workflow, really driving that high bar of success. And finally, services is where you actually distribute your product. And so we have ChatDBT here, which is enabled via an MCP connector, Slack, which we're going to look at in just a moment, and finally, the OpenAI platform where we're dealing with some of these kind of back-end task executions. So now, let's imagine that I'm a rep on the mid-market team, just joined, great job, fun team to be on. This is the same team that Sophie was on right when we started building this product about a year ago. On the mid-market team, you might have hundreds of customers. And you might in a single day have eight customer meetings. So it can be a really challenging job to bring your best to every one of those customer engagements. Now, I just start my day with a go-to-market assistant. And so I'm going to give everyone a quick orientation here. We're looking at a Slack workspace right now. We're here inside a private channel, which is just me and my go-to-market assistant. I just received my Tuesday briefing. So let's imagine it's Tuesday morning. We have Acne, Brookfield, and Redwood. Looks like I got three meetings, and they're all Quite close, like hopefully can fit in a lunch here. We have one first call, meaning it's the first time I'm talking to the customer, two follow-ups, and a little headline of information I should pay attention to. But the real beef of the information here is in the thread, where we get this detailed meeting prep doc. I have some attendees I can quickly click into LinkedIn. This was surfaced via web browse. We have our OpenAI data. And this is where it's really As a rep, I'd have to go to multiple dashboards to understand the state of this customer and what it means and what I should bring today. Here it's actually surfacing for this customer I'm just starting to meet that they're already starting to use our products in the self-serve world. And that's really important information I can bring to this conversation. Finally, we have hypotheses and opportunities. Obviously, my role as a seller is to be very strategic in every conversation. And what my assistant is helping me with here is distilling all of that information to help me show up prepared to this customer conversation. In that recommendation, it had recommended that I create a demo for AcneSearch focused on coding and deep research. So I actually just invoked that here. And it's coming back with a step-by-step script that I can use with the customer. This one's, again, interesting. Because this actually represents one of our best solutions engineers, whose name is Max. Max has been at the company about two years. He has over 100 different demos for ChatGPT that have really landed with customers. And that's sort of the grounding data set here that we've now remixed with this customer's data to create this really useful demo. And so what this allows us to do is to provide a great customer engagement for every customer, no matter the capacity of our technical teams. And so I can take this prompt right here, drop it into ChatGPT. That would render a dynamic web page where we could really walk through the customer this real coding example. I'm going to scroll down. Let's imagine I just got a meeting recap. So that meeting just finished about 30 minutes ago. Now I have some key takeaways here. We have some action items. And it looks like the customer actually asked a question on the call. It's sort of a technical question here about chat completions and the responses API. This is great for me as a rep. Now I know I can be tracking follow-ups inside this thread. But even better is the assistant is actually recognizing this question and sending it directly to our product knowledge skill, which is prepping this answer I can now take directly to the customer. So it's closing out that loop and allowing me to scale my support quicker. But as I'm reading this recap here, I actually noticed that it missed a key next step. It missed the next step. that the customer asked for a prompt tuning workshop for GPT-5. And so my expertise as a seller here is I know exactly what happened in this meeting. This is helping me scale my work. But I also know these things are not 100%. They might be 90%, 95%. They're getting me most of the way there. I'm still using my subject matter expertise to maybe recognize places where it's off. And then I can go down here to say, was it helpful? Click no. I can provide that feedback that it missed this key takeaway. This is going to prompt a regeneration on the back end for this particular task. And so a moment here, I'm going to get my updated, drafted email, which now, fundamentally or critically, is going to have this next step about this PropTuning workshop. Now, this is obviously great for me as a rep. Now I got what I need in my flow of work to support this customer. But even better is this same piece of feedback went to our evals platform and triggered a prompt optimization flow. And that prompt optimization insight was brought directly to our developers inside of a developer-only channel. And now I'm reviewing the scope and impact of this change. And as a developer, I can approve it directly in my flow of work. So what we've done here is we've actually had a top operator, myself in this case, train the system. And that piece of training then gets distributed to the entire organization to really up-level that skill over time. So people trust the go-to-market assistant because they help define it, and they're helping to train it every single day. And we have about 10 skills or so in go-to-market assistant today. And when I look at each of those skills, I see people. I see people from Maggie's team back there who helped build one of our product assistant knowledge skills. I see Max who helped build our custom demo workflow. I see Sophie who built the meetings prep. So those little pieces of excellence inside of our go-to-market org now are distributed to every one of our more than 400 members of the team. Today, our reps are exchanging about 20 messages every single week with the assistant. And they've reported one full day in time saved they can now spend on higher leverage tasks. I asked you, what would you do with an extra full day each week? That's what these kind of products, through very quick revolutions, can drive for your business. And as a final note, as a developer, this has also been a very fun project because the go-to-market team, they have this eagerness to build, this creativity. And when you give them the tools, they can really engage with those tools to scale their best work. So we just got through our first use case. For people who maybe showed up late, we're talking about internal applications of our technology here at OpenAI. We just covered sales, and now we're going to talk about people in HR. So in the people in HR space, we also are dealing with a challenge of scale. We've been rapidly adding employees across the globe, and you have new employees, you've got new offices, new policies, and it's extremely challenging as an employee to pick up on all the institutional context you need to be successful in your role. And so through that massive growth, how do you help employees quickly understand how the company operates? OpenHouse is our solution, and it's built through a very similar structure of connectors, orchestration, and surfaces. But the difference is now we have people and HR systems at the core of this agentic framework. So at the bottom here, you have HR systems, things like Workday, where you're connecting personnel records and who's in each role, what the expectations of their role is. You have some vectorized documents here. And I want to focus on the CMS here. This was sort of trying to solve the problem of you have all these events happening across your company, things like announcements, new updates to maybe key policies. And it's really easy to get kind of lost in the shuffle, right? So the CMS is actually capturing each of those little bits of information happening across our organization. And it's storing them. so that I can easily access them inside of a single system. And we're going to see that in just a second. We have the same spine in the middle, but now different skills. So we have our company knowledge, our people connector. We have career growth, which kind of helps me understand my expectations for my role. And then self-improvement, driven by the same kind of evals and guardrails bit. And finally, surfaces, where we're trying to build this directly into our team's flow of work. with ChatGPT, Slack, where it's added as a Q&A box that can respond in some of these global company channels. And finally, ChatKit, which we're going to look at right now. So we are looking now at OpenHouse. I'm going to walk everyone through what OpenHouse is and then talk about how I used it last month. So on the left here, we have directory, announcements, videos, resources, and support by team. These are all kind of sub-widgets where I can explore some more information. In the middle here, you have in case you missed it. This is where those kind of announcements get bubbled up, right? Slack can kind of be chaos. This is where you go to know the pieces of information you really need to see that are relevant to your role. And then finally, in the middle here, we have this common kind of chat format, which we're going to use right now. About a month ago, I was taking a trip to New York. I was trying to visit a customer for an on-site. And for any new employee who's trying to travel to a new office, the first thing you want to know is, how can I access the office? What's the travel policy? And how can I make this trip successful? So previously, you might message a few of your coworkers, hope you find the office manager. And maybe worst case scenario, you book something that's out of policy, and then your expense report gets denied. Bad outcome, right? Now you can just start with Open House. And so I took that question directly to Open House, and I asked, hey, I'm visiting New York office. Can you help me with the travel policy, how to access the office? I get this kind of summarized chain of thought here, and then I get this nice sort of chat-ready answer, obviously very familiar format to people who are using chat daily. We have our citations, though, here, too, and I'm going to actually click into our office guidebook here over on the right. This is sort of maintained by our office manager. We get all the details here, including where I should sit, how to access the office. Travel policy was back on that main page. And I could actually click into our Slack channel right here, and I would see what's for lunch. So first off, now I can access the office. Obviously, that's great. Probably could have done that just finding the right file. What's really, though, taking this one step further is I had this customer demo. I was meeting with a customer, talking about go-to-market use cases for our technology, obviously a bit familiar. And so I could take that same question right here. And you can see me asking now, is anyone on my team in New York that could help me build a great use case for go-to-market use of our technology? Now look primarily at our directory. So this mapping of kind of all of our employees and these like internal wikis. And it's going to bring to me, I got five different profiles here. And I'm going to click into Joe's. It's highlighting Joe's might be relevant here. I can go into the sources on the right here. And now I see Joe's kind of personalized profile. Every employee at OpenAI is actually filling these out as a part of their new hire onboarding now. We're also kind of pre-populating with some of those kind of HR and people systems data. And so it looks like Joe actually has some expertise in building customer demos. And so what I can do here is I can just click down on the Slack button here to the left, quickly jump into Slack, send Joe a message, say that I'm visiting New York next week, see details right here for the demo I'm trying to drive. So that open house suggested that I was a good person to meet with, asked him if he has time, and he could respond directly in thread. And it looks like he's there. So what happened here is that I was able to tune in to the institutional knowledge that exists across the entire organization. Usually I sort of locked in these silos and what Open House allowed me to do is ask the question, get an answer, connect with a co-worker, and then bring that directly to the customer. Our teams have also been engaging through these same patterns and about 75% of our employees are using Open House every single week. And the reason that it's so high is that it's built on good data. It was designed around real questions, and it's very easy to use. OK, we got through two. We're at the third demo, or the third use case today. We're going to talk about support. Support is this world of massive data and sort of operational complexity. And at OpenAI, we have hundreds of millions of users and millions of support ticket questions every single year with a relentless rate of change. Now, many companies face some elements of scale. Far fewer face maybe scale at the same time as hypergrowth. And hardly any do that while also building a technology that can change the equation. And so we tried to tackle support for more of an engineering and design operations challenge to think about how we could build a model where every interaction improves the next. And so here's a concrete example of the scale we're talking about. You're looking at service tickets over time. And in the middle here, you see the launch of ImageGen. People remember that from a few months ago. We had several magnitudes our normal ticket volume in just a few days. And over 100 million users added in just a few days. Obviously, anyone on the growth team is looking at this being like, great, five star. This is exactly what you want to see. But from a customer support standpoint, this is really unworkable, and it's unscalable with any sort of traditional model. So we started by defining the process. We reviewed all of our conversation logs with specialists, or maybe not all of them. We picked thousands of conversations to review with specialists. We defined gold standards for how should an agent behave here. When should it respond to the customer quickly? When should it escalate for support? When should it tag for added complexity? We then codified all of those standards into knowledge. And we call our knowledge in the support world SOPs, or sort of standardized operating procedures. Those are like an industry standard term. And then finally, we took our SOPs and connected them to knowledge and evals so that we could drive this kind of self-improving loop where every interaction improves the next. See a lot of nodding heads. This is a familiar structure here. Now, we're going to talk about the same framework, right? So we're back to the same slide. This is the third time. And now we have support context embedded. We have customer tickets at the bottom. And we have our kind of vectorized help center articles, our support SOPs. And the really key tie-in here that was important for support is to make sure that as we saw novel patterns with new tickets that were maybe going directly to humans and maybe cases where automation was failing, that that was driving self-improvement loops that directly updated the SOPs so that you have this kind of self-improving system that can really scale with volume. In the middle, we have ticket classification and actions as kind of the core skills, and then distribution inside of the help center, which we're going to look at in just a second. And also the real-time API, which is sort of more of an alpha product right now, or this sort of application of that product, where people can call in and get live support. Now, the really compelling thing about the real-time API here is we didn't have to significantly re-architect this foundation. Because if you get this piece right, you can expand to multimodal, and you can expand to more surfaces. I'm going to quickly show the support page here. And headline here is we're spending 10 seconds here. And that's because the point is that you ask a question, you can get answers in your flow of work. And as a user, all you want from support is to get what you need. And that's what the system can really drive at scale. Some quick hits on the impact. We're seeing about 70% of our tickets now are deflected or sort of handled autonomously by this system. This system is outperforming our legacy system by about 30%. And about 80% of these tickets, when manually reviewed by a QA team, are rated as highly positive. So we just got through three use cases. We talked about sales. We talked about people. We talked about support. And what you saw from all three of these was agents that could collect context, they can make decisions, and they can take actions as of your systems. You are also starting to see this sort of trend where every single department that we talked about today is operating a little bit like a software development team. And we think that trend is going to continue across every department inside of businesses across the globe. And that's going to create incredible opportunities for everybody in this room and for all of our developers to change the speed and shape of every part of your business. Now, to help you do the same inside of your business, you just heard from Sam on stage about Agent Kit. Again, Agent Kit is a sort of like starter kit for deploying these agents quicker. It includes Agent Builder, which is that. kind of visual, you're able to quickly map out the visual logic for your agent. We have ChatKit, which is like that easily embeddable chat component. We have evals, which allows you to drive some of those self-improvement loops and performance at scale. So great package there. Circling back to where we started, I asked the question, how do we use AI to amplify expertise? We challenge everyone in this room to start. with these three items and frame it around a sprint that you can start this week. Number one is find your Sophie. Who is your top operator that is excited to automate and orchestrate their work, understand their work, and move in fast loops? The second is to build inside of familiar tools. You should not be building a separate set of software here. This is embedded into the tools that your teams are already using so that both using the tool and providing feedback on the tool is incredibly native. The third part here is scaled platforms. Pick Agent Kit, build some of these common platforms so that you can really drive your development velocity at scale. I challenge everyone in this room to go back to your companies. and build something that your teams can't live without. Today, we're going to be, if you want to hear more from the builders at OpenAI, I'll be answering some questions on Discord. There's a Discord channel hopefully everyone has access to. We also have the lead developers from each of the use cases that we just talked about today. They'll kind of be hanging out over here. So if you have specific questions, come on up and see us. Hope everyone has a great rest of your dev day. And thanks for joining us. We'll see you around.

Ever wonder how OpenAI is using our own technology? Discover how we build reliable agents for sales, support, finance, and data, then take home a framework for spotting and scaling high-leverage use cases.