Mise en production de l’IA dans les workflows : Lee Spacagna, Solutions Engineer, OpenAI

Hi everyone, my name is Lee and I'm a lead solution engineer here at OpenAI working very closely with financial services customers across AMIA. Today we've got some really exciting things to show you, some of which we only launched last week. Every day when I work with financial services institutions, they always have one question. Where can AI actually change how my business runs? Today there are two paths for AI adoption. first giving chatb and codecs so that employees can use AI in their daily work. Second, there's systems. This is where companies are building entirely new products. They're uh enhancing their customer service. They're improving client advisory and they're working on operational support. So, it looks like this. We've got chat GPC and codeex working from the bottom up using employees as they get more AI literate. And then we've got AI systems from top down. And these are those major transformational uh initiatives. But there's a missing layer in the middle, the automation at the team and the department level. And this is a gap that the new chatbt workspace agents is designed to close. And that's what I'll be demoing for you today. When we say agents, we mean AI systems that we can delegate meaningful tasks to, not just ask questions of. And we can do that by using tools that we already rely on like email, calendar, and those productivity apps that we're using every day. And we want these agents to complete work in the same way that people do. And in the last few months, there's been a huge jump in capabilities. And we had another leap last week with GPT 5.5. Today, agents take on complex work that used to take hours or days, and they can handle them from start to finish. What happens when you need to build something custom and you need to delegate something that your team or your departments are currently working on? Many of you have already built custom GPTs. With Workspace agents, we evolved that into something much more powerful. We've got a new agent builder which brings shared applications, skills, and deployment all to one platform. And this allows these agents to work in the same place that work already happens. So now let's jump into the demo. This is the is a standard attribute interface that I'm sure you're all familiar with. But down the left hand side, you can now see we've got an agents option that we can start with. And now for many teams, the challenge isn't a lack of work to automate. It's that the work is spread across meetings, documents, emails, and other systems. And the decisions all depend on specific context. So today, I'm going to show you how I can quickly spin up an Aentic co-orker. In this case, I want to build my own chief of staff agent. I want it to help me coordinate work, track priorities, prepare meetings, and help keep the team moving. Every function can delegate meaningful work to agents and these can understand the role, use the right tools and they can operate with how the team already works. Here we're going to use one of the uh templates we've got already for the chief of staff uh agent here. And you can see this already has a set of instructions. It already has a set of tools that it's able to work with and it already has capabilities that I can start using. Next, you can see here I want to uh start connecting some of those tools that I mentioned to make sure it's correct for my workflow. In this case, I'm going to use the Microsoft set of uh tools here. So, we've got things like Outlook calendar, Teams, and then my Outlook email. Now, we can see that the instructions are going to be automatically written by another agent. So, you don't need prompt engineering skills. You don't need any technical skills at all. But what it means is that as a business user, you can now use an agent to build another agent for you just with natural language. And that's it. That's the initial version. I haven't written any code and I've got the first version of the agent ready to go. But now I want to customize this to my own requirements. I want it at 9:00 a.m. I want it to run every day. I wanted to look at all of my meetings, look at all of the applications, look at my emails that came in overnight, and generate a daily brief so that I could arrive and be prepared for all of the meetings for the day. So once again here I just give the instructions again in natural language telling it I want it to run at 9:00 a.m. No technical skills needed here at all. And within a matter of seconds we can see that it's able to customize my agents to my team's requirements and work exactly how I like to work every day. So once this is done, we can go and test our agent. And we can now see there's two starter prompts underneath to get me started. If I want to, I can give it its own set of instructions to perform for me. But you can see that there's two already there to go. And you can think of these as capabilities that have already been built into this agent. So let's ask it to prepare the today's brief here. You can see I'm asking it to do a concise brief uh using all of the available information that I mentioned earlier. Highlight priorities, decisions, blockers, and follow-ups and then post these in the CFO team channel in the daily prep channel. And now we can see the agent spinning up. It's going to start grabbing all of those details, connecting into my email, connecting to my calendar, and all those other sources that I mentioned. It's going to check all of those meetings that I've got for the day. It's going to cross reference that with information that might be in my emails. And it's going to pull all the context needed from all of these sources. And the first time it runs, it's going to ask me to give permission to post to the Teams channel. So, it's just going to set that up now. And we'll go and give it the the approval and see how that's worked. And that's it. That's now posted to Teams. So, let's now go and have a look at what it was able to generate for me. So, now over in Teams, we can see in the daily prep channel, we can see that there's an update. So let's go and have a look what it posted and we can see our chief of staff agent from chat GPT has gone and collected all that information and then it's posted that daily brief for me inside teams exactly where I want the information to be for my daily work. So within a couple of minutes we've built an agent from scratch. We've connected it to tools that I use every day. We've given it some customized guidance and we now have a running chief of staff agent for my whole team. But let's go back to the agent and take it a step further. This t this week my team have been burning themselves out running from meeting to meeting and they haven't had any time to prep in between. So now let's add a new capability. I want the agent to proactively research before every meeting like having an expert chief of staff who's the telling me who's there, what's the latest and then what's the goal of that meeting. So for this we need some additional context for some other tools. So now let's go and add some more that are available. I'm going to start off with SharePoint. This is where I'm storing all of the company information and all the information I've taken as notes that's shared across the organization. So, we'll go and add that. And next, I want to add Salesforce for all of that CRM and all that kind of rich information about all the context from that customer. So, we'll go and add Salesforce as well. And now that's done. We've got those two apps connected. You can also add other apps that you use every day here as well or even custom applications that you just you have inside your business. Next is skills. Skills are a way of capturing snippets of information instructions to perform critical tasks. Think of these as an amazing way to capture all of that those tribal knowledge and conventions that are currently trapped in people's heads. And we can turn those into repeatable workflows. You can see that there's two skills already in use here. We've got the chief of staff skill and we've got a final brief formatting skill. But now let's go and add another one that I've already been using across my team. So I've already got a skill here for meeting prep. This tells uh chatbt the way that I want this information to be structured, the key information that's needed, the source of this information, and where I want that information to be posted. So, let's go and add that to my agent as well. Finally, let's save our changes. And now I want to give the agent some more instructions about what to do with these new applications that I've gone and connected. So again, we'll use the agent on the side to have a natural language conversation and give it this additional context to go and update the agent. So here there's all the information. I've just added Salesforce and SharePoint. Add a new capability. I want it to be able to generate these quick meeting briefs in in chat GBT. And all the information I want to give it is just give me the information for the next meeting. So it needs to go through here and make the updates. Um and then it's it will also um add a new starter prompt there for me to use in a second. So now again, let's go and update this. And now we can go and deploy this. and is now available for the whole team. So, let's go and use it. And here is my completed agent deployed and ready. And now you can see we've now got a third starter prompt underneath as well to prepare me for my next meeting. Now, it's going to run pull all of those contexts from those different sources including Salesforce um and SharePoint that I went and added. It's going to go and um check all of the information, put that together into a concise brief in the way that the skill gave the instructions to go and represent that. And personally, I have one of these running every day. Um I have my own agent that checks all of my emails that come in overnight, the important updates from across the business, the things that I said I would do on Slack yesterday or on calls and all the contexts on the transcripts. And now it means that I get that first hour of my day back because I come into the uh into work in the morning and all of my emails have a draft ready to go with all of the contacts from across the business. And it means that I can just go through all of those emails and click send and just approve those and get those out to my customers. And it's completely transformed the way that I work. So now we can see here we're all ready to go for that next meeting. Before the team was understaffed and they couldn't prep for those meetings, but now we've enabled everyone to turn up as if they've been prepped by their own chief of staff agent. But that was just one example there. But this is a pattern that you can apply across the business. We've seen examples of agents like KYC on boarding, AML investigations, relationship management, and more. The opportunity here isn't about one single uh automation project. It's actually about a brand new operating model. Every team can spin up a ro specific agent to take manual work off their plate and help the business move faster. But the next question is is what if we've got thousands of these agents? How do we manage them? And that's where Frontier comes in. Frontier is our platform for deploying and managing agents at scale. It connects to systems usually in silos, things like data warehouses, CRM, and internal applications. It gives AI co-workers the same shared context that the teams currently rely on. And from there, agents can reason over data. They can run code. They can use tools and they can take actions all in a governed environment. And the key thing here as well is as they work, the system improves. They will learn from interactions. They will evaluate their performance over time. And it means that the more they do, the better they get, just like human workers in the business right now. So today it's possible to build in chat GPT codeex and the API and we want to make it easier to deploy out of the box agents, plugins and skills all specific for financial services workflows. This matters because it moves the system towards much more automation. We can use purpose-built agents that plug directly into work and they can handle all the repeatable processes with even less lift and customization. With all those foundations in place, agents become incredibly powerful, allowing you to delegate more workflows to AI over time. And next, Stephanie is going to show you how teams are using them to create transformative impacts across the workforce. Thank you.