Sponsored by Open AI: From Governed Data to AI Systems That Get Work Done With OpenAI and Databricks

Summary

OpenAI and Databricks combine frontier models with governed enterprise data so that AI agents can reason over real business context, support actual workflows, and scale across every function in an organization.
OpenAI's Codex coding agent connects to the Databricks Data and AI platform through Genie to query data, build Delta tables with the Databricks CLI, and autonomously ship production features.
A multi-agent Slack workflow demonstrated in this video shows a treasury liquidity agent powered by a Databricks MCP connection analyzing data and triggering an automated execution package in response to a simulated liquidity crisis.

Chapters

FAQs

Why is governed enterprise data essential for OpenAI agents to work in production?

Frontier models need access to strong enterprise data to perform useful work in a business context—without it, agents lack the grounding to reason accurately over company-specific information. This video explains that Databricks provides the governed data layer while OpenAI provides the frontier models and agent harnesses that operate over it.

What is Codex and how does it use Databricks?

Codex is OpenAI's coding agent that can connect directly to the Databricks Data and AI platform via Genie and the Databricks CLI. In this video's demo, Codex analyzes customer feedback data, builds Delta tables, reviews its own output, and ships a mobile banking feature autonomously.

How does the OpenAI and Databricks integration work technically?

OpenAI agents access Databricks data through connectors and catalog access, with MCP enabling secure tool and API integration between the two platforms. This video shows Databricks as the home for governed context while OpenAI's models and agent harnesses execute work against that data.

What does the treasury liquidity agent demo show in this video?

The Trace demo shows a treasury liquidity agent that uses a Databricks MCP connection to analyze liquidity data and then triggers a multi-agent Slack workflow when a crisis scenario is detected. A Relay agent prepares an execution package and updates the model, demonstrating how frontier AI and governed data work together for finance use cases.

Full transcript

[00:09] All right. Hey everyone, I'm Evan and I lead the demo studio at OpenAI. I'm so excited to be here at the data and AI summit. Thank you very much for having me at data bicks and thank you all in the audience for joining me. Um, so it's very very important uh that the lawyers
[00:23] hear me say this. Uh, I'm actually not really going to talk about the road map. So, this doesn't necessarily apply, but in case I somehow have a stroke on stage and like start talking about the road map um and their forward-looking statements, you cannot legally hold me
[00:35] to it now that you've seen this slide, you all agree unless you leave right now. Um, also, please remember to complete your surveys in the attendee experience app. Thank you. All right, so today we're going to talk
[00:49] about getting more work done with agents. Um, and I'm sure that you've been hearing non-stop about agents in every session, uh, at the conference today and for probably the next couple of days. Um, we're going to talk about really how we bring agents, you know,
[01:01] into the enterprise, get more work done, and we're going to look at a couple of practical examples across engineering, uh, knowledge workers and finance and really everywhere in between. Uh, and of course, we'll be focused on how we can
[01:13] get more out of OpenAI's models to do so. So um I do want to acknowledge right if you're here uh you probably know this but it's worth being said uh data bicks is the home for governed enterprise
[01:26] data. Um that's really important because as we all know um agents uh really don't work unless they have access to good context about your data. Um you know I've been in OpenAI for a while. I remember back in the day when the models
[01:39] weren't even capable of doing simple things like counting the number of Rs in the word strawberry. That was always fun on sales calls. um they've gotten a lot more sophisticated since then. Again, should come as no surprise, but what hasn't changed is their need for really
[01:52] strong uh enterprise data to do things um in the business. So, data bricks is where govern context lives and of course open AAI is focused on pushing the frontiers of intelligence. So, we develop models and we develop agent
[02:05] harnesses to deploy those models and that's why this is such a natural partnership, right? um with the data bricks data that you all have deployed and with our models together we can build some really impactful solutions.
[02:17] Um so speaking of frontier intelligence like I said I've been at the company uh for quite a while. I started in mid 2023. Um and there's been this huge explosion in model capabilities since then and also in the things that we can
[02:29] do with those models. Um so things that models couldn't do in 2023 uh they couldn't really reason right it was like call in response. Um it was like asking a really smart person uh to spend exactly three seconds thinking about an
[02:42] answer to your question. Of course now models can do much more as they think and reason with chain of thought. Um you know at that time they weren't particularly good at writing lots of code maybe doing autocomplete but they weren't going to build an application for you. Um and they certainly couldn't
[02:55] take on tasks that took more than a couple of seconds to do. And now of course we see models working for hours or sometimes days at a time. So there's been this massive explosion in model capabilities. Um but you know for
[03:07] enterprise model intelligence alone doesn't define the value by itself right um you know just simply putting that model out in the world in your company doesn't lead you to agents that can really do anything. And so the next question becomes um you know how can we
[03:21] really ground our agents uh with data from the enterprise um and how can we give agents access to that data in a way that allows them to get to work. So we're going to start to talk about that question with a couple of examples in
[03:33] this presentation. So another way of stating that question is how can we safely deploy agents inside of our enterprises and if you think about what an enterprise agent needs to do it needs to operate within the context systems
[03:46] and controls that the business already trusts and going from that signal take concrete actions. So it's not just answering questions anymore. It's actually taking the right next action. And of course that's why data bricks is
[03:58] so central to the discussion because there's no way to separate this problem from the question of where discover business data come from. So I want to take a look now at the open AAI agent stack. This is what we can layer on top
[04:11] of the data to get work done inside the enterprise. So at the base is models. Um you know GPT 5.5 and our other models are the sort of brain that powers what agents can do and provide that reasoning
[04:23] capability. Then on top of that is the agent harness um which is what provides the uh the tools that allow agents to follow instructions and take actions. And in our case you can see this is
[04:36] built on top of the agents SDK and the responses API. Um next is the sandbox execution layer. So the idea is if we've given our models an agent harness that allows them to take action where can they safely take
[04:49] that action in the enterprise? So this is access to the right files to the right commands and packages. Um that whole layer of execution for code is in this part of the stack. And then finally
[05:01] there's all the operational requirements that enterprises need. So things like tracing evals and of course cost controls to make sure that once we deploy these agents they do what they're supposed to do and they continue to
[05:13] improve over time. So this stack all of these things this is kind of abstracted but there are products that that we build and the products um to access this we're going to go through today would be uh codeex which is our coding agent we
[05:26] built chat GPT enterprise which is a great way uh for large um you know employee bases to access our models and of course we also uh offer our API platform um incidentally uh as we'll
[05:38] talk about in a second here um you can do inference on our models uh both in OpenAI's native AP API platform and also in data bricks which is super exciting. Um so to break this down a bit because there's a lot of boxes here on the left
[05:51] hand side this is all the ways that you can consume models directly from open. So you can see again our products chat GBT and codeex the agents and codeex workflows that we can build and then the
[06:03] way we access data bricks data is through connectors. So primarily datab bricks mcp and data bricks cli and I'm going to talk about both those integration styles as we go through the presentation. And then the right hand side is all the ways that you can access
[06:16] our models within data bricks. So of course you've got genie genie code agent bricks. These can all access OpenAI models. And of course you know uh data bricks is model agnostic. So you have choices there. We hope that you choose
[06:28] our models because they're fantastic but there are choices available to you. Um and then of course all the way down at the bottom everything is powered by data bricks data. So um the point here is that the surface may differ. You may go
[06:40] through our products, you may go through data bricks, but at the base of it all is data bicks providing governed data and openai providing model intelligence. Okay, so once an agent has access to
[06:53] intelligence tools, the harness and focused governed data, it can go way beyond simply just answering questions and again start to take actions. So identifying the relevant signal from the
[07:05] data, identifying the right context from stakeholders, reasoning across steps, and then using tools to take actions. That's something that agents can start to do. And at OpenAI, our goal is to
[07:18] fundamentally change all types of work that gets done on a computer. Um, you know, computer work is kind of an interesting phrase. When I first saw this, I kind of did a double take. Um, this is something that, uh, you know, Greg Brockman, if you saw him talk today, uh, during the keynote, this is a
[07:31] phrase that he likes to use a lot, uh, because he likes to point out the fact that there are lots of different kinds of work that happen in the world. Right now, uh, AI is really good at doing work on a computer. And so, our goal is to deploy our models in the enterprise to
[07:44] really change the way that computer work gets done. And that means bringing agents into every digital work surface. So, it's not just an app. It's not just on your mobile phone. It's in all those places. It's also integrated into your tools.
[07:57] And there might be actually multiple versions of each of these surfaces depending on what kind of work you're doing that is going to be ideal for your workflows. Okay. So, now I'd like to take a look at some examples of these agents in action.
[08:09] And we're going to start with Codeex. Again, um Codeex is our coding agent. And it makes sense to start with because the class of computer work that we've seen the biggest change in with AI in the past couple of months is coding. Um
[08:21] you know if you write software for a living, uh you probably would agree with me that the way that you work today with AI looks very little like how you worked a year ago or two years before that. Um so we're seeing this change happening fastest in that industry. Um and with
[08:35] codeex again it's a coding agent runs on your laptop has access to our models and with connectors to data bricks can access your governed data. It really changes the way that you develop software. Um, now like all coding
[08:48] agents, of course, Codex is really good at the building part, at writing the code. Um, but that's not it. Uh, it started there. So, when I first started using Codeex about a year ago, uh, it kind of felt to me like a really smart
[09:00] developer who was great at writing code, uh, but who never checked Slack, never read email, um, never looked at the project plan, uh, you know, it was up to me as the developer to constantly bring in that context. And it felt a lot like babysitting. Candidly, I didn't like
[09:13] using it very much. Um, but Codex has come a long way with improvements in the model GPT 5.5 with improvements in the harness. Now, we can do the entire software development life cycle directly inside of Codex all the way from
[09:26] starting off by understanding the context through building and testing and deployment. And we're going to show how Codex runs that in this demo in just a minute. Um, of course, data bricks context being very important. You're
[09:38] going to see a couple of demos contained within this next example of where we can actually use the native data bicks genie plugin. This is official. It's available from datab bricks. It's available right inside of codeex and it gives codec
[09:50] access to all of that data and govern context that lives inside of data bicks. It can also do some other things that we'll take a look at. Okay. So, I'm going to show you a demo now of how codec changes the workflow.
[10:02] So, in this demo I'm going to play the role of a product engineer at Blossom Bank. We're a digital bank and of course that means we live and die by the quality of our mobile app and my job is to build the features that are going to improve retention and drive growth. Um
[10:16] but like many of you we're facing a lot of pressure at Blossom Bank to transform the way we work you know as an engineering team as a business and as an industry. And a big part of that pressure is is pushing us to be more
[10:28] responsive to customers. Um you know we now have the tools to respond in near real time to their requests but it's not easy. we face a couple of big problems. Um the first is that we're bogged down in internal process. We have this app to
[10:40] maintain. We spend a lot of time keeping the lights on. Um that's no good. And then secondly, we have a ton of feedback to process. I mean like a lot like more than we can handle a lot because with
[10:52] all this data, it makes it very difficult for us to figure out what we should build. There's a lot of of noise. Finding the signal is challenging. So, I'm going to show you how we use data bricks to make sense of this data and how we use codecs to be more responsive
[11:05] to our customers than ever before. All right, so here I am inside of Codeex and I'm getting started with my day um as that product engineer. And there's a couple of things that I've actually already cleared that would normally be
[11:18] jamming up my time. You know, we hear statistics like you only 20% of a developer's time is actually spent um building the features that they're supposed to be working on. Let's take a look at how we can clear our schedule of some of that busy work. So, one of the
[11:30] things I used to spend a lot of time doing was writing status reports um for my manager. Oh, before I get into that, here's the mobile app that I'm building. Um you can see it's a mobile banking app. It's beautiful, but we know that there's a lot that could be improved.
[11:42] Okay, so one of the tasks that I spend a long time on is writing status reports for my manager. And with Codeex, I've actually integrated it directly with the GitHub repository for Blossom Bank. So, it's able to look, take a look at all
[11:54] the PRs, read all the commit messages, and actually write up this beautiful summary for me that I can just send over to my boss. So, this is great. I don't have to spend all the time digging through pull requests manually to find this context. Um, the next type of busy
[12:07] work that I can get off my plate is, you know, reporting on incident status. So, we have some metrics and KPIs we have to hit here. Um, I again, uh, codeex being integrated with my, uh, incident management system. I can simply pull in
[12:19] that data and I've got this beautifully formatted report that I can send out to leadership along with that update. So, I've already cleared a lot of work off my plate, but when I came into the office this morning, of course, I saw a Slack message from my boss saying, "Hey,
[12:31] can you prepare the mobile status report?" This is a report that goes out every Friday and it informs everybody in the company how we're doing against our goals with mobile. No fun. Would take me a lot of time going through Slack and other systems to find it. But with
[12:44] Codeex, I can automate that as well. So, what I'm going to do is I'm going to use this skill that I built called the weekly mobile brief. Um, and this tells Codeex sort of exactly how I like this data formatted, where I should retrieve
[12:56] it from. And so now all I have to do is prompt it. It can go take a look at all the files on my machine. We can take a look at my integrated systems and give me that status report. And you can see these nice six bullets that we can just copy and paste right into Slack. And if
[13:09] we wanted to, we can even integrate Codeex directly with Slack so it could send it for us. So that's great. It'll save me a bunch of time. And if I really want to put this on autopilot, I can use an automation. So you can think about an automation kind of like a cron job. It
[13:21] has a schedule. It can integrate with multiple systems. It has a prompt and it'll fire at a particular time every week. So now the weekly engineering summary that's sorted. Fantastic. And now I can finally start to build. Well,
[13:34] actually I can't start to build because I don't know what to build yet. That's the problem, right? We said we had all this feedback. So let's take a look at how we can make sense of that data by using datab bricks genie right inside of codeex. So I'm going to go ahead and
[13:46] write a prompt and I'll say hey let's use genie to analyze our recent customer demand across social mentions and feedback submissions. We've got all that social data loaded into our data lakeink and with data bicks genie codeex can
[13:58] access it directly. Now, interestingly enough, here we see this running, um, you'll see that this plugin has access to the MCP server and has some instructions, but at no point during this process do I have to tell Codeex anything about the schema, about the
[14:11] underlying data. Um, you can think about this almost like CEX is delegating to a sub agent with data bricks genie. So, it's it's just writing a natural language prompt um, much like a user would for Genie and we get back the the
[14:23] results that we need all without Codex having to know what's going on under the hood. Um, so in this case you can see that Genie points to predictive budgeting as the strongest retention and growth opportunity. And here is the evidence. It's got the most mentions. It
[14:35] has a high urgency score and the retention lift is just about 5%. And because Codeex is integrated with our GitHub repository, it can actually take that and go one step further. It can recommend specifically what to build. So
[14:47] the way we address this opportunity is building a month-end forecast card. Codeex knows exactly where to put it in the app. Okay. So the next thing I want to do because of course like any good engineer I want to understand the customer pain is I'll ask data bricks to
[15:01] pull a representative tweet to capture the demand and it should be able to do that for me here. Let's take a look. Um okay it's still working and there we go. So here's our pull quote. If Blossom Bank had predicted budgeting it would
[15:14] have told me yesterday that this weekend was a bad idea financially. Okay. This is a perfect expression of the pain that our customers are feeling and now we know what to build to help them solve it. So let's get started building. Well,
[15:27] actually there's one more step. Before I write any code, what I want to do is actually have a plan that I can review. And so I'm going to use another skill that our engineering team has developed called plan mobile feature. And what this is going to do is read the repo and
[15:40] take the information that data bricks returned and turn it into this nice scoped mobile feature proposal. So let's take a look at this. Um, again, Codex is going to access the files on my machine just like I would. It'll read through
[15:52] the context. It'll compare that to the feature request from the previous step. And now I get this nice problem statement um what to build. And then I want to highlight this risks and unknown sections because this is really interesting. You know, a lot of coding
[16:04] agents early on would build exactly what you asked them to. And if you ask them to build the wrong thing, too bad, that's on you. Um, but like a trusted collaborator, Codex is actually able to go a step further and see around corners
[16:16] that I might not have anticipated. So, here's one I want to point out. This sort of um issue around what happens if the user doesn't have a lot of data. So, this predicted budgeting feature could work really well for a user who's got a ton of transactions. We know kind of the
[16:29] pattern over the past couple of months. We can predict where their spend is headed. What happens for a new user or somebody who doesn't use their card all that frequently? Well, that's good. Codex pointed that out and now I can think about it. Okay. So, I'm finally satisfied. Let's actually get to
[16:42] building. I'm going to ask Codeex to implement the feature. And now, I've sped this part up a little bit because, you know, sitting watching an AI generate something is not the most fun thing in the world. But we'll see that it gets started. It looks at the files on the machine and it starts to write
[16:54] some files. You can see what's being changed down there at the bottom. And in just about 3 minutes and 41 seconds, I get back my feature. Okay, but how do I know that this actually works? Well, I got to test it. So, I can do that also
[17:07] right inside of Codeex. Now, one way I could do that would be to look at the code directly, but you know, who's got time for that? So, I'll just go ahead and launch the mobile preview here. And now we can see right here, there's our month-end forecast card. It's implemented on brand inside of our
[17:20] existing design system. Codeex knows what's going on in the existing app and it's able to integrate this in very seamlessly. So, this looks really, really good. And I can see I'm doing a little better than Priya on my spending this month. Okay, so the next step into
[17:34] getting this into production is to push up into the repository. So, I'll ask Codeex to commit the changes to the repo. We'll fire that off. And just a second here, we should get a link to review them directly in GitHub.
[17:47] Okay, let's take a look. There's our PR and you can see that um we get some feedback from the humans. There's some uh you know, CI/CD automation running. Um but in a second, what you're going to
[18:00] see here is that Codeex actually does an automated code review here. So, you can see the chat GPT codeex connector. It does a review and it found there's actually a problem here where we haven't encrypted all of the keys and the
[18:14] outputs in this web hook payload that is going to trigger the budgeting to run. So that's a big problem. We want to fix that before we ship. And because Codeex ran this automated review on the code that it wrote, that's another layer of protection to make sure we don't ship
[18:27] something we shouldn't. Okay, so that's great. Ready to move forward with the deployment. And there's just one last thing I have to do before I can call this project done. And that of course is to update my data bricks lakehouse to track this new feature because I need to
[18:39] report on the adoption. I want to understand if people are using this. So I need to create some new delta tables. Now this is where the next data bricks integration comes in. So what I'll do is I'll say hey using the context from this
[18:51] commit that I just made um use the data bricks CLI and create the delta tables we can use to track these new entities. So I'll fire this off. Um, again, this is sped up a little bit, but what you're going to see in a a second is that um,
[19:04] we start to actually write shell scripts to automate the data brick CLI that's installed on my machine. This is a really cool example of codeex using my computer the same way that I would to script this behavior. And you can see
[19:17] here that it's actually using the standard datab bricks tables create command. And here's the confirmation. And in a second, we should see that all the tables I need are now live and ready to accept this analytics data.
[19:30] So that's the end to end. In just a couple of seconds, I'm going to jump back to the deck here. All right, I hit the button. Um, so you saw just a couple of seconds, Codeex was able to start by
[19:42] leveraging data bricks Genie to make sense of all that noisy customer signal. Um, this is a fancy diagram. It's actually not that fancy, but kind of shows exactly what's going on here. Um so you can see that you know codeex connected directly to our repo. It also
[19:55] is talking to data bricks genie mcp and that allows it to query all the catalog tables that are here. So we were then able to turn that signal into a ship. We kept it all secure with automated code reviews and then we were even able to
[20:09] automate the creation of those delta tables using the data brick CLI directly by automating that on my machine. So this is a pretty different way of developing software. This is all stuff I would be doing
[20:22] manually. And even with early coding AIs that were more like an autocomplete, it would still be me to orchestrate all of these steps. But now with Codeex able to take actions directly on my machine, this is a streamlined endto-end process
[20:34] and I more features that delight customers faster, happy as a product engineer. All right, so now that we've seen codeex change the game for software developers, let's take a look at some of the agentic workflows that business
[20:47] teams can use. So, we'll start off with Chat TPD for Excel and Sheets. And honestly, spreadsheets are a really great example of where agents can deliver a lot of impact for existing enterprise workflows. They have a ton of business
[21:00] context in them already, but there's still a lot of manual work that goes into doing the analysis and we can speed that up with agents. Um, and actually what's really interesting about this product is that it generates native formula. So rather than using Python
[21:13] code to create new analysis, it has inspectable editable formula. So it's more like collaborating with another analyst as opposed to just handing off the entire work to an agent and having no idea what went on. So this is great. Um people love using this. It frees up a
[21:26] lot of time to actually understand the story that it is tell telling as opposed to writing all those formula by hand. Another great example of a way that we can speed up existing workflows is with chat GPD for Microsoft PowerPoint. Um,
[21:39] if you're like most companies, you spend a ton of time taking your great work and putting it into slides and um, spending all that time sort of tweaking and formatting is really where your time is.
[21:51] Telling the story is where you want to be spending your time and using an agent like this can help a lot. So, chat GPT for sheets and slides, they're really good examples of agents which can help with ad hoc tasks. Um,
[22:04] you can kind of think about them like a personal assistant model of an agent. Um, but you know, once business teams get comfortable using these for one-off tasks u in their personal work, they're inevitably going to want to start automating their team's processes. I've
[22:17] seen this a lot working with customers at OpenAI. You get the bug and you say, "How can I scale this out?" And that's where workspace agents come in. They're really fast and easy to build and improve because they use natural language. They're designed for business
[22:29] users to to use. And unlike these sort of personal assistant agents that only work for me, they're always on and they can be triggered programmatically. They can be integrated with chat platforms. So, they really can be invoked anywhere that teams are already working. Um, you
[22:42] know, you can think about them like a digital co-orker that can pitch in to fill gaps in a team's org chart as opposed to a personal AI assistant that everyone has to learn to use independently. So, I'd like to show you a demo now of how this is starting to
[22:54] play out within a finance team. So, hopping back over to the demo machine. Uh, in this demo, I'm going to play the role of Morgan. And I am a treasurer at a software company. So, my
[23:06] job is to make sure that the company is making good financial decisions and never runs out of money. And we're a small scrappy startup, which means that my team honestly has some gaps. Um, I'm doing a lot of this work on my own. I'm trying to scale very quickly. And so,
[23:19] what I'm going to show you is how I can use a team that consists of some agents and some humans working together to get this job done and keep us in a good financial position. So here I am inside
[23:31] of chat TPT and you can see a couple of the agents on the lefth hand side that I've built out. Each of these was built using natural language and I'll show you what one of them looks like. So we'll take a look at trace which is my liquidity assistant and if we edit the
[23:43] agent you'll see how it's configured. So a couple of key things to point out here. Um you can see all the apps that are connected. So apps are like the connectors we looked at in the codeex example and datab bricks also is here as
[23:55] an MCP connector. Um, you can see I've also got access to Google Drive, so I can create spreadsheets and a couple of our SAS applications that we use in Treasury are connected here as well. Then the agent has access to skills, so
[24:07] it's good at doing things like liquidity assessments or modeling our cash runway. And it also has memory, so it can learn over time as it's invoked and get better at its job. Now, the next thing I want to show you is how the instructions
[24:20] work. So these are natural language instructions. You can see here how the various tools are referenced and I have instructions here about exactly how we should use data bricks. So it's going to be our source for receivables and cash flow inputs and then our other SAS
[24:33] applications are listed out here as well. Um and this was all written using natural language of course but I didn't write a word of this. So one thing that I hear a lot um from business users who are getting familiar with this product
[24:45] is hey this is great Evan I'm sure you can write beautiful prompts. How do I do it too? Well that's where the prompt builder comes in. So this thing on the left side here, the way that I built this agent is I simply describe to it, here's my workflow and here's what I
[24:58] want to build, here are the tools I want to integrate with. And in fact, if I have an onboarding guide, something I'd use to onboard a new hire where I've written my process up, I can just upload that PDF and tell this prompt builder, hey, figure out what tools I should
[25:10] connect to. And if I get stuck, so for example, my runbook says, hey, you got to talk to data bricks to get data about your receivables. I don't have a connector for data bricks. It can even coach me on what to ask my IT department to do to get that set up. So the idea
[25:24] really is to empower business users to build process automation for themselves. That's how they do that. Okay. So now that you understand how trace works, let's take a look at how I can put it into action. Um so I come into the
[25:38] office and I see this email and immediately my stomach drops because um this is our CFO telling me that there is a banking crisis barreling towards us. We're on track to breach our minimum
[25:50] cashold uh commitments and I need a three-point action plan with my action response this afternoon. Now, this is literally my worst nightmare. It's the one thing I'm paid to do to make sure the company doesn't ever run into money
[26:02] and I'm facing it right now. But luckily, I've got my agents and my human colleagues to help. So, let's take a look at how they work together. So, um it's 2026. What's the first thing we do in a crisis? Of course, we spin up a new
[26:14] Slack channel. So this is my liquidity tiger team Slack channel and let's invite some colleagues into it. So I'm going to ping Trace and Priya. Again, Trace is my liquidity agent and Priya is
[26:27] my ops colleague who can validate any of the key points that Trace needs confirmation on. And this is super important. Even with the best data in the world, and we hope that data isn't data bricks, of course, there are going to be gaps that agents need to solve
[26:40] for. We don't want them guessing. And so it helps if we can tell them who is accountable to get the information that is missing. So Trace is the agent, Priya is the human. Let's take a look at how they work together. So I'll click into
[26:52] the thread here. We can see that trace checks the bank balances and the ERP receivables, creates a data model in Google Sheets, which we'll take a look at in a moment, and flags a really
[27:04] important point for us, which is that two customer pay payments worth $180 million have just slipped by 7 days. So this explains why our cash flow position is in such a bad place. Okay. So now you
[27:16] can see that trace is pinging Priya and asking her to confirm whether those are actually true delays or they're just some weirdness in the ERP because that's important, right? This could be out ofd uh data. We really want a human to weigh
[27:28] in on this. So let's see what Priya says. Okay, so unfortunately Priya confirms um that this is truly uh a pushed payment. That's no good. Um, but luckily she's confirmed that the
[27:40] revolver capacity we have available today is up to $125 million. Now, I want to take a pause here for a second because I'm throwing a lot of financial jargon at you. The the goal of this demo is not to make you a treasurer. The goal of the demo is to show exactly how
[27:53] specialized these agents can become because this is something Treasury Department deals with all the time. Okay, so bear with me. Priya confirms what's going on. Um, of course, Trace gets back to work with that information.
[28:05] It updates the model with the new data and is now asking if it should model the response options for me. That sounds delightful. So, I'll go ahead and approve it. So, in a moment, let's take a look at the spreadsheet that Trace
[28:19] built out. Here it is. And we'll click through and look at it in Google Sheets. Okay. Um, so I don't know about you, but this looks pretty good to me. And agents sure have come a long way. I remember the first Excel model I got back from
[28:31] chat dbt looked pretty ugly. Now it's nice and formatted. It has all the information up top and if we click through some of the tabs at the bottom u we'll see here's our 10day runway. Um so yep in fact it looks like we're going to run on money just a couple of days. I
[28:44] don't like that red ink but it's good to confirm it. Um you can see all the sources and assumptions that were referenced here. So data bricks right up there where we expect it to be using the Genie connector. But most importantly I get the scenarios modeled out for me. So
[28:58] you can see that three scenarios are recommended here. I've got one where we draw the full revolver. maybe a little bit aggressive, but that probably would work. Another where we we just freeze all of our spending. That's not going to work. The business will shut down tomorrow. And this third model, which is
[29:11] sort of somewhere in the middle. So, that's what um Trace is recommending. But if I dig into that scenario a bit more, we'll see the way that it got to that math is by deferring our hosting payments. Now, I don't know about you, but if you work at a software company,
[29:24] the last thing you want to do is stop paying your hosting provider. You're probably not going to last very long. So, that's a big problem, and we're going to help Trace get better. But the next thing I need to do is ask for a fourth scenario. And I can do that actually right inside of chat GPT for
[29:38] Excel and Google Sheets. So I'm going to load up Chat GPT. We'll open the sidebar. Um and here's that intelligence following me into the workflow where I'm already working. So I'm going to ask uh hey add a fourth scenario which keeps
[29:50] the hosting payment and defers that settlement. Um and you'll see that chatb for Excel gets right to work writing native formula updating the spreadsheet and I get that fourth scenario option. Okay, great. So, it looks like if we are
[30:03] to draw $49 million, which is right in the middle, we can actually stay on top of this if we defer that intercomp settlement. Sounds good to me. Okay, but we did notice that Trace made a pretty critical error here. So, let's take a moment and fix that. So, what I'm going
[30:16] to do is I'll just take a screenshot of the model that was created and I'll hop back into chat GPT and just paste that right into the prompt builder. So now that context is available and I'm going
[30:28] to tell trace, hey, you know, scenario 3 was a miss. Um, please don't ever uh tell me to pause my infrastructure payments unless I give you that exception approval. Um, so you'll see here that uh the side prompt builder
[30:40] gets to work. It updates the instructions and now Trace isn't going to make this mistake again. It really is that easy to update these models operating instructions with a screenshot and a simple prompt. Okay, so now that Trace is in a good position, let's hop
[30:53] back into our Slack channel and round out this exercise. So I'm going to go ahead and share the approved execution plan. So we're going to draw a revolver. We're going to preserve that infrastructure spend. And now I'm
[31:05] tagging in a second agent, Relay. So relay is kind of like an operations agent that can prepare all the necessary documents to execute this transaction. So I'll send the message off. We'll see that relay gets to work in just a
[31:18] moment. Let's take a look here. And here is our completed execution package. So, I've got my board memo to tell every what's going on. I've got this revolver request that I can send over for approval. And um it's updated uh the
[31:31] response tracker to track each of these steps and who's accountable. And now this is really interesting. Um it knows who to tag in for approvals. So, Priya again is the DRRI for confirming that these instructions are good. But then
[31:43] it's also tagging in trace to keep the runway model updated as confirmations start trickling in. And uh I think this is a pretty cool moment as you start to get multiple agents into the Slack channels together to see them start to
[31:56] collaborate with each other. Um this is something where that concept of filling in gaps in your org chart really starting to see that org chart build out with agents um becomes a reality. So very very cool to see this happening in
[32:08] action. Okay, so just to recap in a couple of minutes, we went from this sort of like crisis mode, right? We we have a cash shortfall. We don't know where to go. Um we don't have a lot of people to help us out. And between me and Priya and a
[32:21] couple of our agents, we were able to do the analysis um update our agent and come to a really good place with a plan already in motion. Um but that's just one of the agents or actually two that we have here at side inside of um of our
[32:34] company. Um there are so many more. Um so you can think about all kinds of applications for these things you know customer support sales automation um something as simple as account research. All of these can be started to automated
[32:46] with workspace agents and in fact all these are available as templates right inside of chatbt enterprise. All right so hopping back to our slides to wrap up.
[32:58] So, we've just gone through a couple examples of how teams across the company from software to finance are starting to embed agents into their workflows. But of course, like their human colleagues, agents are only as good as the data they
[33:11] have access to. And that is why we are really excited about the opportunity to partner with data bicks to combine that govern data that data bicks has access to with our frontier intelligence to help you get more value from your
[33:23] agents. And if you'd like to learn more about how this works, come on and visit us at the OpenAI booth. We've got some awesome codeex credits to give you to get you started on this journey. And we'd love to help you build what comes next. Thank
[33:36] you so much. And I'd love to open up for questions. Uh we got a couple minutes here and a mic is down here. Um any questions you have, I'd love to take them.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.