Microsoft Ignite ’25 Keynote | Asha Sharma, President of CoreAI Product at Microsoft (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) Thanks, Jackie. How about a round of (00:00:03) applause for Jackie and Seth? (00:00:06) [applause] (00:00:10) Look, the work that Jackie and Seth do (00:00:12) is uh deeply personal to me. My father (00:00:15) was a surgeon for over 50 years and even (00:00:17) as a child, he would lecture me that (00:00:19) there was a mantra about operating at (00:00:21) the top of your license as a doctor. The (00:00:24) work that Epic is doing to infuse AI (00:00:27) into the flow of how health care (00:00:29) providers deliver care allows physicians (00:00:31) to go one step beyond not just (00:00:33) performing at the top of their license, (00:00:35) but performing at the top of their (00:00:37) potential. We're really privileged at (00:00:40) Microsoft and honored to be able to (00:00:42) provide the platform and tools that Epic (00:00:43) uses to build their AI solutions. We're (00:00:46) actually very excited to share more with (00:00:48) all of you today about how to leverage (00:00:50) the same. Like I said in my opening, (00:00:52) there's a maker in every one of us. (00:00:54) Empowering citizen developers and (00:00:56) professional developers and enabling (00:00:58) them all to collaborate with it to (00:01:00) create amazing AI solutions and then (00:01:02) allowing them to build upon new layers, (00:01:05) new fundamental building blocks that (00:01:07) embody context and intelligence is what (00:01:10) it's all about. Here to share more about (00:01:12) ubiquitous innovation, please welcome to (00:01:14) the stage Asha Sharma. (00:01:19) >> [music and applause] (00:01:25) [music] (00:01:29) >> Every person in every role now holds the (00:01:32) power to create right inside of the (00:01:34) tools they already know and love. From (00:01:37) the soil of a Nebraska farm to the code (00:01:39) running a global bank, innovation is no (00:01:42) longer the domain of few. It has become (00:01:44) the superpower of many (00:01:48) work will be defined by human creativity (00:01:50) and agents that help it scale in this (00:01:52) era. And at the center of this shift (00:01:55) across individuals, domain experts and (00:01:57) developers is Microsoft. (00:02:00) Let's start with the individual maker (00:02:02) inside of all of us. (00:02:05) Every one of us has had an idea for an (00:02:07) app that would make our lives just a (00:02:09) little bit easier. But today, turning (00:02:12) that idea into something real takes (00:02:13) someone technical and often months or (00:02:15) years of work. (00:02:18) Not anymore. (00:02:20) Today, we're announcing App Builder. So, (00:02:23) anyone can create an app in minutes (00:02:25) right inside of E365 C-Pilot. (00:02:30) All you do is start with your data you (00:02:32) already have, your charts, your files, (00:02:34) your meetings, all powered by work IQ. (00:02:37) You design the experience you want in (00:02:39) natural language and you share it with (00:02:41) your co-workers just like you would a (00:02:43) doc. This is the beginning of something (00:02:45) new. Software evolving from being made (00:02:48) for people to being made by people. (00:02:53) Now innovation doesn't just stop at the (00:02:55) individual builder. Some ideas require (00:02:58) more domain expertise, more data, and (00:03:00) deeper connection across your business. (00:03:03) Code-Pilot Studio is the only low code (00:03:06) agent builder that understands your (00:03:07) business right out of the box. So, it (00:03:09) uses work IQ, workflow, security, (00:03:12) identity to take real actions and not (00:03:14) just generate basic prototypes. (00:03:17) Today, more than 230,000 organizations (00:03:21) use C-Pilot Studio, including 90% of the (00:03:24) Fortune 500. (00:03:27) Now, for decades, developers have lived (00:03:30) at the frontier and they have been the (00:03:32) ones to solve the world's hardest (00:03:34) problems. And that doesn't change in (00:03:36) this era. What changes is the leverage. (00:03:40) We're bringing them new AI powered tools (00:03:42) to continue to build the future of (00:03:43) what's possible. And they do this on (00:03:46) GitHub. (00:03:48) This year, someone new joined GitHub (00:03:50) every single second, which is the (00:03:52) fastest growth we've ever seen. (00:03:56) GitHub Copilot started as a pair (00:03:58) programmer and now it's become an entire (00:04:00) fleet of coding agents and now it is (00:04:03) also the single largest contributor to (00:04:06) GitHub's codebase. (00:04:09) But we know that developers want choice, (00:04:12) enterprises want control, and everyone (00:04:14) wants fewer subscriptions. (00:04:17) That's why at GitHub Universe, we (00:04:19) announced Agent HQ, a single place to (00:04:22) pick agents from GitHub, OpenAI, (00:04:24) Anthropic, Google, Cognition, XAI, and (00:04:27) more. All with consistent governance and (00:04:29) observability as part of C-Pilot (00:04:32) subscriptions. (00:04:35) Now, when innovation becomes ubiquitous, (00:04:37) everyone gets a new set of powerful (00:04:39) tools to drive progress forward. And in (00:04:42) a retailer like Zava, that means the (00:04:44) frontline associates in the aisle, the (00:04:46) regional managers are running (00:04:47) performance across the stores, and (00:04:49) developers are writing the systems to (00:04:51) code every single line. (00:04:54) That's exactly what we're going to show (00:04:55) you. Please welcome Midi, Lydia, and (00:04:58) Kyle to the stage. (00:05:03) [applause] (00:05:06) Today at Zava, the store is packed and (00:05:09) sales are strong. But as the store (00:05:11) manager, I am scrambling behind the (00:05:14) scenes. Every shift change means (00:05:16) tracking down who's free, having (00:05:18) multiple forks to negotiate swaps, and (00:05:21) calling in favors just to keep things (00:05:23) running. I really need some help. So, I (00:05:25) reached out to Lydia over in the store (00:05:27) operations team, who says there's no (00:05:29) bandwidth for this right now. There has (00:05:31) to be a better way. Lydia says, "I (00:05:34) should check out Microsoft 365 C-Pilot, (00:05:37) where I can build apps, agents, and (00:05:39) workflows." (00:05:41) Now, I want to build an app to solve my (00:05:44) staff's scheduling challenges. So, I'm (00:05:46) going to go ahead and ask it to build me (00:05:48) one. I'm going to ask it to create an (00:05:50) app with a dashboard and rich visuals so (00:05:53) I can easily manage staff scheduling. (00:05:56) Now, App Builder understands what I'm (00:05:58) asking for and it immediately begins (00:06:01) generating the app right here inside (00:06:03) M365 Copilot. I'm really hoping it uses (00:06:06) all the documents, spreadsheets, notes, (00:06:09) and emails I use today to manage (00:06:11) schedules. And here is the app it built. (00:06:16) And I can see here a dashboard for (00:06:18) shifts across time, a manager view just (00:06:22) for me, a visual for store coverage, and (00:06:25) even a table for our team members and (00:06:28) their shifts. Most importantly, App (00:06:31) Builder built on top of Zava's Microsoft (00:06:34) 365 data, and it understood the business (00:06:36) the way I do. I did not need to learn (00:06:39) any new tools. I just described what I (00:06:41) needed. No more frantic calls. No more (00:06:45) guessing who's free because I've got an (00:06:48) app to do it for me. (00:06:50) Now, I'm so excited with what I've built (00:06:52) here that I'm going to share it with (00:06:54) Lydia to see where it can go. (00:06:57) And the second I saw it, I knew this (00:06:59) wasn't just going to help one store. All (00:07:01) of my stores could use this. Plus, I'm (00:07:04) going to address another challenge. When (00:07:06) someone can't make a shift, and chaos (00:07:08) ensues. Enter Microsoft Copilot Studio (00:07:11) to help me build an agent to fix that. (00:07:13) So I say to Copilot Studio, when someone (00:07:16) emails that they're sick, I want an (00:07:17) agent to secure a replacement team (00:07:19) member. Instantly, Copilot Studio jumps (00:07:23) into action and creates the full agent (00:07:25) framework. Copilot Studio has native (00:07:28) access to work IQ, including Microsoft (00:07:30) SharePoint and the Microsoft Graph. So (00:07:32) it's grounded in rich context. First, I (00:07:36) need to determine when I want the agent (00:07:38) to be triggered. (00:07:41) When a team member emails ouruling inbox (00:07:43) to say they're sick, I want the agent to (00:07:45) jump into action. So, we'll go ahead and (00:07:48) choose that. (00:07:52) Go ahead and give it an appropriate (00:07:53) title. (00:07:56) And just like that, my trigger is added. (00:07:58) Next, I want to go ahead and I want the (00:08:01) agent to interact with a staff member to (00:08:03) confirm their availability. To do that, (00:08:06) I'm going to go ahead and add a co-pilot (00:08:08) studio flow. This one's called contact (00:08:10) replacement team member. So, we click on (00:08:13) that and just like that, that's added to (00:08:15) my agent. Finally, I want the store (00:08:18) manager in the physical store to be (00:08:20) updated about which staff member will be (00:08:22) covering the shift. I'm going to use an (00:08:24) existing agent that my colleagues built (00:08:26) in Microsoft Foundry, which drastically (00:08:29) speeds up my efforts. (00:08:34) All we need to do to integrate C-Pilot (00:08:36) Studio with the Microsoft Foundry agent (00:08:38) is give it a name, add the description, (00:08:41) and the ID. (00:08:44) So, let's play this out. It's 7:45 a.m. (00:08:47) on a Saturday, and Whim, a team member, (00:08:50) says they're sick. Before anyone sees (00:08:52) the email, the agent jumps into action. (00:08:55) It starts by suggesting a replacement (00:08:57) team member based on skills and (00:08:59) availability, then contacts them to (00:09:01) confirm they can cover the shift. (00:09:03) Finally, it goes ahead and contacts the (00:09:05) store manager and lets them know that in (00:09:08) this case, Luca has confirmed he's going (00:09:10) to be able to cover the shift. And now, (00:09:12) my Zela pop-up stores have an app and an (00:09:15) agent working together to ensure every (00:09:17) shift is covered inside Microsoft 365 (00:09:20) Copilot and Copilot Studio. And now onto (00:09:23) the next task on my to-do list. I want (00:09:26) to handle this message that Ryan sent. (00:09:28) It looks like the new e-commerce site (00:09:30) isn't currently optimized to work for (00:09:32) mobile devices. So, let me go ahead (00:09:36) file that as an issue with GitHub and (00:09:39) let Kyle from our development team (00:09:40) figure that one out. (00:09:43) Thanks, Lydia. I'm a developer at Zava (00:09:46) and I work on our online retail store. (00:09:48) We use GitHub for our entire software (00:09:51) development life cycle. While I'd (00:09:53) normally be making my coffee right now, (00:09:56) I'm going to start by burning down (00:09:58) issues in our backlog. And I've asked (00:10:00) Copilot to tell me what's on my plate. (00:10:03) GitHub Copilot gives me a ton of choice (00:10:06) with models from all the Frontier Labs. (00:10:08) And as of earlier this morning, that (00:10:11) includes Gemini 3 Pro. And we even have (00:10:14) a custom fine-tuned model deployed to (00:10:18) Azure Foundry. So here's the issue from (00:10:21) Lydia. Improving the website to make it (00:10:24) more mobile friendly. Okay, honestly (00:10:26) this is important but it's kind of (00:10:28) boring to do. So I'm going to ask (00:10:31) Copilot to start work on this. And (00:10:34) because this is a UI task, I'm going to (00:10:37) use the custom coding agent that our (00:10:40) team built, focused specifically on (00:10:42) front-end changes based on our (00:10:44) frameworks and styles. (00:10:46) Let's also ask Copilot to include (00:10:49) screenshots so I can easily see the (00:10:51) difference without having to run the (00:10:53) app. Copilot gets to work right away (00:10:56) planning and writing code, tests, and (00:10:59) opening a pull request for my team to (00:11:01) review and merge later. I can manage (00:11:04) this task in all of my agent sessions in (00:11:07) one place. Agent HQ. Okay, Copilot has (00:11:11) that task in hand. So, let's go be a (00:11:14) team player and work on some security (00:11:16) remediation for our project. Our (00:11:19) security team uses Microsoft Defender (00:11:21) for Cloud to monitor our security (00:11:23) posture and find risks in production. (00:11:26) Today we are announcing a new code to (00:11:29) cloud integration between defender and (00:11:32) GitHub advanced security so we can f (00:11:35) filter for vulnerabilities using runtime (00:11:37) data like sensitive data runtime issues. (00:11:41) Our security team has created a campaign (00:11:44) to tackle these vulnerabilities. They (00:11:46) choose which vulnerabilities to focus on (00:11:49) and a deadline to get it done so we have (00:11:52) an achievable target. I'm going to (00:11:54) select all of these vulnerabilities and (00:11:57) assign them to C-Pilot. Copilot will (00:12:00) generate these fixes all in a single (00:12:03) pull request for our team to review. (00:12:06) Now, I also spend a bunch of time in VS (00:12:09) Code. And of course, I use coding agents (00:12:12) to help with my work here, too. Here I (00:12:15) can see the agent sessions related to (00:12:17) the code I'm working on across local (00:12:20) chats, CLI sessions, and even thirdparty (00:12:24) agents like codecs. I can even start new (00:12:28) tasks from right here and they'll run in (00:12:30) the cloud. Now, I realized I haven't (00:12:33) written any tests for this code. So (00:12:34) before my team calls me out in a pull (00:12:37) request, let's add some tests so we can (00:12:39) make sure that we're using O on all of (00:12:41) our pages. (00:12:43) At Zava, we have developers using (00:12:46) Copilot in a lot of different ways. And (00:12:48) now all enterprises can see how C-Pilot (00:12:51) is being used within their organization (00:12:54) with Copilot metrics. I can see usage (00:12:56) statistics, what models my team is using (00:12:59) or not using, and even the programming (00:13:02) languages that Copilot is assisting with (00:13:04) most. Okay, let's go back to where we (00:13:07) started with agent HQ. We kicked off a (00:13:10) bunch of coding tasks and I can see all (00:13:12) of that work here. Now, I want to take a (00:13:15) look at the mobile task I created (00:13:17) earlier to resolve Lydia's issue. I want (00:13:21) to give it a little more context and I (00:13:23) can just do that while it's running and (00:13:25) let it know that I only care about phone (00:13:28) sizes, not tablets right now. So, in (00:13:31) only a few minutes, I got C-Pilot to (00:13:35) remediate security issues, start making (00:13:38) our online retail store more mobile (00:13:40) friendly, and improve our testing so I (00:13:43) stay off the naughty list this year, all (00:13:45) while keeping me in control. What we (00:13:48) just saw was Midi using App Builder to (00:13:51) solve a scheduling problem. Lydia using (00:13:54) Copilot Studio to scale it with an agent (00:13:56) workflow across their region. And I used (00:13:59) GitHub agent HQ to tackle three tasks in (00:14:03) 2 minutes. That's the power of (00:14:05) ubiquitous innovation for each of us. (00:14:07) Back to you Asha. (00:14:09) >> Thanks team. Incredible work. [applause] (00:14:15) Now we're with this new work and this (00:14:17) new world of creation, people that build (00:14:20) AIdriven products need a very different (00:14:22) foundation. One built from models, (00:14:25) tools, and intelligence. (00:14:27) That's why we built Microsoft Foundry, (00:14:30) an open and modular app server for the (00:14:32) AI era. (00:14:34) Foundry has powered nearly every single (00:14:37) product that you've seen on the stage (00:14:38) today. And as we mark our first year, (00:14:41) the momentum is unmistakable. (00:14:44) Over 80,000 organizations are building (00:14:47) on Foundry and today we are the number (00:14:50) one AI platform in the world. (00:14:56) [applause] (00:14:59) Now we find ourselves in a multimodel (00:15:01) world. Foundry has become the home for (00:15:04) the leading models like Open AI, (00:15:07) Mistrol, Deepseek, Llama that your teams (00:15:10) use every day. (00:15:14) But there's been one frontier provider (00:15:16) that we've been missing. (00:15:19) This morning, we changed that. (00:15:22) Anthropic models are now live on (00:15:24) Foundry. (00:15:27) [applause] (00:15:28) Please join me in welcoming Mike, the (00:15:30) chief product officer of Anthropic. (00:15:42) Hey Mike. (00:15:43) >> Hey Asha. (00:15:44) >> Thanks for being here. What a what a (00:15:47) year it's been together. (00:15:48) >> It's been an amazing year together. I am (00:15:50) so excited to be here and so excited for (00:15:52) the news. (00:15:53) >> Amazing. Do you want to let's tell (00:15:55) everybody what we launched this morning. (00:15:57) >> Absolutely. So our partnership together (00:15:59) is really about enabling developers and (00:16:02) enterprises to build intelligent agents (00:16:04) with cloud no matter where they are. And (00:16:06) that's why today we announced the (00:16:08) availability of Cloud Sonnet 4.5, Haiku (00:16:11) 4.5 and Opus 4.1 models, all our tier (00:16:14) models in public preview on Microsoft (00:16:16) Foundry. Cloud on Foundry enables (00:16:19) customers to build, scale, and secure (00:16:22) productionready applications and (00:16:23) enterprise ready agents in one trusted (00:16:26) platform. (00:16:27) >> Now, we've been finding ways to work (00:16:29) together over the last year. Um, but (00:16:31) we've been cooking up some exciting (00:16:33) things. Do you want to show everybody (00:16:34) what we're going to be rolling out over (00:16:35) the next few months? (00:16:36) >> Absolutely. Um, so we've already (00:16:38) partnered together to bring the cloud (00:16:40) family of models to GitHub copilot's 26 (00:16:42) million users. We've been collaborating (00:16:44) closely on the MCP standard for tool (00:16:47) calling and we've been offering cloud as (00:16:48) part of M365 copilot chat for billions (00:16:51) of users. Now we're taking the next big (00:16:54) step by bringing cloud models to all of (00:16:56) Microsoft Foundry. In this example, (00:16:59) we're using Sonnet 4.5 and skills to (00:17:01) build a product plan for our favorite (00:17:03) line of socks for Zava. Claude builds (00:17:06) the plan by using Work IQ to find (00:17:09) context about market details and (00:17:11) potential pricing. And then it uses the (00:17:13) Zava branding skill to create potential (00:17:15) taglines that feel on brand. Next, we (00:17:17) use Claude to create a PDF by using the (00:17:19) PDF skill. And this quickly creates a (00:17:22) beautiful, fully designed executive (00:17:24) summary to share with the team. And then (00:17:26) finally, and this is cool, we'll use (00:17:28) M365 to send the PDF back to the two of (00:17:31) us. So this is Microsoft's enterprise (00:17:34) knowledge combined with Claude's agentic (00:17:36) capabilities and skills. (00:17:38) >> The thing I love the most about this is (00:17:40) Claude skills working together with Work (00:17:42) IQ. I feel like we're the only two (00:17:44) companies in the world that can bring (00:17:46) that type of use case together for (00:17:48) enterprises. (00:17:49) >> Absolutely. And this goes beyond (00:17:50) knowledge work to coding as well. So (00:17:52) thanks to our partnership with GitHub, (00:17:54) Cloud has become a first class agent, (00:17:56) which means your developers can (00:17:58) collaborate with Claude as part of (00:17:59) GitHub's agent HQ. So example, you can (00:18:02) start by assigning Claude an issue just (00:18:04) like you would a teammate. Claude (00:18:06) instantly picks it up, creates a branch, (00:18:08) opens a pull request, and then from (00:18:10) there you interact directly with that (00:18:12) PR. You can ask questions, request (00:18:14) changes, iterate together. When it's (00:18:17) ready, you can merge, and that's it. So (00:18:19) this is Claude working alongside you (00:18:21) like a co-orker directly in your team's (00:18:23) workflow. It'll write code, learn from (00:18:25) your feedback, and help you move faster. (00:18:28) >> Amazing. You know, you have been so (00:18:31) thoughtful about your model (00:18:32) advancements, safety, everything you're (00:18:34) doing in coding and enterprising and (00:18:36) everyone you partner with. Um I'm sure (00:18:38) everybody here and everybody at home (00:18:40) would love to know why is our (00:18:41) partnership different than anyone (00:18:43) else's? (00:18:43) >> From the very beginning, there just felt (00:18:45) like there's been a lot of shared DNA (00:18:46) and trust across both companies. And I'm (00:18:48) really excited about what happens when (00:18:50) you combine the power of our trusted (00:18:52) models with Microsoft Foundry and M365. (00:18:55) So we're making Cloud available to (00:18:57) enterprises that have already invested (00:18:58) in Fabric, in Foundry, in M365, and (00:19:02) removing barriers like navigating (00:19:04) separate vendor contracts and billing (00:19:05) systems so developers can get started (00:19:07) with Claude right away. I can't wait to (00:19:10) see what you all build. Uh stay tuned (00:19:12) for much more to come. (00:19:13) >> Thanks so much, Mike. (00:19:14) >> Thanks, Asha. (00:19:15) [applause] (00:19:20) Now what's unique about this moment is (00:19:22) that now Azure is the only a hyperscaler (00:19:26) to offer both open AI and entropic (00:19:28) models to fully empower our customers to (00:19:31) build the best AI applications on the (00:19:34) planet. (00:19:37) >> [applause] (00:19:40) >> So with all this choice, picking the (00:19:43) right model for the right job to be done (00:19:45) becomes even more critical. That's why (00:19:48) today we're also announcing the general (00:19:51) availability of our new model router. (00:19:56) This model router automatically selects (00:19:58) the best model based on accuracy, (00:20:00) performance, cost or balance. It (00:20:03) enforces US and EU data boundaries by (00:20:05) default and customers are already seeing (00:20:07) 50% lower latency and 15% higher (00:20:12) quality. (00:20:14) So with models multiplying tools (00:20:16) everywhere, we've never had so much raw (00:20:19) power. But raw power is not enough. (00:20:22) Models don't understand your margins. (00:20:24) Tools don't understand your customers. (00:20:26) And most solutions only understand the (00:20:28) what of the work, not the why. (00:20:32) The why lives in documents, lives in (00:20:34) chats and meetings and web pages and (00:20:36) files. And in every organization, this (00:20:39) sits in two oceans of data. One that's (00:20:41) structured and one that's unstructured. (00:20:44) And between them lies this intelligence (00:20:46) gap. Agents should not sit on top of (00:20:49) these systems. They need to live in (00:20:51) between them. They need to be able to (00:20:52) see across silos, reason through (00:20:54) context, and act on behalf of people. (00:20:58) That's why today we're also announcing a (00:21:00) new intelligence layer. One that knows (00:21:03) how your company works, understands your (00:21:05) business logic, and can take real (00:21:07) actions in the real world. (00:21:10) It is made up of three connected planes. (00:21:13) Work IQ, Fabric IQ, and Foundry IQ. (00:21:19) Now, Ryan talked about work IQ. So, (00:21:21) let's start with Fabric IQ. Fabric IQ (00:21:25) builds on 20 million semantic models in (00:21:28) PowerBI that encodes your business (00:21:30) logic. Now those models are going to (00:21:33) extend beyond analytics and into (00:21:35) operations so humans and AI can share (00:21:38) the same understanding of your (00:21:39) organization in real time. (00:21:42) And because Fabric IQ is integrated with (00:21:44) Microsoft 365, (00:21:46) every single user benefits in PowerBI, (00:21:50) in Excel, in Teams, in Copilot, all with (00:21:52) security flowing automatically. (00:21:55) To take us deeper, please welcome Zia. (00:21:59) >> Thank you, Asha. Imagine this. You need (00:22:02) to understand if Zava should open a new (00:22:04) pop-up store. Our store operations (00:22:07) dashboard has an agent built on Foundry (00:22:09) taking advantage of our data in (00:22:11) Microsoft fabric. You start by asking (00:22:13) for markets with high demand signals, (00:22:15) but the agent didn't understand what (00:22:17) demand signals meant or where the more (00:22:19) relevant demand data is. Results by just (00:22:22) dropping an agent on top of your data (00:22:24) can be mixed at best. AI needs an (00:22:26) understanding of what the data means and (00:22:28) how things relate to take further (00:22:30) action. With Fabric IQ, you can adopt (00:22:33) the language of your business defining (00:22:35) entities and connect them across all (00:22:37) your data in one lake from real-time (00:22:40) data, semantic models, and even mirror (00:22:42) data from sources like Snowflake, (00:22:44) Oracle, or Google BigQuery. You're not (00:22:46) just analyzing data. You're giving AI (00:22:48) and your people a shared understanding (00:22:50) of what it means so they can act on it (00:22:52) in real time. Similar to what semantic (00:22:55) models did for business intelligence, (00:22:57) Fabric IQ is doing for AI, making (00:23:00) results consistent, accurate, and (00:23:03) transparent. With the model built, you (00:23:05) can now see a live unified view of your (00:23:08) organization and visualize how (00:23:10) everything connects from relationship (00:23:13) graphs to supporting reports and other (00:23:15) business data. You can also trigger (00:23:18) rules to take action and leverage agents (00:23:20) to help run your business through (00:23:22) automated decision-making. Using (00:23:24) Fabric's geospatial capabilities, you (00:23:27) can instantly see how shipment delays (00:23:29) impacts demand across regions and where (00:23:32) opportunities are emerging. After adding (00:23:34) Fabric IQ as a knowledge for our agent (00:23:36) and foundry, it looks like Boston is the (00:23:39) strongest candidate for the next pop-up (00:23:41) store. We now have confidence in the (00:23:43) results provided by our AI solution. It (00:23:46) has a deeper understanding of our (00:23:48) business and provides much more relevant (00:23:50) guidance. And look at that jump in (00:23:53) quality. Same question, same agent, but (00:23:56) now it understands us. Adding our (00:23:59) business knowledge turns a generic (00:24:00) response into something actionable and (00:24:03) tailored. Now that we've published (00:24:05) Fabric IQ, it's ready to fuel other (00:24:07) solutions and any agent, including (00:24:09) Foundry. From unified data to unified (00:24:12) intelligence, this is IQ in Microsoft (00:24:15) Fabric. (00:24:17) >> Thank you so much, Zia. [applause] (00:24:22) Now, if Fabric IQ is shared business (00:24:25) logic, Foundry IQ is the intelligent (00:24:27) connection point across all of your (00:24:30) structured and unstructured knowledge (00:24:32) that your agents rely on and take action (00:24:34) on. (00:24:35) This is the first large-scale (00:24:37) implementation of context engineering (00:24:39) extending rag built for agents. And (00:24:42) unlike traditional rag, Foundry IQ (00:24:44) doesn't just retrieve, it plans, it (00:24:47) reasons, and it iterates across work IQ, (00:24:50) fabric IQ, blob storage, the web, and (00:24:53) more. (00:24:56) So to put it to the real test, let's (00:24:58) build a multi- aent system endtoend live (00:25:02) on stage. Please welcome Elijah to bring (00:25:05) it all together. (00:25:08) [applause] (00:25:11) >> As a developer at Zava, my job is to (00:25:13) make sure that our store ambassadors can (00:25:15) focus on customers while our AI agents (00:25:18) watch the store. Let's dive in. (00:25:22) Here we can see our workflow with three (00:25:24) different agents. We're going to talk (00:25:25) about all of them today, but first I (00:25:27) want to talk about our Zava store (00:25:29) operations agent. This is the agent that (00:25:32) knows everything a store manager could (00:25:33) possibly need. Let me show you how I (00:25:35) built it. (00:25:37) Every agent in Foundry starts with a (00:25:39) model from the Foundry catalog. Foundry (00:25:42) gives us sameday access to the latest (00:25:44) open-source and Frontier model releases, (00:25:47) so we're always working with the best (00:25:49) the moment it's available. The Foundry (00:25:52) catalog even lets me compare two models (00:25:54) against each other to see which one is (00:25:56) best for my use case. Today I'm going to (00:25:58) be using DeepSeek V3 which gives me the (00:26:01) perfect blend of speed and accuracy for (00:26:03) our low margin retail environment. Now (00:26:06) let's go ahead and create our agent. I'm (00:26:08) going to name it Zava store operations (00:26:10) and I already have it ready. So let's go (00:26:11) check it out. So you can see here first (00:26:13) I gave the agent some instructions and (00:26:15) then I connected a knowledge base using (00:26:18) Foundry IQ. We're going to talk a little (00:26:20) bit more about Foundry IQ here in a (00:26:21) second, but first, let's give it a real (00:26:23) task and help plan a sale for the (00:26:25) upcoming CIM marathon in Sacramento. I'm (00:26:28) going to use the starter prompt and kick (00:26:30) that off. While that's running, I'm (00:26:32) going to use Foundry IQ. Our knowledge (00:26:34) base and Foundry IQ contains a bunch of (00:26:36) different pieces of information (00:26:37) including our product catalog, um, store (00:26:40) inventory, employee handbook, and local (00:26:42) market intelligence. Foundry IQ and has (00:26:45) it's possible to do this from a using a (00:26:47) wide variety of uh data sources (00:26:49) including Azure Search Index, Blob (00:26:51) Storage, Bing, SharePoint, One Lake, and (00:26:54) Fabric IQ. Fabric IQ, which Zia just (00:26:57) showed us, gives our agents realtime (00:27:00) business context on our store (00:27:01) operations. Now, if we go back to our (00:27:03) agent, we can see here that it created a (00:27:06) full plan for our marathon sale, pulling (00:27:08) in info from previous events using (00:27:10) Foundry IQ. But let's take it a step (00:27:12) further. I'm going to ask, "What should (00:27:14) I do if there's an item on the floor in (00:27:16) the store?" And what's going to happen (00:27:18) is our agent will then pull from Foundry (00:27:20) IQ to pull information such as our (00:27:23) safety policy documents and other (00:27:25) information to give us the answer. (00:27:27) Foundry IQ supercharges our agents to be (00:27:29) able to make critical business decisions (00:27:32) quickly. So, not only does this agent (00:27:34) help us plan, but also react to what's (00:27:37) happening in the store. Speaking of (00:27:39) reacting, let's go to our next agent, (00:27:41) which our video processing agent powered (00:27:43) by the Foundry video indexer. This agent (00:27:46) watches our store in real time and emits (00:27:49) inventory events like if there's low (00:27:51) inventory or if whoops, somebody knocked (00:27:55) something on the ground. I don't know (00:27:57) who knocked that shoe box over, but (00:27:58) we'll come back to that. Zooming in on (00:28:01) this agent, you'll see that Foundaries (00:28:04) open ecosystem lets me run Microsoft (00:28:06) agent framework, langraph, which I'm (00:28:08) actually using here today, or other (00:28:10) frameworks on agent service. This is all (00:28:13) possible through our new hosted agents, (00:28:15) which handles the deployment, life cycle (00:28:17) management, and autoscaling for you. (00:28:20) Foundry's open ecosystem also gives me (00:28:22) access to over 1,400 enterprise ready (00:28:25) tools and MCP servers. And because it's (00:28:28) all Atoa compatible, it interoperates (00:28:31) seamlessly with agents builtin Foundry, (00:28:33) Copilot Studio, or any other compliant (00:28:36) system. Let's go ahead and check out our (00:28:38) store camera feed. And if we see here, (00:28:41) there is a shoe box on the ground. And (00:28:44) it it says right here, and our agent (00:28:46) picked it up saying hazard shoe box on (00:28:48) the floor. Pretty cool. So now finally (00:28:50) I'll add our policy and notification (00:28:52) agent which uses our new agent memory to (00:28:55) remember previous conversations and (00:28:56) personalize messages to store managers. (00:28:59) And so here we can see our agent memory. (00:29:01) Pretty great. And so returning to our (00:29:04) workflow, if the video agent deems an (00:29:06) event requires action and I just got a (00:29:09) team's ping. We're going to get back to (00:29:10) that here in a second. So first our (00:29:12) video processing agent will monitor the (00:29:14) situation and if it has an event that it (00:29:17) feels like it requires action it'll pass (00:29:19) that to the Zava store operations agent (00:29:22) which will in using Foundry IQ will add (00:29:24) context around store layout safety (00:29:26) policy and risk level and finally the (00:29:28) policy and notification agent man not (00:29:30) notifies the store manager and Asha I (00:29:33) don't think you sent me that ping so (00:29:34) let's go see who it is and if we flip (00:29:37) over to Microsoft teams there it is (00:29:40) cleanup on aisle 6. Our workflow just (00:29:43) notified us that there's a white (00:29:45) cardboard box currently lying on the (00:29:47) floor. Fully automated real time end to (00:29:50) end. We just saw how easy it is to build (00:29:54) multi-agent workflows with the allnew (00:29:56) capabilities of Microsoft Foundry. Only (00:29:59) Foundry brings together multi-agent (00:30:01) fleets, hosted agents, and live (00:30:03) enterprise knowledge with work IQ, (00:30:06) fabric IQ, and Foundry IQ across any (00:30:09) store, any region, and at any scale. And (00:30:12) it works with any model, any framework, (00:30:14) and any tool. And with that, back to (00:30:16) you, Asha. (00:30:18) >> Thanks so much, Elijah. (00:30:20) [applause] (00:30:24) >> With every platform shift, it (00:30:26) redefineses how we create. (00:30:29) Computing made it personal, cloud made (00:30:31) it connected and AI makes it ubiquitous. (00:30:34) And that is changing how every company (00:30:36) needs to operate. What you saw today (00:30:39) expands what work can be. Every person (00:30:42) can now shape intelligence to fit their (00:30:43) job, their team, their ambitions. And (00:30:46) they can stop scaling by effort and (00:30:48) start scaling by capability. (00:30:51) And they can do that with the tools that (00:30:53) they already know across M365, Copilot (00:30:56) Studio, GitHub, Fabric, and Foundry.