We are in a moment unlike any other that is bigger than any one product, company, or technology. It's about ensuring the benefits of AI reach us all. That's why we're building a frontier ecosystem, one where you own your intelligence, continuously improve the outcomes that matter, and accelerate your advantage in a world that never stops moving. This is how we create technology that amplifies human ambition and turn intelligence, grounded in trust, into real-world impact. More people building, more people innovating, more people shaping the future together. Welcome to Microsoft Ignite.
All right. Good morning. Good morning. It's fantastic to be back here at Ignite, and welcome. We have a fantastic show for you lined up for the rest of the week, and I want to kick it off by highlighting what we are going to unpack throughout the week. This week is about two things. First, how do we capitalize on all this new capability that's growing exponentially? And second, how do we build the capability inside of our organizations to deal with the consequences of all this capability that's growing exponentially? So last week, in some sense, was really a reminder of how critical digital infrastructure is as the dial tone for the world. We know the trust you place in our services and take that responsibility extremely seriously. And in that context, I know we have work to do to provide that resilience that you can count on, the world can count on, and we are committed to ensuring that we learn and we continuously improve. And I also want to recognize-- yeah. But more than that, I want to recognize the hard work that each of you do inside of your organizations to ensure that continuity, that resilience. And so, really, a big, big thank you from the bottom of my heart for all that you do for your organizations and the entire ecosystem.
I always find it grounding helpful to always start with our mission, which is to empower every person and every organization on the planet to achieve more. And in today's world, that means one singular thing, which is how do we collectively build out that frontier ecosystem? We need to be able to parlay all of these technological advances into real outcomes, new points of leverage for people, the organizations that we build, as well as the society at large. That's the goal. And for us, that comes down to three things: one, keeping the human at the center of it all, ensuring that you have ownership over your organization's intelligence and learning and trust. AI in the flow of human ambition is perhaps the biggest design challenge we have in front of us. Right? How do you do that? How do you make sure that individual roles, organizations, systems, and products are all being designed to amplify human intent and ambition and agency, so that it can flourish going forward? And we are seeing this change, in some sense, unfold when you look at, say, a software engineering team today. As an example, developers are not just writing code or doing PRs, but they are building these systems around code that give them new leverage. You can think of software engineering scaling from building software artifacts to building software factories that build those artifacts. The important thing here, though, is to remember still that that human intent, that judgment, that experience, skills, expertise, all very much matter. And it's not just about the point -- it's mostly about the change in the point of leverage. What you did before changes, but your expertise matters, your skill matters, your craft matters. And this is not just limited to software teams. Every function will go through a similar change across their work, their workflow, and their work artifacts.
The second pillar is intelligence. In the age of AI, the key question is how do individual firms continually build and own their learning and tacit knowledge? For an economy to remain an economy where many firms exist and thrive, intelligence, by definition, needs to be broadly distributed versus just collapsed into a couple of models. And the third pillar is trust. I mean, it's an interesting time, right? Scaling laws continue to work. You have all these emerging capabilities from each scale up. And yet, we find ourselves in a place where alignment is elusive, mechanistic understanding of these models is elusive. And so the question is: how can you even trust all this powerful technology? And the reality is, by definition, you can't. And in other words, it becomes an architectural imperative to separate the supply of intelligence from the authority over it. And so our goal is to build a stack that addresses these three architectural imperatives.
So let's start with the first one: Copilot. Copilot is the new operating system for work. It brings together all of the form factors of human-computer interaction, whether it is Chat or Cowork or Code or Autopilot, all into one experience. It also brings together the disparate categories that we have all referenced: knowledge work, coding, business process applications. Apps are just plugins to Copilot. That's why we, in fact, have brought all of Office, all of its richness, built into Copilot. And underlying this is a model system that matches the right intelligence to the right context and the right form factor to accomplish any task that you may have. Home gives you one place to start, whether you need a quick answer or want to really complete a complex, multi-step task. And code brings you an infinite SaaS factory to every individual inside the organization. You can build your applications, your dashboard, and much more, and all of this can be hosted, managed securely inside of your tenant. An autopilot is an autonomous agent that keeps moving your work. Twenty-four by seven, it continues to work in all the places where you and your teammates collaborate today. You essentially have an additional teammate. Underlying all of this is this very powerful database, Work IQ. You can think of this as, perhaps, the most important data asset you have. It has a deep understanding of the people, its relationships to other people, your emails, your meetings, chats, projects that make up all of the fabric of work. And today, we are thrilled to announce that all of this that I just talked about is broadly available to you to use inside of your organizations.
And so we are going further. We are extending Copilot to specific roles. That means if you are in sales, legal, finance, service, we are organizing the skills, the connectors, the plugins to all of the business applications so that you can access them from the Copilot surface areas. This includes, in fact, third-party systems, too. In fact, we are announcing today business plugins for Dynamics 365. It's just not Dynamics. We are excited to partner with Salesforce to bring Salesforce to Copilot. And you'll hear a lot more from Judson shortly. And, of course, we also recognize the need to continue to have the right form factors, the ergonomics for heads-down work, whether you are an analyst in Excel, a legal professional in Word, a designer in PowerPoint, a scientist using Discovery, a doctor or nurse using Dragon, or software developer in GitHub. Now speaking of GitHub, at Universe a couple of weeks ago, we announced two new primitives. Projects lets you define a goal, connect one or more repos, and have Copilot coordinate task sessions and recurring work, all driving towards that goal. It's pretty magical. And second, we're also creating a new composite model family in GitHub Copilot Fusion. It orchestrates multiple models to deliver frontier-level results at a significantly lower cost. In fact, just today, Copilot Fusion was ranked as the number one coding agent on Terminal-Bench, scoring five points over Claude Core at less than half the cost. It's exciting to see that. And, of course, we're going beyond that as well. Up until now, when you think about Copilot, it's been about the single-player world of people and their coding loops, or agent loops, and SWANs. But time has come for a true multiplayer form factor, where people, their agents, and other people are driving towards that common goal with shared context. This is Copilot Workspaces, let's take a look.
Yeah, I'm really, really excited about -- You know, even in the few weeks that I've been using it, it's really changed my working habits. Just like how when you first had Chat show up and it became the place where you started all work, I think Workspaces today are going to be very similar. What you just saw, though, was just what points to, essentially, organizational intelligence. All that work we have done building out that database in Work IQ lights up. The cloud enabled us to move client software to have this very rich representation of organizational data, enabling deep collaboration. So the cloud was about moving from just client software to essentially this cloud data gravity that allowed that collaboration. And now AI is allowing us to take that data that you have in the cloud and connect it to intelligence. And that's what we're doing with Microsoft IQ. The idea behind IQ is to help every organization build, retain, and continuously improve your own intelligence across the organization. It brings together your action space, that means all of your data estate, all of your application estate, the context, that means all of the IQ that you have created over the years, and the additional intelligence that you have with a multi-model system. Essentially, we are taking the system we built underneath Copilot, and we are externalizing it as a first-class platform that embodies your organizational intelligence as an asset that you can continually own, continually build on. We are bringing together all of the data you have, whether it's the data underneath Microsoft 365-- that's Work IQ, Fabric IQ, Foundry IQ, the data across the web with Web IQ, and of course, third-party context too. In fact, Microsoft IQ comes with hundreds of new connectors we're announcing today, so that agents that you build can act with full enterprise context where businesses get done. And all this manifests as simply as business plugins that can be shared and inherited across the entire organization. And finally, there is a multi-model system. Now, for us, it's about making sure that we are able to map intelligence to a given task with the best economics. That's the optimization problem we are trying to solve. This is something we can uniquely do in the context of the real-world task by tuning the product, the harness, the tools, and the model system all as one end-to-end optimization.
Now speaking of models, we are very excited about a new class of models called Decision Models. These models allow you to embed intelligence into software as a core primitive. They are quickly becoming key to how software is built. In fact, just like how databases back in the day, or relational databases back in the day, changed software engineering, I think of decision engineering similarly. And today, we are thrilled to announce Microsoft Decision 1.1, which is now generally available and ready for enterprise deployment. We have added multi-modal support to it. We have improved performance, making decisions more accurate and responses faster. Already, over 4,000 customers have used it. In fact, Epic is using it to make decisions across their patient experience, clinical care, as well as their back office apps. Copilot system itself uses it to route users to that right mode, model, the effort level with millisecond precision. And our Xbox team is using it to sort tens and thousands of open-ended pieces of feedback into what people are saying about all the different game launches. And here's the fun thing, it's more than 10x faster and 200 times less expensive than using an LLM. And Foundry, of course, has all of the popular closed and open-weight models. In fact, today we are very excited to bring GPT-6.2 to Foundry. It's a great new model, sets a new bar for AI professional work on a benchmark for complex professional work. GPT-6-2 Astra outperformed its predecessor by more than half the estimated cost. And we also brought Claude Fable 5.2, Mythos 5.2 last week. And today, we are also announcing Meta's latest models are all coming to Foundry as well. So a great choice of models for you.
And, of course, we're also innovating at the frontier with our own model family, the MAI model family. That's both compute-efficient when it comes to pre-training and token-efficient when it comes to the deployment. Our goal is to put additional points on the frontier curve, really informed by these real-world RLEs around task performance that matters to you inside of all the products. And we have great momentum; MAI Transcribe 2 requests are up 12x since it launched, it launched at OpenRouter. MAI Code Flash on GitHub Copilot last month had 3 million users. MAI Cyber now powers over 40% of the scans across Perception code scanning. And we are going further. We are announcing MAI Licensed Weights, meaning trusted partners will now be able to get access to Microsoft AI model weights through our Frontier Co. And we are committed to doing what it takes to ensure that everyone can own their own intelligence and thereby drive the broad benefits of this platform shift. And today, we're also introducing lots of new models. And Mustafa will come up and talk a lot more about this in more detail. But I wanted to highlight two in particular. First, Code. We're introducing a new flagship family of agentic coding models, Code 2. Code 2 is at the frontier of cost performance for everyday enterprise coding tasks. It excels in taking code to production, multi-file implementations, repo-wide reasoning, and UI at a fraction of the cost. And Code2 Flash is designed for fast, high-volume work like code explanations, targeted edits, rapid iterations. And it really brings near flagship quality with much lower cost. And we are taking the same approach with Cyber as well. Cyber 2 Flash is our most capable security model yet. It's flash-sized model that delivers performance ahead of frontier models at a fraction of the cost per task. And just as we have done with coding, we are co-designing the model, the tools, and the systems all as one system. And we are integrating Cyber 2 with Perception. And once you have the models and your context in Foundry, our managed agent harness is available to you to be able to deploy and build these long-running agents. The same harness we use to power Copilot and GitHub Copilot is now available to you. And you can connect it to a continuous improvement loop that learns from experience. You get to define your own private benchmarks and evals. And then Foundry measures your agents against them and improves the agent or the model underneath it based on what you see in production. That means every execution is an opportunity to learn, improve, and deliver a better outcome. We're creating ultimately, that always improving, always self-healing system. That's how your organizational intelligence becomes an asset and more valuable over time and becomes your IQ.
So now let's move to the third pillar, which is all about trust. It starts by making sure that you have observability at every layer of the stack. That's where Agent 365 comes in. It's the control plane that extends your existing governance, identity, security, and management systems to the agents across your enterprise. We already have 50 million agents in Agent 365. And it includes FinOps, too. As AI scales, every organization needs a clear line between what it spends and the value AI creates inside the organization. And that's why today we are very excited about announcing Agent 365 Business Analytics. You can simply think of this as an ERP system of AI usage. You can see exactly what's driving the spend. You can connect the spend to the outcomes that drive business value. You can optimize the ROI across different scenarios. And the next thing I want to talk about is how do you take an environment you've set up from Copilot to IQ to all the agents registered in Agent 365 and secure them? That's where Perception as our security stack comes in. We are building this end-to-end trust platform. We are helping customers both secure their AI deployments and use AI to strengthen their security posture. From Agent 365 and Entra, giving every agent a unique, durable identity, to MXC and Cloud Containers providing strong tool and network boundaries, to Purview governing all of the data and information flows and Defender disrupting and containing threats in real-time. In fact, we saw a fantastic example of this recently. An independent security researcher used ARTEX, that's in the news everywhere nowadays, an autonomous AI attack framework, to test an environment protected by Defender. So the agent kept trying to move laterally across the network, but Defender automated attack disruptions really kept containing those compromised accounts and isolating devices, even as the agent continued its attack. Attacks like this one are, unfortunately, increasingly agentic and autonomous. And that's why I think it's so important that we have that machine speed in our own agentic response. That's why we built Perception as this complete agentic security system. And the best way to think about Perception is that you have these Red team agents that simulate attacks, Blue team agents that detect attacks, Green team agents that fix and remediate. A good example of a blue team agent would be what you had in Defender that protected against the ARTEX attack. And another example would be a Pen Test Red team agent. It takes, say, an entire surface area of something like a web app, the auth, the APIs, injection attacks, business logic, and creating all of the evidence you need to confirm that you're not exploitable. In fact, it's not just any individual agent. It's the playbooks you can compose, teams of agents to tackle the complex jobs that trigger and take action autonomously. Perception also helps essentially connect the signals, the sensors, the actuators across all of the assets, right? You have identity, your client, your cloud, your apps, of course, agents now, and really secure your end-to-end environments. And that's why I am very excited to share that Perception is now generally available.
At the end of the day, we believe there should never be a choice between innovation and safety. That means in addition to the work we have been doing across our products, we are also working across the ecosystem to ensure that everyone can innovate with speed and scale safely. Today, we can see almost everything that models touch, but you cannot see what happens inside a model. And really, what's worse is that we know that what a model tells us about itself is not even independent evidence and, therefore, cannot be trusted. And one model monitoring another model has the same issue. So you essentially have these unacceptable nested set of black boxes, which we now need to make transparent and verifiable. That's the technical challenge. And so, therefore, we need standards that allow us to examine what AI sees and does, verify its actions independently, and control what it's allowed to do. And that's why today we are also announcing the Microsoft AI Transparency Initiative. There are three parts to this very important initiative. First is we will expand our collaboration in NIS Agent Standards Initiative, drawing on its research to help the development of open industry-wide standards for AI transparency, security, and governance. Second, we are taking an ecosystem approach. That means together, we will enable AI systems that are accountable, independently verifiable, and safe by design, regardless of which model it is or who the provider is. Already, 40 organizations are working with us on defining and applying these standards. And third, we are going to incorporate the same standards across all our products. That means our MAI models, Foundry and Copilot. Our ambition through all of this work, the research work, the product work, our ecosystem work, is to help establish an open industry standard that makes AI safer, more transparent, and more accessible.
Now, underneath every one of these experiences that I've just talked about is infrastructure to deliver them at global scale. And when it comes to infrastructure, it's always shaped by the primary workloads of any new platform era. Right? That means these are being shaped. Today's infrastructure is being shaped by large-scale training and, of course, now agentic inference workloads at scale. A great example of this is what we are seeing with GitHub. GitHub is seeing exponential increases: the pushes, the commits, the actions, the PRs, all being driven by agents. And we're rebuilding, that means all of GitHub's core, starting with the Git architecture, simplifying the system primitives. In fact, decoupling storage and compute, minimizing the coordination by sort of shrinking that critical path of a push to the smallest step of integration. After doing all of that and testing-- in testing, we pointed, I think, 5,000 agents to one repo, and we're able to sustain 7,200 writes per second. That's 180x the current throughput. That's the kind of improvement in scale we need for an agentic workload. And, of course, what worked for GitHub now needs to be pervasive throughout the Azure platform. And that's what we are doing in Azure for Agent Primitives. We now have more than 500 plus data centers, 80 Azure regions. And we are continually working to make them more efficient, more sustainable, and partner deeply with the communities in which we operate. We absolutely believe that technology should strengthen and not strain any community and the systems on which they depend on. And that belief underpins our community-first AI infrastructure commitments. And we're doing all of that sustainably, from matching 100% of our annual global electricity consumption with renewable energy, or reusing or recycling 92% of all the servers and components, and of course, all of the cooling designs that conserve water. And customers are also asking for more control over where and how they run their most important workloads, and that's where Sovereign Cloud options provide a continuum of security, control, resilience capabilities without compromising on innovation. And, of course, we are also excited about Windows extending all of that infrastructure to the Edge. We now have very capable models that can run locally. And we really see the opportunity with Windows to deliver unmetered intelligence. We are working across the hardware ecosystem to bring those capabilities to a new generation of PCs with new devices. You can run, in some sense, last year's frontier models on your laptop and last quarter's models on your desk. And we are also rearchitecting, of course, the core compute, storage, and networking primitives in Azure to execute, connect, share context at that agent scale I talked about when it comes to GitHub. On top of all of this, you have these micro VMs now that can come up in less than a second with full sandbox isolation guarantees on top. And all of this can run inside of confidential computing enclaves. You can then put an agent grid on front of it that brings identity and policy-aware networking controls, where each agent -- you can sort of control what each agent can access, where it can connect, in fact, which other agents can it communicate with. And then this context cache will even accelerate and model loading. Of course, you can now then take all this and wrap it into AKS or AKS Flex or even hosted agents inside a foundry, giving you all of the agentic primitives for you to be able to build your next generation of applications.
We're also reimagining the data layer. Agents need databases that can keep up with their speed and scale, while giving them the context they need for all of the reasoning and actions. And that's why we are building HorizonDB, our new distributed -- yeah, it's exciting -- new distributed managed PostgreSQL database with ultra-low latency, scalable reads, built-in Search, and AI capabilities. I mean, just the performance is mind-boggling. It's 8x faster performance than self-managed Postgres. And today, I'm excited to share that HorizonDB is generally available in Azure. And this next part is what I've been working on for, I think, 25 years of my life, which is we are bringing all of HorizonDB into Fabric. This represents the complete unification of transactional and analytical data for agentic era. Think about what this means for a second. In the past, you had your transactional databases running your business and then a separate set of analytical systems to understand what was happening. But now imagine an e-commerce business processing millions of transactions, every order, every purchase, every inventory change, all happening obviously in real time. And then with HorizonDB and Fabric coming together, agents now can reason over those live transactions along with petabytes of historical records in real time. And now you can just simply go to Copilot and just tell it, "Tell me what's happening in my business right now." And Copilot can create a live dashboard of your real-time enterprise. So this is really, really exciting that this vision that we have had for 20 plus years is now finally coming together and for you to be able to build the next generation of systems.
Underneath all of this is also unbelievable systems in silicon innovation. We are building a heterogeneous fleet, a platform where we are able to tap into the best innovation coming across the ecosystem. And we are also advancing our own hardware systems, co-designing the data center, the servers, the silicon, all mapping to our workloads. We just announced Maia 300, which will deliver over 40% more tokens per watt compared with leading GPU systems. And Azure Boost and Azure Integrated HSM take that critical infrastructure and security functions and put them into purpose-built silicon. And today, we are very excited to announce the general availability of our next generation of Cobalt 200 VMs, which are our highest performance general-purpose VMs with leading price-performance features. Thank you so much for that. So that's just a very quick walk across all of the layers of what we are building. And when you look at all of what is coming, what is clear is we are entering an era of ubiquitous intelligence. But the key imperatives for us are to ensure that human ambition remains central, a firm's control of its own intelligence remains central, and trust in technology that gets broadly diffused is maintained. This is the promise of the Frontier ecosystem, and that's what we are focused on building together with you. And now, let me hand it over to Judson to take all of this and show you how we are bringing it to life. Thank you all very much, and have a fantastic Ignite.
Well, good morning, everyone. Good morning, and welcome to Ignite. It's awesome to be with you all here in the arena, and everybody watching online. The rest of our keynote this morning is going to be anchored around these three durable assets of frontier transformation and the journey that lies ahead for all of us: human ambition, intelligence, and trust. And we're going to start with human ambition. Microsoft Copilot is a platform for human ambition. It enables amazing everyday AI experiences for all of your people, and on top of that, high-valued experiences using the best state-of-the-art models. Simply stated, anything you can do in Claude or ChatGPT, you can do better with Copilot and for less. Now my ambition since the beginning of my career has always been about helping people and organizations solve problems using technology. So I thought a great way to start off this keynote would be to show you how I use Copilot to support my ambition. So to do that, let me take you back a couple of Sundays ago, very specifically Sunday, November 8. I've been super busy preparing for this event, meeting with customers. And, let's just say, I've been a little bit behind on my email and the Teams messages I've been getting outside of the context of Ignite. So, I'm here in Microsoft Today, which is an awesome new feature inside of Copilot. It appears both in the Outlook canvas as well as in the Copilot app. The cool thing about Today is that it scans my entire email inbox, all of my Teams messages, and shows me what's most important. Now you'll see a lot of things redacted and grayed out here. That's because this is my real environment and not a demo. And so we have to hide a few things. But here you see that Today has surfaced up this email from my Chief of Staff, Caroline Sanderson. You'll see this email is calling to attention that in January I have the lead keynote at NRF, which is retail's biggest show, and I really haven't done anything to prepare for it yet. So I'm going to go into Copilot and actually get started on some preparation without bothering my team because they're super busy working on this event. So I went into Copilot and I created a project. Projects are super cool because they allow you to chat with your business and use Cowork for advanced tasks. So I can use all the models that I need and only the models I need to get real work done. You'll see that I kick things off with a simple prompt. If you look here, I've just said basically, "Tell me what's going on at NRF and what the hot topics are going to be." If you scroll down, you can see that Copilot has done some really cool things here. First, it has scanned all of my Work IQ, all of the external information, and it's used hybrid intelligence here on my Surface Laptop Ultra, which means it's pulled in content that's local to my device, my speaker notes that I downloaded before finding out that I was speaking at the event. And so what it's done here is it's pulled together a great synopsis. It says there are seven big topics ranging from agentic commerce through to the return of the physical store experience and also supply chain and sustainability. So I chat back with Copilot and further prompt, "Hey, Copilot, given that Ignite is just two months prior to this event, wouldn't it be great if we took the seven stanzas of NRF and merged them with the three here at Ignite: human ambition, intelligence, and trust." So Copilot goes off and creates an awesome outline for me. The cool part about this is Office is embedded in Copilot. This is not a sidecar. This is the full-blown Office. You see all of the sensitivity labeling carrying through my ribbons and even the ability to track changes. You can see that it's got some news around Dynamics 365 supply chain. So that's really cool. I can close that out and then actually open up a richer body of work and use Copilot Cowork to take this a bit of a step further. You can see the prompt I've added here is a bit more robust: "Tell me more about what's happening, why people love retail or retail experiences, what the retail craft is all about, and build me a storyboard using Word, a presentation I can share with my team once we get done with Ignite, so we can get started on the keynote prep, as well as a tracking document that shows me the top customers attending NRF and what they're doing with Microsoft." -- because we want to stay situationally fluent as we deliver the content. So here's what it built for me. It multitasked. It used the models that it felt were best for each task, producing great work in the most efficient way. So you can see the PowerPoint presentation it built for me. Again, this is PowerPoint. I can go in and edit this document, make it to my liking. You can see the rich storyboard that it built and that it picked up on my notion to really pull out the human ambition here, why people love retailer, what is the merchant's craft in this whole retail experience. And it's giving me a minute-by-minute outline of how to run the keynote that I can share with my team. Now, that context for them is so powerful because normally, we would have to spend meeting after meeting so that I could get those ideas out of my head and into the context of a show. So this is super beneficial. Lastly, it's pulled together this customer tracker in Excel. Again, there's things redacted here because this is from Work IQ and Fabric IQ, pulling together what we do with these customers and how important they are to us as partners, whether they're speaking at NRF. So fantastic. There's one last thing I've asked Copilot Cowork to go and do, and that is to take this awesome storyboard and turn it into a 3D model using Code. And I'll come back to that and show you what it did. But meanwhile, I want to breathe life into this whole thing. I want to keep the project going in an ongoing way. So to do that, I bring in my autopilot. Now, I've named my autopilot "Amelia" because I'm a little bit of a flying junkie. And Amelia is going to go do some really cool things for me. She's going to do two major bodies of work on an ongoing basis. She's going to keep that spreadsheet I just showed you up to date with everything that's happening in the market, what's most important to my customers. And she's also going to attend all of my product meetings. Big part of my job is taste-making our products and working through product roadmaps. So Amelia is going to keep track of everything we're working on and help me understand what's best to present at NRF once we get to January. She's also going to inform the rest of my team about all of this work. So with that simple prompt, Amelia's gone off and created routines for herself, so she'll do these things on an ongoing basis. Amelia doesn't just work for me. She can work here in Teams. Here you see the NRF planning channel that I've created. She's told the team the three things she's going to go do. And what's really cool is in Copilot, I can set up a workspace that maps to this NRF planning channel in Teams. And the really cool thing here is, I've got autopilots and my people working together; my autopilot Amelia, my teammates' autopilots, and they can work on curating this content. Again, this is all on a Sunday, not bothering anybody, me working on my own to help prepare the team once we get done with Ignite. So, if we step into Code, I'm going to take that Markdown file that Copilot Cowork built for me. And you can see here I've just given it a simple prompt: "Build me a 3D demo based on the Markdown file." I can actually give it some cues about the colors and the fonts and the background and taste-make it to my delight. And you can see the demo that it's built here. Really pretty cool. Here's a story about how human ambition comes to life in the retail experience. Here's the intelligence behind all of it: the data sources, the things that belong to the retailer, their unique IQ, the shopping agents, merchandising agents, logistics agents, and the trust platform that enables the retailer to manage, govern, secure, and get business value out of all of these assets, and all three layers coming together. Now, the really cool thing about this demo is that even though I didn't write a line of code, there is real code behind it. So I can hand this JavaScript file to my professional developers. They can dump it into GitHub Copilot and plumb the backend with databases and hit the ground running. Now, again, the awesome thing with all of this is I did all of it without bothering my team on a Sunday. I spent a couple hours doing it. I can track my usage. If you look in here, you can see that on that Sunday, November 8, I spent about 10,000 credits, or about 100 bucks. Now two hours of my time, $100 worth of tokens, Copilot using only the models it needed to use, optimizing the experience, and my team hits the ground running for NRF without me bothering them. So that's how Copilot empowers my ambition. To hear more about how Copilot can empower your ambition, please join me in welcoming to the stage the Executive Vice President of Copilot, Jacob Andreou.
Thank you. Thanks, Judson. Now what's so cool about what you just saw is that wasn't a special build. This is how we're getting our work done in Copilot right now. And you'll notice what he didn't have to do. He didn't jump into a developer tool to write code. He didn't jump between five different applications. He did everything right in Copilot. Now, a ton of what you just saw is already in your people's hands, and some of it is coming soon. But let's start with Home. Home is where Chat and Cowork come together in a single composer. This means that whether you have a quick question or you're kicking off an entire project, Home is the perfect starting point. And you also saw Office. We've all seen Anthropic and OpenAI embrace documents, slides, and spreadsheets, but their simplified viewers and editors force compromise on power and collaboration. And real work needs both. Knowledge work already lives in Word, Excel, and PowerPoint. It's where your team is already collaborating. And that's why we brought the full functionality of Word, Excel, and PowerPoint directly into Copilot. This means that you can collaborate with AI and your team at the exact same time. This includes track changes in Word. It includes real advanced formulas in Excel, and, of course, the ability to at mention a teammate to bring them into the doc that you're working in. This means that your whole team doesn't have to blow up their workflow, doesn't have to switch which apps they're using, and all the files stay at rest in your tenant with security labels and compliance policies fully intact. The best part is all of Office in Copilot is included in your Copilot subscription. That means that Copilot is the only way to draft a PowerPoint presentation without spinning a meter. You also saw Judson use Code. Code allows anyone to describe what they want to build in their own words. Copilot picks the best approach, whether it's a tracker, an app, or a dashboard, and it builds a full working app. Code is fully sandboxed locally or hosted securely in your tenant by default. It makes hosting and sharing as easy as sharing a Word document. Now, none of this innovation matters unless it really holds up at-scale. And few are at the scale of Chevron. And so I'd like to welcome Ryder to show us how the new Copilot is transforming work at Chevron.
Thanks, Jacob. At Chevron, powering the world to enable human progress is our purpose. Society depends on us to provide reliable, affordable energy, and to do this while protecting people and the environment. One of the ways we achieve this is through a engineered basis of design. It's the blueprint our engineers use to ensure we execute and operate safely and reliably. However, building this can take months. So to improve this, we turn to Copilot, where I typically start and end my day. Let me walk you through the demo. We started in Copilot Chat and asked it to write a Cowork prompt to build a basis of design. You can see here's the prompt. It's 300 lines of technical instructions to Cowork. And what enabled this is Work IQ in the background sourcing relevant files. And with one click, I'll send this over to Cowork, and work begins. Cowork's already gone through hundreds of engineering files. It's highlighted to the engineer right here where they need to focus their time to build out the basis of design. And here at the bottom are first two initial drafts: a basis of design document and an Excel workbook. However, these two documents were based on a relatively small data set compared to our complete enterprise data. So next, scroll up here, we ask Cowork to enrich both deliverables with our engineering standards and with all of our relevant data across the company. And what enables this is Fabric IQ. You can see here in the bottom right corner in our plugins, Fabric IQ. And without changing any of our data architecture, it's compared our basis of design to thousands and thousands of company-operated designs and operational reports. So I'll scroll down, takes just a second here, to our deliverables. You'll see our two updated deliverables here. And I'll click on our basis of design, over 80 pages of technical information on a basis of design, based on our company template, informed by our company standards, and pulling in from across the company, data and our institutional knowledge located in multiple systems across the company, including a system from a company we recently acquired, all enabled by Fabric IQ. I'll close out of this and get back into our session. And now that we've completed that task, we ask Cowork to turn all of this into a reusable skill. And I'll go ahead and just scroll down to the output. You can see it working here. And here it is, bottom of the screen, assets, basis of design, skill. Now, anyone on the engineering team can use this skill. In fact, any engineer in the company can use this skill to create a basis of design. We're almost finished. We go back into Cowork. We give it a prompt to summarize the work we've done, post the two deliverable files into our engineering team's channel for further review. And with one click . . . we're done. And I never left Copilot.
So, that's really awesome, Ryder. So I bet the question, though, that everybody's asking in the room is, how much did all of that work cost you?
Well, Judson, costs always matter. But in Cowork, I can hit "forward slash cost". Let's see. Oh, in this session, 4,500 credits, or around $45, which is significantly less than if I had stood up a coding team to build this as an agentic solution. And our engineers' time went from months to minutes.
Fantastic. Thanks so much, Ryder.
So, you saw how Ryder visualized his cost with just one simple command. But it's really cool to be able to see this from the entire perspective of the IT organization. I'm here with Steph Waters from Chevron's Enterprise Data and AI team. Steph, show us what this looks like across all of Chevron.
Will do. Agent 365 gives IT visibility into the AI spend across the organization so we can see where investment is going and manage it at scale. At a glance, I can see what we've used, what remains, and what's driving spend across our services, agents, policies, and models. And I can drill down into our users. And here I can see Ryder. And I can see the models he's using as well as the services and the use cases and the value that he's getting for using Copilot.
Now importantly, IT needs to set the spend limit before anyone can use Cowork. And once that limit is reached, a user must request an increase to continue to use Copilot. That prevents runaway spend. And alerts let IT step in before any issues occur.
I can also set which models are available for use, as well as create ones from templates. So ultimately, I'm not just managing cost. I can see the value that we're creating and make better decisions about where to optimize and where to invest next.
Super. Awesome job, Steph. Thank you. Jacob, back to you.
Those guardrails help keep cost in control, but Copilot works to maximize how far each of those dollars go. We do that in three ways. First, we do it with the Copilot subscription. The Copilot subscription is the best way to give everyday AI to every person at your company. It includes Chat, all of Copilot, and Office, which means that you can do drafting, summarizing, and analysis, all without spinning a meter or generating additional charges.
Everything you saw is also powered by auto routing. Auto routing selects the right mode, the right model, and the right effort, all from a single composer. That means that you get the best possible response. And when you're using Cowork, we help manage the costs of each request under the hood. It also benefits from Work IQ. Work IQ grounds everything that you do in your company context and how you work. The most impactful way to keep costs under control is to make sure that you have the context to do the job right the first time. So Copilot knows your work. It's able to pick the best model for every task. And all of that is included in a high-value fixed subscription. So whether it's everyday AI in everyone's hands or advanced AI for your very hardest tasks, Copilot saves you money. This is AI built for work.
So everything you've seen so far is available now. And I want to take a second to pivot to some of the really cool things rolling out over the next couple weeks. This is when Copilot goes from helping you with your work to actually moving your work forward. Judson started this off by showing you Today. Today is our new proactive, personalized experience in Copilot. It's a command center that spans across Mail, Calendar, Teams, Meetings, Tasks. It shows you what you missed, what needs your attention. And Today doesn't just surface the work. It actually helps move things forward. The draft is already written. The proposed schedule change is already made. Everything is waiting, always for your approval.
You also saw Autopilot. Autopilot is your digital teammate. It's the first ever enterprise-grade long-running agent. It lives in your tenant. It has its own identity, its own memory, its own email, and its own computer. And that means that Autopilot is not just limited to M365 apps. It can also use all of the legacy applications that your business runs on. Autopilot watches channels. It can follow up on threads. And it can push recurring work forward, even while you sleep.
We're also really excited to announce Workspaces, which you saw in Satya's keynote at the top. The future of collaboration is bringing people and AI into one place. And Autopilots come to life in workspaces. Until now, AI has been mostly single-player. And workspaces makes it truly multi-player. In a workspace, people and agents come together around a common goal with shared context. The Workspace automatically pulls in the files and the people needed to get something done, so that conversations and decisions all live in one place. Everyone has full context, they know where things stand, and they know what to do next. We think all of these things can make Copilot an amazing way to serve every person in your business.
But what sets your company apart likely lives in the functions. It lives in legal, in sales, in finance, in HR. Each of these functions in your company has its own systems, its own data, its own way of doing things. Copilot is a platform for human ambition, and plugins is how it supports each function in your business. It brings together the skills, the connectors, and the tools that each of your teams relies on. Plugins is how Copilot stops being a generalist and actually becomes an expert in the job to be done. IT approves plugins once, and it works everywhere Copilot does across Chat, Cowork, Code, and even Autopilot. Please welcome Andrew from BlackRock to show you how they're putting the new Copilot to work. Andrew.
Thank you. Thank you, Jacob. Sales is a relationship business. It's about knowing your clients, understanding what matters to them and bringing the best solution to them at the right time. Today, the challenge isn't having the information to do that, it's connecting the information across conversations and systems. And that's where AI gets really powerful. It brings those signals together and gives our Relationship Managers a much better starting point, so they spend more time with clients and less time searching for information.
At BlackRock, the opportunity isn't simply adding AI to the way sales works today, it's reimagining how we manage client relationships. Now we are working with Microsoft to shape the Copilot and Dynamics experience around the needs of our teams and our clients to enable our sellers to mine insights, take action, and accelerate commercial outcomes. So let me pass it to Nirja to show you what we're working towards.
Thanks, Andrew. Hi, everyone. I'm Nirja Trivedi from Microsoft. Let me show you what we're building with BlackRock. So let's say I'm a Global Sales Lead. My day starts in Copilot, specifically in Today, the command center for my work. Here, information from Dynamics 365 Sales is pulled in, along with information across my work systems, as well as my business systems, like financial health and pipeline performance. So here, for example, I see from Dynamics 365 Sales that my team is at 73% of their goal. In order to hit our goal, we need about 3x pipeline coverage, but we're only at about 2.2x, so we have some work to do. Copilot spots this gap and recommends a cross-sell play around our new private markets fund. So I'm going to dive into this a bit deeper with Cowork.
Cowork will pull in information across Dynamics 365 Sales and services plugins, like my service history and client information, to identify the top clients that are most likely to commit. These clients represent about 110 million dollars of potential commitment. More than that, the top clients are in decisions and could commit by December. So I want to turn this insight into action. So I'm going to jump back over to Cowork. I'm going to ask Cowork to help me build a package for each client with the standard information from BlackRock's knowledge base, a personalized email, and an invitation to meet with our private fund specialist in New York. I also want to account for some of the travel time in between. Lastly, I want to send each package to the client's Relationship Manager to review.
Cowork will pull this information from Dynamics 365 to route it to the specific client Relationship Manager. More than that, nothing goes out without that Relationship Manager's explicit approval. And if I want to, I cannot only read from the CRM, I can write back with my approval as well. To see this in action, I'm going to hand it over to Efe.
Thank you, Nirja. My name is Efe Abugo, and I am a Relationship Manager on Nirja's team. In my Today, Copilot has pulled Nirja's request right to the top of my view with those three personalized emails to send out to my clients. And below, I can also see other recommended actions I could take today. So I'm going to hit "Start Task," and Cowork is going to begin building out those three emails for me to send. And Work IQ pulls in meeting and email context from Microsoft 365. And with Dynamics 365 client activity and history, I'm able to actually see personalized emails based upon conversations we've actually had.
Now, two of my clients are actually going to be in New York next week, where our Private Fund Specialist will be. So Cowork has proposed meeting times based upon his actual availability. Now all three emails look good to me, so I'm going to hit "Send" on all of them, and they're all going to go all at one time. Now, this is a lot of work to stay on top of. So to help me keep track of everything, I am going to put Autopilot to work. Autopilot will work on my behalf, keeping me updated on key items, as well as helping me move critical work projects forward.
So here, I'm going to ask autopilot to help me track my fundraising outreach opportunities, client news, and service cases. And if a client hasn't responded within a week, I want it to draft a follow-up for me to review and approve. And additionally, I want to get a weekly view of relative fund performance and firm-level news. So here we see Autopilot has created the two routines. First, we have that daily check where Autopilot will look across Dynamics, Outlook, and Teams, and we have that Monday morning intel brief. And the guardrail is clear: Nothing goes out to my clients without my approval. Autopilot is operating within the supported controls IT has already put into place. And if we come out here and navigate to the right, I can see all of the routines I currently have running, including the two that we just created together. So at any point, I have visibility into the actions Autopilot is performing on my behalf. Now, let's turn it back to Nirja to see how things are going.
This is great, thanks Efe. All right. So it's been about two weeks, so let's check in on the progress of my team. I can see in just two weeks, we're up from 73% all the way up to 81%. We also have $59 million of new pipeline generated from this play. So you can see here that Copilot really becomes the front door to my business processes. Client knowledge that once only lived in sellers' heads is now intelligence my whole team can act on. And with Dynamics 365 plugins, sales and service intelligence is now directly in the flow of my work, so I can keep the momentum to have a strong close to the quarter. Back to you, Judson.
Thanks, Nirja and Efe. Really appreciate it. Well done. So, Microsoft Copilot, again, is a platform for human ambition. And plugins enable you to wire in these experiences for your people and how they work, whether they're sales professionals or other walks of life. Now, you would have expected us to have plumbed that demo with our own assets to see Copilot working well with Dynamics. But what you might not have expected is that we're doing the same thing with Salesforce. We're announcing three major new product innovations in conjunction with Salesforce today. First, Copilotforce, which is the new Salesforce plugin for Copilot, the Salesforce Cloud Sales Cloud in Teams, and the integration of Data 360 in Fabric with zero copy. To share more about that, please welcome to the stage President of Engineering, Rohan Kumar.
Rohan, welcome back to the Ignite stage.
Well, thank you, Judson, and good morning, everyone. After nearly 28 amazing years at Microsoft, it's wonderful to be back at Microsoft Ignite, now representing Salesforce as we bring together our power to shape the future of enterprise AI. It's clear that the foundations of computing are changing. And people actually had to adapt to software, navigating applications, moving through systems, and actually piecing together their work. And what we're seeing, essentially, is the ability for AI to become the interface of the future, applications becoming capabilities, and enterprise users basically spending their time focused on their business outcomes that they are trying to seek instead of navigating software.
You know, this is exactly the vision we landed at Dreamforce earlier this year when we introduced AI Force, bringing the power of Salesforce closer to wherever you choose to work. And for millions of business users across the world, that place is Microsoft 365. Salesforce has your complete customer context, business understanding, and workflows. Microsoft 365 is where your people collaborate and get their work done. Why should they not work together? And as Judson mentioned, we are so excited to introduce Copilotforce. We are bringing the best of Salesforce directly into Microsoft 365 Copilot and Microsoft Teams.
With this ability, it's not just about your CRM data. It's about your complete trusted customer context, your business understanding, and the ability to take action, all with Salesforce security and permissions. So instead of asking AI what's the next best action with the customers, you can actually get it done, get it executed using Copilotforce. This is the part of the agentic enterprise: intelligent, open, and interoperable. And we are thrilled to partner with Accenture, both Salesforce and Microsoft, as we build Copilotforce to bring this part together. And we are thrilled that they are deploying this across their sales teams. Back to you, Judson.
Thanks, Rohan. That's right, you can clap for that. So, we're really excited to have Accenture share what this is like in their environment. Please welcome Tony Leraris, the Chief Information Officer of Accenture.
Thank you, Judson. At Accenture, we all work in Teams. It's where more than 800,000 of our people meet, collaborate, and increasingly use Copilot to deliver impact every day. Now, with Copilotforce, our sellers will have customer context built right into the tools they already work in. Accenture is a design partner, and we're customer zero. We helped build it, and we helped prove it on ourselves first. Let me show you what that looks like. Here's Priya, one of our sellers. She relies on Cowork to track her sales priorities for the day. Copilot reasons over her entire Salesforce pipeline, from open opportunities to the latest customer activity on each account, and within moments briefs her on the top deals and where to focus.
First up, a new RFP from Kestrel Foods that needs her attention. With a quick tap, she uses Cowork to set up the deal room to tackle the whole pursuit, the opportunity in Salesforce, a Teams channel, a close plan, and a weekly deal sync in everybody's calendar. Day one, and we're already organized. Now, let's jump to the deal room in Teams. Here, the team is grounded on the customer opportunity, with the RFP and the close plan ready to go. Let's hop over to the Teams chat, where Priya has shared with the team that they've made it to the shortlist. Without needing to switch apps, and directly in the chat, she asked Atlas, her Autopilot, to update Salesforce. Using the Salesforce plugin, and grounded on Work IQ, it can pull insights across CRM emails and chats and meetings, and proposes five updates to send to Salesforce. One click, and Salesforce is current.
Now, this isn't just for sellers. A week later, Priya's manager opens Copilot to see how her team's updates have moved the forecast. Using the pipeline review skill, which is one of more than 38 Salesforce skills available in Copilot, it combines CRM data with context from email and Chat to give an instant snapshot of a region. You can see it flags the two biggest deals at risk with a recommended next step for each. With a simple confirmation, she sends the account manager a nudge in Teams tailored to the opportunity in Salesforce. That's sales leadership in the flow of work. Copilot, Teams, and Salesforce working as one. With Copilotforce, our sellers will spend less time on process, more time with clients. And by proving the value on ourselves first, we can help every company do the same. Judson, back to you.
Thanks, Tony. Intelligence is the second durable asset that we all need to build on the frontier transformation journey. Not artificial, yours. We all need to build, retain, and continuously improve the assets inside of our organization so that intelligence compounds from within. Last year, we launched Work IQ, Foundry IQ, and Fabric IQ to release to you the fundamental building blocks for this intelligence platform. Now, over the last year, we've been thrilled to see so many customers adopt this same concept. You'll see a Lilly IQ, a Toyota IQ, and a Levi IQ. We really want to bring this whole concept to life for you by showing you Unilever IQ. Please join me in welcoming to the stage Deepalok Roy Chowdhury from Unilever to talk to you about Unilever IQ.
Thank you. Thank you, Judson. Every day we have 3.7 billion consumers use a Unilever product, and that helps us generate 20 million orders and 5 million claims a year. At that scale, we need intelligence at the core. At that Unilever IQ, how our people operate, how our business runs, and how we generate value for our customers. Sharoon, you want to showcase the trade claims agent we have been building?
Absolutely. Thank you, Deep. Here is the trade claims agent. It works where our team works. You can see it has already processed 14 claims on its own, and now it's asking me for help for the Northwind claim. Now, let's take a closer look. Looking at the first line, you can see that the customer is billing us. However, we notice a discrepancy with the contracted rate, more than $2,000 over.
Two thousand dollars on one line? I don't think a human would have picked up out of the millions of lines that we look at. And so, I am sure the agent did some magic to get this going spontaneously. How did it do that, Sharoon?
Great question. The agent in this case, ran a three-way match. You can see that it looked at the deal sheet from Foundry IQ and matched it with the claim file and point-of-sales data from Fabric IQ. Now, if I'm working on this claim, I'm trying to understand: Well, has this happened before? So now, you can see that the agent goes through all my past claims and finds four similar claims with the exact same pattern. But even more interesting, it looked at Work IQ to find the people behind those decisions, and that is Unilever IQ in action: our contracts, our data, and our people, all in one place. So Deep, should Unilever pay this high rate?
Hey, hold on. I don't want an agent to make a judgment. Let's bring in some people in and have the agent and people work collaboratively. So let's have Daniela call up Jordan, who is the Account Lead for the Northwind account, and see what he has to talk about it.
Exactly. So seems like Jordan has already responded here, and he's recommending to settle this as a one-time adjustment. Now, I can apply the override here, but here is the best part: Jordan's judgment doesn't just stay in a Chat thread, it's saved back to Unilever IQ so that the next time this comes up, the agent knows how to handle it and also helps with auditing. Now, finally, for the four items that are still pending, the agent drafts a letter leveraging all of this context from Unilever IQ and clearly articulates whatever is needed to close this claim out. That's nearly $12,000 protected in value on this claim alone.
And $12,000 on one claim? Think about the millions we save on the other claims if we find such discrepancies. Great job.
That's right. Now, all of this comes back to Unilever IQ. Every decision our team makes is added here. Our best analysts' tacit knowledge is now part of it, and every claim makes it smarter. Back to you, Judson.
Thank you. Well done. So we're taking this a step further today by launching Microsoft IQ. It's your enterprise context management system that enables you to create these durable business plugins that you can use with Copilot and across your agentic estate. Now, of course it works with GitHub, Foundry, Fabric, and Azure, but it is also open, model-diverse, and heterogeneous at every layer of the stack. Here to tell you more about it, please welcome to the stage EVP of Azure, Arun Ulag.
All right. Thank you, Judson. What Unilever built, and what every company here can build, are agents that understand exactly how your business works. Today we are excited to announce Microsoft IQ, the enterprise context management platform that allows you to bring together all of the context across your entire organization to a single place, so that it can be used consistently across all your agents. This includes all your collaboration data, both humans to humans and humans to agents. It includes all your business data, your unstructured data, data from the web, as well as context from pretty much any enterprise system that you already use.
It's really easy to get started because you already have the building blocks. For every Microsoft IQ customer, Work IQ makes your collaboration data available. This includes your Teams chats, your email, your documents, your spreadsheets, your SharePoint sites and so much more. WorkIQ enforces the permissions in Microsoft 365, so every agent can only see what it's supposed to see. For every Power BI customer, Fabric IQ makes all of your metrics and KPIs that you've already defined in your Power BI semantic models available to your agents, again, with all of the security and permissions intact. Power BI already connects to all your data, not just from Azure, but from AWS, GCP, Databricks, Snowflake, and so much more. So it's ready to go without having to rebuild your semantics in yet another platform.
With Foundry IQ, you can build knowledge bases that bring together everything your organization knows, from contracts and precedents, to technical manuals, case studies, and beyond, making that knowledge accessible to agents. And when you need to selectively bring in content from the internet, for example, regulatory guidelines or the news on specific topics, Web IQ can take care of that for you. Microsoft IQ goes beyond the Microsoft ecosystem, with 350 plus connectors to third parties, including ERP and CRM systems. You can bring your entire business context into one unified platform. And it works wherever work happens: in Copilot for your people, in Foundry for your agents, and because it's completely open, in assistants like ChatGPT and Claude as well.
This context is your intellectual property. It's what makes your organization unique. It is your IQ. Our goal at Microsoft is to ensure that your IQ stays yours. And it compounds in a way where the benefits accrue to you, not to Microsoft, not to the model companies, and definitely not to your competitors. So now, I'm going to invite Eren Orbey to show you a demo, so you can see Microsoft IQ in action, unifying context for the business process, packaging up the context into a business plugin, and using it seamlessly in Copilot. Eren, over to you.
Thanks, Arun. You just saw what a trade claims agent can do with the right context. And Microsoft IQ is the platform that helps enterprises manage that context. Our company's knowledge used to be scattered across different sites, handbooks, systems, and imaginations. And now, it's consolidated and codified here in a brand-new interface as our IQ. These are business processes that we've minted with building blocks from Microsoft plus context from our own systems. So let's take a look at that trade claims process. It sits under legal and compliance standards that apply company-wide. The legal team has vetted certain skills and policy documents that point agents to canonical sources like these. And then within legal, we have guidelines for contracts, and then specifically for commercial contracts. And down here in trade claims, we automatically inherit from that hierarchy.
So everything above flows through: skills, documents, data, and official web sources. And then our claims team can add bespoke agent skills for checking a charge or escalating exceptions to the right approver. These skills sync up with the responsibilities that we've defined for people in our organization. And we've pulled in the live Teams thread where we triage escalations day by day. Foundry IQ comes in to help us index files like contracts and deal sheets for agentic retrieval. And Fabric IQ brings in operational and analytical data with OneLake from across clouds in open formats, plus semantic models like this one, which define metrics and measures atop that data. Every source here can be customized with instructions telling agents when and how to use it. So they load only what they need and no more, saving tokens along the way. And as a subject matter expert, I can easily evolve this process by bringing in assets from our enterprise and beyond.
Of course, there's a Copilot experience to help me evolve the process further. And all of this context can be permissioned so that only select stakeholders on our team can view and edit it. At the top level, this context can also be certified by IT as a source of ground truth for agents, as we see here. And put it all together, and our trade claims context can be published as a durable, interoperable business plugin. It's governed in Agent 365 and accessible in Copilot, in Foundry, and in other applications that we use. So here's what that means at the most essential level for a claims analyst. When I ask Copilot why one retailer's claim looks high, our IQ steers it straight to trusted sources like these. The response that we get charts the charges that don't match up according to our data and our contracts. It lays out next steps based on our process. And for anything that needs human judgment, Copilot drafts an email to the approver specified by, you guessed it, our context. With Microsoft IQ, every agent and every analyst can follow the same process that we defined, getting answers grounded in in-house expertise. That's the power of curated enterprise context. Happy Ignite, folks. And Jay, over to you.
Thanks, Eren. Now, if you ask any CIO about agents, you're going to hear the same thing: building them is easy, but running them in production is not. A production agent needs things like identity, authentication, sandboxing, access to enterprise data, monitoring, red teaming, and the list goes on. And that's where most agents stall. So the question isn't how to build an agent, it's how to build and run thousands of them, trust them, and make them better every day. We built a foundation to help you do that: a comprehensive end-to-end platform. GitHub is where you build, Fabric is where your data is organized, and Foundry is where your agents run in production and keep getting better.
Today, I'm going to cover a couple of key features in Foundry. Let's start with models and model choice. Foundry has more than 11,000 models, including OpenAI, Anthropic, Meta, Mistral, open-weight models, and our own MAI models. But the number isn't the important point. It's having one place to evaluate them and switch when a better one ships without having to rebuild your agents. Now, with our model router, you don't have to pick just one. Simple requests go to small, fast models, complex ones go to the frontier. Our data shows that this is up to 65 percent cheaper than sending everything to a frontier model. This allows you to strike the right balance between the quality of the output and the cost of getting to that output.
The next big question is: Once I have my agents up and running, how do I improve them over time? After all, your business isn't static. You're working hard to improve it all the time, and your agents should be doing the same. Foundry helps you do that in three ways. First, agents on Foundry get better the more you use them. Foundry allows you to feed agent traces and user input back into the system, creating a continuous learning loop. Second, Foundry looks across every run for ways to make your agents better. The system finds and tests improvements, shows you how each one trades off across quality, speed, and cost. And then you get to decide what ships. Finally, Foundry provides a range of other capabilities that allow you to improve progressively; from lightweight skill improvements to agent configuration, all the way to training your own model using your own data. And all of this learned intelligence stays with you and only you.
Alaska Airlines runs this kind of loop, and they are here with us today to show us what they've built. Three things to watch: first, how they define better. Second, how far optimization takes them. Third, what training adds. Please welcome Charu Jain from Alaska Airlines and Seth Juarez from Microsoft.
Thanks, Jay. Airlines are in the business of time. Being reliable and running on schedule is how we earn your trust and deliver a remarkable travel experience. Our approach to innovation starts with the guest. We look for pain points across your journey and work to solve them. So we're working with our Seattle neighbors at Microsoft to put powerful tools in the hands of the people who are closest to our operation. What you're about to see starts with routine maintenance. Every maintenance report gets an ATA code that helps us track the work across our fleet. The more precise and consistent that coding is, the easier it is to see patterns and identify issues we may be able to address proactively, rather than having them show up close to departure. So this is very great work, and I'm excited for you all to see it. So let's go see it.
How are you doing, Charu?
I'm good, Seth. What have you been up to?
I am so excited. When I heard about this scenario, I generated a whole bunch of synthetic data that matches exactly what you're doing. Now, you talked about ATA code. ATA stands for Air Transport Association, and they have chapter and verse of every single code that maps to every single -- you can see it up there. There's a lot of them, but it's also hierarchical, so it's fairly complex. So do you want to see the system we built to help out with that?
Okay. So this is the maintenance reliability software, and what you can see is that during a general inspection, identified the number three break is near its wear limits.
And that's routine maintenance.
Absolutely, this is routine. It's the system working. But here's the thing, right? The engineer coded it as a 32-00 in the logbook, which is technically correct. It's part of a landing gear. But 32-40 turns out to be a little bit more precise.
Yeah. See, these precise codes will help us make better decisions and solve the right problem.
A hundred percent. And so since it's text coming in, we thought maybe we could make some agentic software to look at the text and see if it could figure out the ATA code. But again, the important bit is that the engineer is always in charge, so they can look at this and then confirm whether something is correct or not. Isn't that cool?
So let me show you how we built this. This is Microsoft Foundry, your place to build, optimize, and train your particular agents inside of one service. So let me go to the agents pane here. And what you're going to see is the Alaska ATA agent is right there. There you go. And what it is is it's an agent that we built, but we need to know how it's working. So there is something called "Traces" that enables us to see inside of these agents. And so I already have a Trace already open up, and you can see it right here. Notice that you can see the entire run of this particular agent, including getting your Alaska IQ, which includes these ATA codes plus some other stuff that we're going to talk about later. It's really cool.
So Seth, how hard was it for the model to understand airline maintenance codes?
Yeah, and see, this is the thing that's most important. Just because we've made an agent doesn't mean that we know it's working. So what we actually did is we created a rubric that grades this agent. Foundry did this for us, but we can validate. And let me show you three things that it checks. This is like a rubric, like any school grading rubric. And notice that the first one is correct ATA code is super important, but you also have, as part of your Alaska IQ, your internal house codes, which are amazingly super important as well, the way you do stuff, which is good. So we ran this. I made a naive agent the first time. Do you want to see the results?
Okay. So it turns out that I did not -- this is not as good as I thought it could be.
I think I could do better. Okay. So there's two things you can do whenever you're making an agent better or more accurate. You can correct the assets of the agent, like your instructions, your tool descriptions, or you can fix the model. So let's do that. Let's do the first thing first. Let's see if we can fix the assets. So, inside of Foundry, we have something called the "Foundry Optimizer" and what it does is, it looks at your agent, and it iteratively goes through fixing the system prompt and the tool descriptions to give us a better one. And notice that it actually changed a lot of stuff. It changed a lot of stuff. And do you want to see how much better it did? Okay, so here it is. Notice that it's getting way better accuracy. We were using a Frontier model, GPT-6 Astra, and I have to come clean, okay? So we did a little analysis of this thing -- buckle your seatbelt.
Oh, hold on, Seth. That is way too --
I know, but here's the thing. Here's my explanation. It turns out that because we're using so much more tokens pre-loading for the instructions, because we're giving more instructions, the input and output token cost went up a little bit. I think we can do better.
Let's do it. Okay. So the next thing we can do is we can actually fix or train a model to know more. So one of the things that I do when I'm looking at this, a hint is if you have a lot of stuff in your instructions, that could be an indication that you could train a model to have those instructions in its weight instead of the instructions. So that's what we did. If you go over here, notice that we actually did a training job on MAI Code 2 Flash, so that it would actually do a little bit better. And notice, if you look at the monitor as it goes up, you're going to see a chart that actually sees the things increasing. And are you ready for what we were able to do with this model? 97% exact ATA code match and 93 on your coding group. But here, it gets better. It gets better. Everyone, buckle your seatbelts one more time.