Satya Nadella on AI Agents, Rebuilding the Web, the Future of Work, and more

The new theory of the firm in the AI era is that firms must perfect the loop of fine-tuning, deployment, market feedback collection, and continuous improvement—not just acquiring access to foundation models.

“And that I think, is sort of the new theory of the form in my sort of, you know, sense of how one needs to go and you really are going to have to perfect that loop.”

The core question about competitive advantage in the AI era is 'what is the edge of a firm'—and the answer is that a firm's knowledge and data edge is the real question going forward.

“In some sense, what is the form, going forward and what is that edge? The knowledge edge of a form is the real question.”

The sustainable competitive advantage for enterprises using AI is not static proprietary data, but rather the ability to create a virtuous feedback loop: take commoditized foundation models, fine-tune them with internal knowledge and data, deploy outputs, capture market signals (customer feedback, marketplace validation), and reinforce the model through reinforcement learning in the real world.

“And the sustainable advantage is to get a new sample to then use these reasoning models with your data to then be able to do RL in the real world. Right? So which is the reward function from the market, reinforcing the application of your knowledge to fine tuning the new sample. Right. That's the virtuous cycle. So the world keeps getting better in terms of its model capability. Forms are about taking that. That's ultimately a commodity that you bring into the firm. You then because of all the knowledge work you do internally and the data you have, fine tune it, put out the output into the world, get the signal. It could be a customer thumbs up. It could be marketplace thumbs up, whatever. That's the reinforcement. Reward. And then you go back.”

Society celebrates tech companies far too much compared to the actual impact of technology, when the real measure should be whether technology solves concrete problems like education and healthcare efficiency.

“I just want to get to a place where we are talking about the technology being used and when the rest of the industry across the globe is being celebrated because they use technology to do something magical for all of us”

The key to successful organizational transformation is simultaneously changing three things: how we work (internal operations), what we work on (products and focus), and how we go to market (business model and customer strategy).

“How do I fundamentally change three things how we work, what we work on, and how we go to market. Right. And that muscle of fundamentally reinventing the production function simultaneously with the product and the innovation, as well as the go to market and the business model is a hard thing.”

Organizational transformation ultimately comes down to culture and capability building, which allows companies to take more shots on goal and practice the discipline of hunting for new concepts.

“I think ultimately it comes down to culture and capability building. That allows you to take more shots on goal. And that, I think, is what one needs to do as any one of us faced with these shifts”

Even massively successful products like Windows eventually stop growing, requiring the company to build new businesses, which means competitive survival requires continuous capability building and 'taking more shots on goal'.

“Even something as big a hit as windows, you know, stopped growing at some point. And so therefore, it meant that we had to build other businesses. But I think ultimately it comes down to culture and capability building. That allows you to take more shots on goal.”

Real success is when technology use across industries and globally is celebrated—when healthcare providers, educators, and other sectors using technology to solve problems are the heroes, not the technology companies themselves.

“I just want to get to a place where we are talking about the technology being used and when the rest of the industry across the globe is being celebrated because they use technology to do something magical for all of us, that would be the day.”

The story that matters is not AI companies celebrating themselves, but the World Bank study showing Copilot creating educational impact in Nigeria—this is the real reason Nadella joined the tech industry.

“That's the story to me, right? That's why I joined the tech industry.”

Knowledge workers should not be trained through formal training classes on AI tools, but rather through diffusion of general-purpose tools into their daily workflow, which naturally changes work artifacts and processes as people discover utility (the 'PC penetration model' where legal adopted Word, finance adopted Excel, teams adopted email-based forecasting).

“So a little bit of sort of what I'm seeing even inside of Microsoft, whether it's with GitHub Copilot or whether it's with the M365 copilot or anywhere else, it's just the diffusion of the tools.”

Education has been a longstanding problem where society has sought one meaningful tech intervention for decades without success, but AI agents finally offer real statistical evidence of impact, as demonstrated by World Bank research in Nigeria showing that providing people with Copilot or similar agents measurably improves educational outcomes.

“the decades all of us have spent saying, give me a one tech intervention that makes a damn difference in education, right? Finally, you have real statistical evidence, thanks to that world Bank study in Nigeria that, you know, you give people copilot their or any other agent. It doesn't matter. Something changes.”

At Microsoft, the hardest organizational challenge is not building great products but managing continuous transformation across how work is done, what products are built, and how they are taken to market, requiring constant reinvention of production function, product innovation, and go-to-market simultaneously.

“at Microsoft we are not like there's one trick company. In other words, at some level the hardest thing for us is being adjusting. But when I came to Microsoft, you know, at that time, it is unclear, like our existential competitor was Novell. And so over the years, Microsoft has had to learn in the hard way. Sometimes we've gotten it right, sometimes we've gotten it wrong. How do I fundamentally change three things how we work, what we work on, and how we go to market. Right. And that muscle of fundamentally reinventing the production function simultaneously with the product and the innovation, as well as the go to market in the business model is a hard thing.”

The workflow for CEO-level information preparation has inverted due to reasoning models: rather than receiving prepared reports from account teams through email and OneNote, executives now prompt AI agents to pull data directly from multiple sources (web, email, documents, CRM, supply chain systems) and receive consolidated reports, making them more employable and empowered.

“if I had if, say, a somebody had to get prepped for a, let's say, a customer visit, you know, you would sort of go to the account team. The account team would write up a report, the report would show up in the email that would get loaded into one note, and I would read it. In fact, that was pretty much the sort of, you know, I was preparing for most of my career. Those reports and then eventually, you know, reading those reports a lot in the last even 4-5 months, thanks to these things like reasoning models like researcher, it's completely inverted. I just go prompt myself, hey, I'm meeting with the CEO of XYZ Corporation. Pull all the stuff I need to know. It pulls from the web. It pulls from my emails, my documents. Most importantly my CRM system, my supply chain system.”

GitHub Copilot's coding agent eliminates the distinction between filing bugs and fixing them for developers; developers must now actively participate in fixing bugs alongside the agent rather than just reporting them to a development queue.

“even for me and it's no longer sufficient to for me to just file bugs. I got to get involved in even fixing bugs, right? To the, I mean, a programmer that I can work with.”

The agentic web requires a composable technology stack where every layer (from Microsoft 365 Copilot to Foundry to standards like Natural Language Web and Model Context Protocol) is open and pluggable, allowing orchestration of multiple agents and middleware that respect the original ethos of web openness.

“I feel to do that, you got to build up a real stack right where every layer is open. It composes. There are standards that protocols, and I feel like they're finally there. And so what is exciting to me is to have the stack from sort of, Microsoft 365 Copilot to Foundry, all compose, with some of these things like NL web and MCP to create this agentic web.”

The Stanford tumor board multi-agent orchestrator example demonstrates the practical value of AI agents in healthcare by automating the collection and synthesis of data from pathology labs, medical literature (PubMed), and other sources, which are then organized by multiple coordinated agents and delivered to clinicians in their existing tools (Teams).

“One of the best things, that we talked about was that Stanford demo, right? Stanford Medicine, think about it. Right. Something high stakes, like a tumor board meeting. How do you use AI to have better tumor board meetings? That's a real thing that one can go to work on. In order to make that the case, though, you really have to get data from pathology from, you know, multiple labs from PubMed and all getting orchestrated by multiple agents. Right. And then have it show up where users are. In that case, it was teams.”

An engineer in Microsoft's networking team built a multi-agent orchestrator using low-code/no-code tools to automate manual DevOps work (fiber operations) that was consuming excessive organizational resources through email and phone calls, demonstrating the power of empowering non-specialized developers with agent-building tools.

“I was talking to an engineer in our networking. I went right. I mean, I was talking about, man, the AI when is growing like crazy. The number of fiber operators we work with is like some, you know, insane number. And by the way, all of the DevOps on it, right. You know, if there's a fiber cut somewhere, it's all a manual, right? Literally it is emails and calls and so on. So she just said, okay, I can't deal with this. I have to really automate this because otherwise, you know, all of Microsoft will be just working on DevOps of fiber. And so she just built an agent and in fact, a multi-agent orchestrator. Right. And by the way, she just did this using some of the low-code, no code and the foundry tools.”

Getting tools into employees' hands is more important than formal training; diffusion of general-purpose tools naturally leads to behavior change and workflow innovation without requiring classroom training.

“one of the things that we're emphasizing is getting the tools in. Right.”

The underlying capability layer—having data, multiple models, agent orchestration layers, and reasoning models that can decompose intents into multiple model calls—is what's genuinely exciting about the agentic shift, more so than the interface layer.

“But the interesting thing is the underlying capability, right where you have, hey, you have data, you have multiple models, you have these agent orchestration layers, you have these reasoning models now that are capable of taking intents and decomposing it into multiple calls to multiple models. That's I think, the exciting thing.”

The key insight about what happens when 90-95% of code is generated by AI is not about developer replacement but about the need for a massive increase in development capacity to address the software deficit.

“yeah, it's a couple of different things, you know. One is you got to remember, like when I look at the amount of deficit we have, right... the reality is we need a lot more software development, in order to be able to really reach, the goal of being able to work down that deficit.”

We as a society celebrate tech companies far too much versus the actual impact of technology, and this misalignment prevents us from seeing where real value is created.

“I think we as a society celebrate tech companies far too much versus the impact of technology.”

The true value of AI is global economic growth and productivity gains in sectors like healthcare and education, not AGI benchmark achievements which represent 'nonsensical benchmark hacking.'

“you mentioned that AGI is just some nonsensical benchmark hacking, and the true value of AI is... global economic growth.”

Proactive agents should execute high-level user intent and allow inspection of work afterward, rather than following a task-by-task model or acting without user visibility, requiring a balance between autonomy and user control.

“I'm going to have a high level intent, and then someone's going off and doing a bunch of work and I can inspect.”

The transition from role-based UI design (applications focused on functional domains) to intent-based UI design (AI systems that understand and execute high-level goals) represents the next evolution in human-computer interaction.

“So that's where I think some of these proactive agent stuff comes in. And I want to be in control though, like, for example, that's why I think the, the cool agents, for example, the computer use agents have to sort of get to a point, one that can't be fast and better, and more reliable. But the more importantly, they have to be changed as part of some intent, of mine. Right. It's not like I'm going to go give task. Task, task tasks. I'm going to have a high level intent, and then someone's going off and doing a bunch of work and I can inspect.”

Excel is the greatest and most ubiquitous programming tool in the world because everyone learned to program in Excel, and a similar pattern will repeat with AI tools where general-purpose AI becomes the programming language.

“she just built an agent and in fact, a multi-agent orchestrator. Right. And by the way, she just did this using some of the low-code, no code and the foundry tools. Right. That ability to empower people, with tools and agency to go and look for the workflow closest to them, that I think is what needs to happen, just like how we were able to use, you know, I think I always say Excel is the greatest programing, most ubiquitous programing tool on the world. And all of us learned how to program in Excel, essentially. And that's the thing that I think is going to happen again.”

We need to build scaffolding for the AI age similar to how Microsoft built integrated platforms like Outlook (combining email, calendaring, and contacts) and Teams (combining chat, channels, and meetings), now creating a UI for AI that brings together chat, search, agents, and notebooks in one interface.

“When we say we are definitely trying to build a scaffolding for the AI age, right? Just like, say, back in the day, we built teams, right? You remember? Or even back in the day. We built outlook right before outlook came together. Calendaring was a separate app, contacts was a separate app and email was a separate app. And so then we said, oh, maybe we should build it all together.”

Knowledge work today is defined by the abstraction level at which people operate; as abstraction rises, what was specialized becomes universal (typists, slide makers), freeing workers for higher cognitive tasks.

“So therefore, I think the levels of abstraction, whether it's managing agents, if I look at even the simple, simple workflow from when I joined Microsoft in 92 to now, if I had if, say, a somebody had to get prepped for a, let's say, a customer visit...”

As platform shifts mature, the conversation moves from discussing individual applications to discussing the entire platform's architecture, standards, and ecosystem—this shift in conversation focus signals a maturing platform transition

“you start talking about the platform, not just about the one app or the few apps and how they got built. Right. Because you're about to sort of start scaling, things in a much more. I'll call a generalized way where as a developer, you can start seeing how to build it.”

Code generation by AI—even at 90-95% of new code—should be understood not as eliminating developers but as changing the nature of development work from writing boilerplate to orchestrating multi-file edits, reviewing outputs, and maintaining system integrity

“Or I can highlight a bunch of code and have it explain it to me and it'll draw a diagram. Then you know agents right. Where I can assign synchronously some tasks. So multi file edits, the full repo changes that I want done.”

How PCs penetrated the enterprise demonstrates that different roles discovered domain-specific uses: legal discovered Word for contracts, finance discovered Excel for modeling, but ultimately people realized they work in teams, leading to integrated tools and collaborative workflows.

“how PCs became standard issue. Right? Is it fascinating to study even how PCs penetrated the enterprise? First, you know, someone in legal said, wow, this is great to write contracts up because I need word, you know, with all of its footnotes and what have you finance obviously loved Excel, right? Because, wow, it can help me do models and so on. But then it turned out that people don't work in their silos. They work in teams.”

The human must remain in the loop for all AI agent operations, particularly in software development where agents must submit work for human review before any CI/CD automation occurs, preventing autonomy that could cause harm.

“ultimately the human is in the loop. I think we overstate the autonomy here, right? Because even before it does any CI CD thing, it comes back for a human review. Right. And it can be, you know, something that you can even automate with another agent. But ultimately, there is a workflow here where people inside of a development organization are going to work with these AI agents to essentially work down the deficit we have.”

The Stanford tumor board meeting case study demonstrates that high-stakes AI applications require orchestrating data from multiple sources (pathology labs, PubMed, etc.) through multiple agents and delivering the result where users already work (e.g., Microsoft Teams).

“Something high stakes, like a tumor board meeting. How do you use AI to have better tumor board meetings? That's a real thing that one can go to work on. In order to make that the case, though, you really have to get data from pathology from, you know, multiple labs from PubMed and all getting orchestrated by multiple agents. Right. And then have it show up where users are. In that case, it was teams.”

Copilot fine-tuning gives enterprises the ability to take their proprietary data and knowledge and use it to fine-tune the Copilot system, creating a sustainable competitive advantage.

“Copilot fine tuning, because the idea is to be able to take what you have is a form, which is your knowledge, your data, and use it to essentially tune, the copilot system.”

Healthcare represents the highest-priority domain for agent-based productivity gains because it represents 19-20% of US GDP and much of the cost comes from workflow inefficiency, not care quality. Multi-agent orchestrators that help providers navigate complex digital systems and reduce administrative burden could deliver better care at lower cost.

“To me, take health care. What is it, 19-20% of our, GDP in the United States is in health care. A lot of that cost the inefficiency. If you talk to any of our providers, is all in the workflow, right? They're working so hard to provide unbelievable care, except sort of, you know, and they're using all these digital systems, but they're really not able to, tame the complexity and reduce the cost.”

There exists a global software development deficit: the world has unfinished projects and technical debt that require more software development capability than exists in the market, making AI code generation tools necessary to expand capacity rather than displacing human developers.

“One is you got to remember, like when I look at the amount of deficit we have, right. If I sort of looked at, let's call what people say is tech debt or it debt, right. Which is if you look around the world and say the number of projects that are still unfinished, the reality is we need a lot more software development, in order to be able to really reach, the goal of being able to work down that deficit.”

Organizational transformation requires culture and capability building that enables companies to take more shots on goal, combined with hunting for new concepts, rather than focusing narrowly on one business or product.

“I think ultimately it comes down to culture and capability building. That allows you to take more shots on goal. And that, I think, is what one needs to do as any one of us faced with these shifts”

A stable existing business becomes a tailwind for transformation, providing the financial and operational stability to pursue new agent-era opportunities, whereas a declining business must use transformation as a survival strategy.

“If you have a stable business. Fantastic. Then this is tailwind. You can you can take and get more leverage out of it. If you have a business that's in decline, you can then use this as an opportunity to reinvent yourself.”

Healthcare represents 19-20% of US GDP, with a large portion of costs driven by inefficiency and workflow complexity rather than care quality, and multi-agent orchestrators could reduce costs while maintaining or improving care quality.

“take health care. What is it, 19-20% of our, GDP in the United States is in health care. A lot of that cost the inefficiency. If you talk to any of our providers, is all in the workflow, right? They're working so hard to provide unbelievable care, except sort of, you know, and they're using all these digital systems, but they're really not able to, tame the complexity and reduce the cost.”

Case studies of successful companies do not help because you cannot learn transformation by observing others; you must do the hard work of transformation yourself.

“The reality is, case studies don't help. You have to do it yourself right now. Somebody said to me once, which is you don't get fit by watching others go to the gym. You have to go to the gym. And that, I think, is what this change is all about, right? It's not about admiring someone else. But it's about really doing the hard yards yourself.”

Forecasting workflows changed from being a manual, multi-step process (faxes, interoffice memos, manual entry) to one where a single Excel spreadsheet sent via email allowed all area leads to enter data simultaneously, creating an integrated forecast.

“pre PC, I you know basically faxes went around. Somebody did an interoffice memo. And then eventually, you know, you got to a forecast where, you know, once the PC became standard issue, you just took an Excel spreadsheet, entered some numbers, sent an email, all your area leads entered numbers, and you had a forecast. It changed the work. The work artifact in the workflow.”

Excel is the most ubiquitous programming tool in the world, and most people learned programming fundamentals through Excel; the same pattern will repeat with AI agents as they become ubiquitous.

“I think I always say Excel is the greatest programing, most ubiquitous programing tool on the world. And all of us learned how to program in Excel, essentially. And that's the thing that I think is going to happen again.”

The human remains in the loop in AI agent workflows; agents do not achieve true autonomy. Even before CI/CD execution, coding agents return results for human review, and this review step can itself be automated by another agent.

“ultimately the human is in the loop. I think we overstate the autonomy here, right? Because even before it does any CI CD thing, it comes back for a human review. Right. And it can be, you know, something that you can even automate with another agent.”

Case studies of other companies' transformations provide no actionable guidance; each organization must conduct its own transformation work because the specifics of organizational culture, legacy systems, and market position are unique and cannot be copied.

“The reality is, case studies don't help. You have to do it yourself right now. Somebody said to me once, which is you don't get fit by watching others go to the gym. You have to go to the gym. And that, I think, is what this change is all about, right? It's not about admiring someone else. But it's about really doing the hard yards yourself.”

Computer-use agents must become faster, more reliable, and more trustworthy before they can be widely deployed, beyond just being functional.

“computer use agents have to sort of get to a point, one that can't be fast and better, and more reliable.”

There is a significant 'tech debt' or 'IT debt' problem: the world has more unfinished software projects than the current development capacity can handle, creating a supply-demand deficit in software development capability.

“One is you got to remember, like when I look at the amount of deficit we have, right. If I sort of looked at, let's call what people say is tech debt or it debt, right. Which is if you look around the world and say the number of projects that are still unfinished, the reality is we need a lot more software development, in order to be able to really reach, the goal of being able to work down that deficit.”

GitHub Copilot's code completion and explanation tools (highlighting code to receive visual flowchart representations) enable individual developers to work more effectively by adapting the interface to their cognitive preferences without requiring external training.

“I can highlight a bunch of code and have it explain it to me and it'll draw a diagram. You and yesterday I like took some code and sort of just highlighted and said, just give it to me as a flowchart, like just that simple thing I just because I do better with visual stuff.”

Even major successful products like Windows eventually reach saturation and stop growing, which is why organizations must build the capability to continuously create new businesses rather than rely on any single product as permanent revenue source

“Even something as big a hit as windows, you know, stopped growing at some point. And so therefore, it means it meant that we had to build other businesses.”

Computer use agents must be fast, reliable, accurate, and most importantly, they must be aligned with the user's intention rather than operating independently, requiring humans to retain agency and control.

“they have to be changed as part of some intent, of mine. Right. It's not like I'm going to go give task. Task, task tasks. I'm going to have a high level intent, and then someone's going off and doing a bunch of work and I can inspect.”

Platform shifts (like the transition to AI agents) require organizations to move beyond studying successful case studies and instead build the capability themselves through direct experimentation, analogous to getting fit requires going to the gym rather than watching others exercise

“Somebody said to me once, which is you don't get fit by watching others go to the gym. You have to go to the gym. And that, I think, is what this change is all about, right? It's not about admiring someone else. But it's about really doing the hard yards yourself.”

Nadella's workflow for customer preparation has completely inverted with AI reasoning models: previously he received reports prepared by the account team and loaded into OneNote; now he prompts an AI to pull from web, emails, documents, CRM, and supply chain systems for comprehensive, real-time reporting.

“if I had if, say, a somebody had to get prepped for a, let's say, a customer visit, you know, you would sort of go to the account team. The account team would write up a report, the report would show up in the email that would get loaded into one note, and I would read it... those reports and then eventually, you know, reading those reports a lot in the last even 45 months, thanks to these things like reasoning models like researcher, it's completely inverted. I just go prompt myself, hey, I'm meeting with the CEO of XYZ Corporation. Pull all the stuff I need to know. It pulls from the web. It pulls from my emails, my documents. Most importantly my CRM system, my supply chain system.”

We have spent decades saying we want a tech intervention that makes a real difference in education, and finally with AI agents like Copilot there is statistical evidence (World Bank study in Nigeria) that providing people with AI agents creates measurable change in educational outcomes.

“think about the decades all of us have spent saying, can we have one tech intervention that makes a damn difference in education... Finally, you have real statistical evidence, thanks to that world Bank study in Nigeria that, you know, you give people copilot their or any other agent. It doesn't matter. Something changes.”

A Microsoft networking engineer, faced with massive DevOps manual work (fiber cuts requiring emails and calls), built a multi-agent orchestrator using low-code/no-code tools to automate the process, exemplifying how tools empower individual contributors to solve workflow problems.

“I was talking to an engineer in our networking... the AI when is growing like crazy. The number of fiber operators we work with is like some, you know, insane number. And by the way, all of the DevOps on it, right. You know, if there's a fiber cut somewhere, it's all a manual, right? Literally it is emails and calls and so on. So she just said, okay, I can't deal with this. I have to really automate this... And so she just built an agent and in fact, a multi-agent orchestrator. Right. And by the way, she just did this using some of the low-code, no code and the foundry tools.”