The third coming of Microsoft AI CEO Mustafa Suleyman

[ResidentialBusiness]

Microsoft employees stream down a hallway by the dozen, smartphones and paper coffee cups in hand, many clad in heavy coats on this frigid February morning. The setting is idyllic—Lake Washington is in full view through floor-to-ceiling windows—but they stride purposefully. As they do, they pass a digital sign with a tersely worded call to action:

All squads ship

Competing/differentiating

Growing work every sprint to double Successful Sessions

ABS
(Always Be Shipping)

Despite the profusion of Microsofties on the premises, this isn’t Microsoft’s sprawling Redmond campus. Instead, these staffers have taken over a Hyatt hotel in Renton, another Seattle suburb. They work for a division known as Microsoft AI—MAI for short—and have traveled from corporate outposts as distant as the U.K., Switzerland, China, and India to attend a team off site.

Mustafa Suleyman, MAI’s CEO, instituted these conclaves upon arriving at Microsoft just under a year ago—part of an unorthodox mass hiring in which the software behemoth absorbed most of the staff from Inflection AI, the startup Suleyman cofounded in 2022. The gatherings take place roughly once every seven weeks, and part of their purpose is unblinking self-assessment. Ahead of the meeting, around 80 squads of 6 to 15 people apiece have rated themselves on their success in hitting recent deadlines on a scale of red, amber, or green. The results aren’t great—but Suleyman sees that as progress in itself.

“It’s taken a few cycles to get people to be basically honest in terms of their scoring,” he tells me shortly after presiding over the event’s keynote presentation. “And this time, there was lots of red—like almost 45% red. I think that was a really, really good moment; and we sort of stood up and owned it. I was very proud of the team.”

Though being CEO of something called “Microsoft AI” sounds like a job of nearly unlimited purview, Suleyman does have a more specific remit. He’s charged with using AI to transform the company’s consumer properties, including the free Copilot chatbot app available on the web and in versions for Windows, Mac, iPhone, and Android. The progress MAI measures in squad-size chunks all levels up to a much higher goal: building an AI companion that not only answers questions accurately but performs tasks based on a deep understanding of your needs. And not just an anodyne chatbot, but a warm and relatable persona you’ll enjoy spending time with.

In its present form, Copilot has barely begun to hint at this wildly ambitious vision. But as Microsoft pushes forward, “You’ll start to see Copilot become a platform that enables a personalized AI companion for you,” promises Suleyman. “It’ll have its own name, have its own visual representation, have its own personality, and really be your sidekick. What we are building is your second brain, your aide, your consigliere, your reliable chief of staff in your pocket.”

It’s lofty talk, but Suleyman—whose round, wire-frame glasses help give him the presence of a particularly glib owl—has a knack for explaining AI in a compelling fashion. His cautionary 2023 book on the subject, The Coming Wave, was a New York Times Best Seller; his 2024 TED talk, “What is an AI Anyway?” has been viewed 2.7 million times. More importantly, his bona fides include cofounding not just Inflection, but before that DeepMind in 2010. The London-based company made computing history when it created software that taught itself to play the famously complex Chinese board game, Go, better than any human. Then it developed an algorithm for predicting how proteins fold themselves, a transformative tool for drug discovery.

Both of those landmark feats reached fruition after Google acquired DeepMind in 2014; Suleyman left DeepMind in 2019 and exited Google altogether in 2022, shortly before founding Inflecton. In 2023, Google merged the company with another AI arm, Google Brain, to form Google DeepMind, with Suleyman’s fellow cofounder Demis Hassabis as CEO. The combined operation is responsible for the Gemini large language model now used in many Google products, putting Suleyman in direct competition with his former colleagues. (Suleyman says he remains friendly with Hassabis, but argues that competition fuels creativity, noting that in February, he recruited the engineers responsible for one of Google’s best-received uses of AI: its uncanny “Audio Overview” synthetic podcasts.)

In terms of raw users, Copilot has some catching up to do. According to data from intelligence company Similarweb, the consumer version—which is distinct from the one that’s part of the Microsoft 365 productivity suite—had a desktop and mobile web audience of just 15.6 million in January. That was far behind OpenAI’s ChatGPT (246 million), Chinese upstart DeepSeek (79.97 million), and Gemini (47.3 million), though ahead of Perplexity (10.6 million) and Anthropic’s Claude (8.2 million). This data doesn’t include people who use Microsoft’s free Copilot app, but market intelligence firm Sensor Tower says that ChatGPT currently has 30 times the monthly active users of consumer Copilot.

Microsoft is not wholly dependent on Copilot to reach consumers. Bing, another part of Suleyman’s portfolio, may only have 4% of the search market to Google’s 90%, according to StatCounter, but with an audience of 174 million people in January, it’s larger than any AI bot except ChatGPT, per Similarweb. His group also oversees Microsoft’s Edge web browser, which comes bundled with Windows and could become a potent AI delivery system of its own. And over time, Suleyman’s AI companion vision might give Copilot more market traction by clearly differentiating it from ChatGPT. (The two products share many technical underpinnings thanks to the Microsoft-OpenAI partnership.) Still, the relative tininess of Copilot’s current user base shows that the vision of Microsoft’s signature consumer-AI effort has yet to transform into the kind of mass attention the company cares about.

“For Copilot to be truly useful for you long term, it needs to not only be able to ingest all your long-form documents and your email and your calendar and your context, but it needs to not forget what you’ve talked about a couple of sessions ago. And we are getting really good at that now.” [Photo: Carlton Canary for Fast Company]

Then again, it’s not like anyone else has truly figured out consumer AI. The industry’s often-clumsy stabs at it—such as Google’s gaffe-ridden AI Overviews and short-lived Meta bots personified by the likes of Snoop Dogg and Paris Hilton—can feel like answers to questions nobody asked. “Every single company is trying to understand what the market wants at this point,” says Divya Kumar, Microsoft‘s general manager of search and AI marketing. “We’ve barely scratched the surface.”

But Suleyman can’t depend on Microsoft’s rivals flailing forever; if MAI doesn’t create the first true AI consigliere, somebody else surely will. Hence, the intensity he brings to managing and motivating his team, a job he describes as “building the cultural flywheel that then builds the product.” That goal was apparent in a February email to his staff—written shortly after DeepSeek’s stunningly efficient LLM shocked the AI industry—in which Suleyman predicted more surprises ahead and called for a great hunkering down.

“What MAI needs from everyone this year is extreme focus,” he wrote. “The competition will be unlike anything we’ve seen. This is for real. This is the time to do the best work of your lives.”

---

The son of an English nurse mother and Syrian cab-driver father, Suleyman landed on AI as his life’s work not because he was in love with the technology for its own sake, but because he saw its potential to make the world a better place. At 19, he dropped out of the University of Oxford, where he studied philosophy and theology, to help start a telephone counseling service for Muslim youth. He then served as a human rights policy officer for London Mayor Ken Livingston and cofounded a consultancy dedicated to driving societal change on a global scale.

Suleyman was only 25 when he and two friends, who had the training in computer science he lacked (Hassabis and Shane Legg), started DeepMind in 2010. The company was a bet on their conviction that evermore-powerful supercomputers would lead to an epoch-shifting moment when AI would surpass human cognitive ability across an array of disciplines. Legg called that phenomenon Artificial General Intelligence, or AGI.

Early on, their faith that AI‘s future would be extraordinary was so contrarian, Suleyman says, that “we never really talked openly about our ambition to build AGI—that was always something that we whispered in hushed tones to each other and among a very small group.” Eventually, though, the entire field adopted AGI as a concept—and a goal.

DeepMind indeed made historic progress, putting Suleyman at the heart of the AI revolution just as it was taking off. But in a major career setback, he left DeepMind in 2019—reportedly not by choice but precipitated by complaints from employees that he had a bullying management style. In a 2022 podcast, he accepted the criticism—”I really screwed up”—and said that he’d since worked with an executive coach to become a better boss.

By the end of his DeepMind tenure, Suleyman says, he was itching to get AI out of the lab and into the real world. Rather than leave Google altogether, he spent another two years at the company as VP of AI product management and AI policy. Among his responsibilities was working with the Google Brain team, which had developed an LLM called LaMDA. At the time, ChatGPT didn’t exist; even OpenAI‘s GPT LLM hadn’t proven itself capable of powering radically new AI experiences.

LaMDA “was at GPT-3 level performance, at least a year earlier than GPT-3,” Suleyman remembers. Leveraging it into new Google features would have been a bold, attention-grabbing move. But it also would have been risky and required sign-off from many internal stakeholders. Suleyman struggled to rally support.

“That was really on me,” he says. “I was the one trying to get this out the door—persuade the lawyers, persuade the policy people, persuade Google Search. And for some reason, there was just a series of mental blockers in the company.”

Concluding that this effort had reached a standstill, Suleyman departed Google in January 2022. Officially, he was joining venture capital firm Greylock as a partner. Barely more than a month later, however, he returned to AI with the launch of Inflection. Suleyman and his cofounders, DeepMind principal research scientist Karén Simonyan—now MAI’s chief scientist—and LinkedIn cofounder and Greylock partner Reid Hoffman quickly lined up $225 million in funding. Suleyman didn’t spell out the startup’s exact plans beyond acknowledging they involved making it easier for humans to communicate with computers: “It feels like we’re on the cusp of being able to generate language to pretty much human-level performance,” he told CNBC.

What that meant became clearer in May 2023, when Inflection introduced Pi, its chatbot. Short for “personal intelligence” and available on Facebook Messenger, Instagram, and WhatsApp, as well as the web and as an iPhone app, Pi was a rough draft of Suleyman’s notion of an AI companion—optimized for engaging conversation rather than purely informational utility. “I chatted about philosophy with it for what turned out to be 2 hours,” wrote an impressed Reddit user. “I kept waiting for it to ‘break’ and say stupid random stuff like [ChatGPT does] but it kept going coherently.”

As Pi was establishing itself, Suleyman found himself in an ongoing dialog with Microsoft CEO Satya Nadella about their companies’ respective futures. Microsoft had invested in Inflection and was providing the startup with cloud services, so it was only natural they’d talk. But by the winter of 2023—a period when Sam Altman’s brief ouster at OpenAI highlighted how vulnerable Microsoft had left itself by tethering its AI vision to an outside partner—Nadella was proposing scenarios involving Suleyman joining Microsoft in some capacity. 

Suleyman was willing to listen. Microsoft already had a deep technological platform, a large consumer footprint it knew how to monetize through advertising, and the ability to shovel its formidable resources to high-priority initiatives—assets Inflection couldn’t match on its own. Furthermore, Suleyman’s confidence in Inflection’s initial business plan, which involved building high-cost computing clusters to train its own in-house LLM, Inflection-1, had been shaken by new developments such as Meta’s open-sourcing of its Llama AI model, which made a world-class LLM available to any company that wanted to use it. “I just did not predict that a public company would make the crown jewels available to everybody,” he says, calling the realization of how that might impact the competitive landscape “painful.”

Nadella, too, had reason to reassess Microsoft’s AI strategy, particularly on the consumer front. For all the benefits the company had reaped from its investment in OpenAI, the tantalizing sense that it might help Bing bite into Google’s dominance in sudden and dramatic fashion hadn’t panned out. Nor had Copilot become ChatGPT’s peer in traffic and name recognition, despite being based on some of the same underlying GPT technology. “As of right now, it feels to me as an outside observer that they haven’t gotten nearly the leverage that they would’ve wanted on the consumer side,” says tech investor and writer M.G. Siegler.

Microsoft product manager for model personality Rachel Taylor: “The way that you show up for Gen Z versus my mom, who’s just turned 70, has differences in style and delivery, and it should feel different as time goes on.“ [Photo: Carlton Canary for Fast Company]

Suleyman is quick to underline that he could have continued pursuing his AI vision at Inflection, which had raised a total of more than $1.5 billion from investors. On top of that, in November 2023 he met for eight hours with SoftBank CEO Masayoshi Son, who “made an enormous offer to raise a gigantic round at a huge price.” But over a January 2024 lunch at the World Economic Forum in Davos, Suleyman says, Nadella made an irresistible offer: “Come and drive Microsoft through the next decade.”

”One of the most compelling pieces of what he said to me was, ‘We completely missed mobile,’” referring to Microsoft’s Windows Phone being thrashed by Apple’s iPhone and Google’s Android, Suleyman remembers. “‘If we completely miss AI in the way that we missed mobile, it’ll be existential for the future of the company.’”

Suleyman says it was a wrenching decision: “We were 70 people. The product was growing like wildfire. We had an amazing roster of investors.” But a couple of months after Davos, a deal was announced.

Acquiring Inflection would likely have raised antitrust concerns. Instead, Microsoft simply hired Suleyman, Simonyan, and most of the team they’d assembled. (LinkedIn currently lists 59 ex-Inflection staffers now at Microsoft.) Microsoft won’t say how much it spent on this gambit; reports put the price tag at $650 million—or maybe more—to license AI models, pay off investors, and compensate Suleyman and other new employees. Along with existing Microsoft employees and additional ones Suleyman would recruit, the Inflection alums are part of MAI—“a Russian doll of new teams,” he says. The company also declines to provide a current head count, but last November, Wired’s Steven Levy quoted Suleyman saying “about 14,000 people” reported up into his team. (Inflection still exists, under new management—Suleyman no longer has an ownership stake—and has pivoted from building an AI companion to selling technology to enterprise customers.)

Microsoft general manager of search and AI marketing Divya Kumar: “One of the aspects of Mustafa that I really appreciate is his consumer ethos. Everything he thinks about, whether that is product engineering or marketing, he thinks about from the lens of the customer.” [Photo: Carlton Canary for Fast Company]

Suleyman’s unsatisfying experience trying to commercialize Google’s LaMDA did leave him wary of big-company bureaucracy. “The most important thing was figuring out how, culturally, we kept some distance from some of the old patterns in Microsoft that we were worried would slow us down,” he says. “This was something Satya and I talked about a lot. He was adamant that we would have all the necessary freedoms to operate and we wouldn’t be impeded in any way.”

The evidence that Nadella was good to his word includes the fact that Suleyman divided his team into small squads of employees who tackle their work in six- or seven-week cycles, an approach he brought with him from Inflection. MAI controls its own tech stack and is responsible for its own recruiting. Despite being part of the ultimate Microsoft shop, it even does some of its collaborating over Slack (“I would say we mostly use Teams,” Suleyman emphasizes when I ask).

Nearly a year into his new job, Suleyman remains keen to shield MAI from certain Microsoftian tendencies. In a November email to staffers, he dutifully praised Microsoft’s consensus-driven culture—but stated that MAI’s objective should be “respectful disagreement, followed up by a complete commitment to the outcome.” That bias toward speedy decisiveness is tempered by the self-awareness Suleyman has carefully fostered since getting into trouble at DeepMind. In the lead-up to one team meeting, for instance, he asked his direct reports for examples of recent moves he’d made that proved problematic so that he could share them more broadly.

Suleyman still doesn’t hesitate to describe himself as “relentless.” But he adds that he’s now “much more practical and more realistic. In some ways, I’m actually a lot more patient and balanced as well.” That’s not far from how MAI staffers describe him: “He’s a kind and empathetic person who still manages to have a tremendous level of urgency,” says Kya Sainsbury-Carter, corporate VP of Microsoft Advertising, a direct report, and an 18-year veteran of the company.

---

Back at Microsoft AI’s off-site meeting, a standing-room-only crowd has assembled for a workshop on understanding Generation Z. The presentation includes TikTok videos in which young people rhapsodize about ChatGPT and excerpts from interviews with high schoolers talking about their busy lives. In an empathy-building exercise, participants discuss what they learned from this exposure to the next generation and how it might be applied to improving Copilot.

Such anthropological inquiry is serious stuff at Microsoft, which turns 50 this year and is intent on forging a relationship with consumers who weren’t around for the glory days of the PC. Gen Z “demands more than real innovation and crisp aesthetics; they want authenticity, social responsibility, and seamless digital experiences,” wrote Suleyman in a January email to MAI staffers. “We need to create experiences that truly resonate with these users.”

Copilot design director and Inflection AI alum Matt Pistachio: ”We still move quickly as a startup. We still meet together like we used to do. We still have the same vision. We’re just doing it at scale.“ [Photo: Carlton Canary for Fast Company]

Of course, at the scale to which Microsoft is accustomed—and hopes Copilot will reach—even delighting an entire generation of consumers wouldn’t be enough. The company must consider the needs and desires of individual customers spanning a wide swath of humanity, a test that some on its Windows team have likened to ordering pizza for 1.5 billion people. Indeed, Copilot design director Matt Pistachio—one of the Inflection employees who joined Suleyman at MAI—gets most animated when telling me how AI can empower his mother, a Lebanese technophobe. “She can talk in Arabic,” he explains. “She can talk in her broken English. She can just say what she wants and she has access to computing.”

As another Inflection alum, MAI’s product manager for model personality Rachel Taylor, puts it, “Your AI should feel different than mine.” So how can Microsoft even begin to attack the problem of creating an AI companion that teeming masses of people might find indispensable—but each in a slightly different way? Suleyman divvies the necessary elements into three buckets that he’s been talking about since the days of Inflection’s Pi: IQ, AQ, and EQ.

IQ covers a companion’s raw skill at working with facts—“to answer any question accurately, superfast, be grounded in the real world, [and] provide evidence and citations,” he says. AQ references the power to take action on behalf of a user—or, using one of the tech industry’s favorite current buzzwords, to be agentic. And EQ is about the companion’s emotional intelligence—its ability to “make you feel empowered and make you feel supported and make you feel smarter and more capable.”

Building out these elements could keep Suleyman’s team busy for years. “A lot of breakthroughs need to happen to be able to get to the vision that he’s got in his mind,” says S. Somasegar, Madrona Venture Group managing director (and, previously, a 27-year Microsoft veteran). But some ingredients are falling into place, at least as first drafts. Starting in October, MAI began rolling out the meatiest changes to Copilot since Suleyman’s arrival. Thanks to voice mode, you can talk to the AI rather than type, and it’ll talk back. A feature called Think Deeper, based on OpenAI’s “reasoning” o1 model, takes 30 seconds to generate its answers, but is optimized to deliver richer, more sophisticated explanations and advice than the stock model. Copilot Vision lets you carry on spoken conversations with the AI about web pages, say, so it can help suss out pertinent details in an Airbnb listing you’re skimming while planning a vacation.

A key focus is memory: Copilot knowing you better the more you use it rather than every session being a Groundhog Day-like new start. In February, Microsoft quietly shipped an update that lets it weave topics and ideas from past sessions into new ones: “You have a sense that there’s a compounding value,” says Suleyman. Microsoft is also working on giving its AI companion enough social grace to master group chats, tailoring its responses to the interests and attitudes of each human in a session.

The challenge of making all this work is not just technological. Any AI companion worth its salt will certainly need a world-class LLM under the hood. But the MAI employees at the Hyatt off-site include “educators, therapists, linguists, comedy writers, advertisers, designers,” Suleyman tells me. Instead of building on the tech industry’s past 20 years of consumer experiences—products such as Facebook and YouTube that aggregated massive amounts of user-generated content, with its rough edges often part of the appeal—he is aiming to tap AI to help attain a level of polish that software has rarely had.

“We now have this new raw material to make beautiful experiences with, which is much closer to the raw material of Hollywood or game design,” he says. “We, as humans, love the feeling that arises when you listen to a beautiful piece of music or watch an epic movie or see a director of photography wash color over a scene.”

Once again, Suleyman’s rarified description of his aims is running ahead of anything MAI has actually shipped. But some of his aspiration to engage on an emotional level was visible in a blobby, smiling onscreen animated character I glimpsed at MAI’s off-site meeting—an early, unannounced manifestation of what Copilot would look like if you could see it as well as chat with it. The character’s cartoony vibe also happens to scratch an itch Microsoft has had since at least 1992.

That’s when the company became smitten with research by Stanford professors Clifford Nass and Byron Reeves, which showed that people attribute human qualities to computers and other forms of media. Taking their conclusions as an argument for making software interfaces more anthropomorphic, Microsoft released a quirky Windows add-on called Microsoft Bob. After that flopped, the company doubled down with Office 97’s Office Assistants, including that iconic pest Clippy.

In the 1990s, Microsoft tried to add a dash of anthropomorphic companionship to software with Microsoft Bob (seen here) and Office Assistants such as Clippy. They went on to be among the company’s most famous failures.

When it turned out Office users didn’t actually want productivity aid from cartoon characters, Microsoft deemphasized Clippy and company—and eventually removed them altogether. More recently, it has good-naturedly embraced the talking paper clip as a totem of failure. If consumers really do find an animated version of Copilot to be irresistible, Clippy can feel free to have a long-delayed last laugh.

But applying AI to a synthetic persona also brings risks that were unimaginable in Clippy’s day. Much of the air went out of Microsoft’s triumphant reveal of the first version of its consumer Copilot in February 2023 when it turned out to be a slightly terrifying loose cannon, most famously telling New York Times writer Kevin Roose that it loved him and he should leave his wife. (The company moved swiftly to tamp down its new chatbot’s wild side.) More recently, and far more alarmingly, Google-backed Character AI has been sued by the mother of a teenager whose suicide, she claims, reflected his unhealthy emotional attachment to the startup’s bots. In another suit, families say their children’s conversations with Character AI bots went in dark directions involving self-harm, violence toward others, and sexualized content.

Microsoft is hardly blithe about AI companions’ potential to go awry. ”We’re setting the standard for these things existing,” says Taylor. “And so we have to be totally sure that we’re comfortable with them existing in the world.” Her colleague Pistachio adds that the company is building its AI to calmly steer sessions in a responsible direction: “It’ll be like, ‘Okay, I think we’ve gone a bit too far—I don’t think we should be joshing around this much.‘“

Suleyman, whose book The Coming Wave takes AI’s perils at least as seriously as its promise, told me repeatedly that keeping it safe is not just vital but the whole point of his career. “The goal of the next 50 years,” he declares, “has to be to make sure that this technology remains subservient to humanity.” Yet that hasn’t led him away from high-stakes applications of the technology. Former DeepMinders are involved in a new MAI healthcare group, whose mandate he paints in only the broadest strokes for now: “Four billion or so people don’t have access to high-quality medical or health advice on a daily basis—I think it’s just an amazing opportunity.”

Ultimately, Suleyman says, his mission “is to make sure that [AI] genuinely does always remain something that makes our lives healthier and happier. The goal of civilization, in my opinion, is to relieve people of the obligation and pressure to solve shelter, food, community, work, well-being.” Even now, as he builds a business at a titan of capitalism that’s older than he is, he’s still the 25-year-old enthralled by AI’s potential to do good.

---

View the full article