Author Talks: How AI-powered teams could transform the future of work

In this edition of Author Talks, McKinsey Global Publishing’s Mike Borruso chats with Melissa Valentine, associate professor of management science and engineering, about Flash Teams: Leading the Future of AI-Enhanced, On-Demand Work (MIT Press, October 2025), coauthored with Michael S. Bernstein. Valentine shares how flash teams can optimize team assembly and workflow to provide a new managerial superpower for business leaders. This approach transforms traditional team structures, enables more data-driven decision-making, and drives better project outcomes. An edited version of the conversation follows.

What problem were you trying to solve in writing the book?

I was at a social event, and I met someone who mentioned spending several hours per week trying to integrate her point-of-sale software, custom website, and inventory management.

I said, “You should not be spending your time on that. You should get a flash team.” I showed her how to use a flash team for the first time. Within a couple of days, she had solved the problem and asked, “Why have you never told me about this?”

Similarly, I attended a retreat for senior business leaders at which I told a smart, successful, and innovative CEO about flash teams. She said, “That’s super cool. I would have no idea how to do that.” I was struck by her response.

Based on those two stories, my coauthor Michael Bernstein [professor of computer science at Stanford University] and I realized that people didn’t know what flash teams are or how to create them. Given our position at Stanford Engineering, we are aware of tools and capabilities. We felt that people need to know what’s possible with teams, and with flash teams, specifically, in a practical way.

How are flash teams different from agile teams?

Any flash team with experts who are very good at agile methodology would be very successful. Agile is such a useful methodology. The flash team’s approach is very complementary to agile. It’s good at being iterative and adaptive, which is essential to flash teams.

Where flash teams build on agile is that agile is more general. Flash teams incorporate new tools that agile hasn’t necessarily applied. There are millions of people in online labor markets who are registered, expert, and can join the team. There are all types of AI tools or other data tools that can help inform the assembly and management of the team and augment what’s possible with agile.

There might be an opportunity to learn where AI can be helpful. Agile is a great framework. The agile framework was developed before AI tools became so much more accessible, so rethinking where AI can help in places it wasn’t noticed before could prove beneficial.

What’s AI adding to agile that humans can’t?

New technology makes the search for teams much more automated and expansive. For example, if there’s a large labor market that has ten million people registered and they have lots of skill data, credentialing data, and data about when people are available, then you can find people and more diverse expertise more quickly. That is one piece of the team assembly.

When you’re working on multiple software platforms, you’re collecting a lot of data on what’s happening. You can draw on that data to inform your decisions. For example, one of the book’s chapters is about optimizing team assembly. We write about that as a new kind of managerial superpower. You might have a point of view about who will work well together. You might want to optimize for familiarity or for different skills. In the past, it was unlikely that people were optimizing team assembly by using data in that way.

Also, when you’re managing the team, there’s a lot more that’s possible when using AI. We write about how you can receive recommendations on the team, role structure, and workflow. AI can be part of your team’s learning journey as well. For example, you can have the team reach a milestone and then have AI analyze how things went. It can recommend whether to keep the same team structure or try a new one.

There is a new, fun function that enables you to conduct AI simulations on your team. For example, you could say, “Here are the five roles I have. What’s the likely bottleneck? What’s the likely conflict that will come up?”

You could even say, “Here are the five individuals. There is a new kind of AI technology people are inventing right now. What’s going to happen with this set of individuals?”

Where does the data come from to enable flash team analyses?

Software teams and remote teams have become very expert at documentation. If you’re already familiar with that world, you’re already aware of what it would take to form a flash team. There are two kinds of data that we focus on: documentation and trace data. Documentation is the record of decisions and of everything that has been done. Trace data is more passive and slightly easier to automate. For example, with one of our early proofs of concept, we had a data set that reflected every task that had been completed and the time it took. That could become a template of the workflow, potentially used to recommend the next workflow. Over time, that archival trace data becomes the flash team’s library that helps recommend the future work.

And the possibilities expand with emerging smart technologies?

One of the companies that is creating flash teams at an enterprise scale shared a prototype of a client pitch, where the clients describe what they want. Then, using natural language, users can prompt the “smart conference room” about what they’re seeking. On the wall, profiles of people who are available start to appear.

There are also instant flash teams informed by voice, along with an automatic display of these profiles. People are envisioning a lot of cool stuff that starts to get into the science fiction realm.

Could large organizations build their own internal labor market platforms?

Since we began conducting this research, we’ve presented to more executive audiences. When the C-suite realizes, “We have 400,000 employees; we could do this,” we always respond, “Yes, exactly.” What is the data that exists internally that tells you who “everybody” is? What are their skills, their availability? How do you start to use that more dynamically? That seems to be a future capability that’s very generalizable for any business.

From the worker side, are flash teams just for people who like freelancing?

In management research, we discuss extreme cases you can learn from. For example, leadership in crisis tells you about daily leadership to some degree.

The generalizable principle that illustrates why flash teams can be good for both workers and organizations is that companies have a more strategic and data-driven point of view on staffing and on making use of employee expertise.

One example relates to an organization, a retail tech company that had a very advanced data science team. They have 100 data scientists, one of whom—Molly—happened to be an expert on causal inference.

For each data science algorithm they started, they had to consult with Molly. Every data science team ended up making use of Molly. She became a flash expert who joins, consults, and rolls off teams.

For someone like Molly who has a very rare, valuable expertise, what tools could be used to help manage her time and engagement more dynamically and more flexibly across all these teams? Right now, Molly manages her time using her calendaring system. Yet you can start to imagine a more strategic, holistic, data-driven approach to understanding expertise and modeling where people fit in. That means approaching the process from the point of view of people who are in large enterprise systems.

In our situation, we worked with freelancers more. Oftentimes, freelancers complete a lot of solo work. For them, flash teams present a chance to work on something of higher value. Actually, a lot of the people we interviewed after our engagements were very much into it. Working on a team gives them bigger projects. They’re working with people and solving problems together. It’s more fun. The vision offers something to both large enterprises and freelance settings.

Once you start meeting people who are good at freelance, especially in online labor markets, you can understand what a gift this type of work is for them.

They can be quite flexible. They can work from anywhere, they choose their clients, and it can be profitable for some people. Some companies are establishing enterprise flash teams now. One company we worked with completed business models for them. Given the flexibility, they talked a lot about how they were able to keep people in the workforce who perhaps otherwise wouldn’t have access to the workforce.

Will flash teams work in any industry?

We see flash teams most in software, creative industries like film, and professional-service firms—any industry that’s very project-based. Flash teams are obviously very well suited for that. Yet I have had great conversations with people who are, for example, in healthcare and are seeing the opportunity to rethink how they collaborate.

For example, there are even teams of clinicians that help with a diagnosis. There’s an academic cancer center that created what I view as an internal flash-team system. They realized that, for cancer patients, having very diverse experts speak to the plan and the diagnosis was better. There are even auxiliary services, such as dermatology, nutrition, and palliative care. If all those people can come together and speak to a patient’s case, especially to very complex patient cases, then that will be a better plan for the patient. Actually, I saw that academic cancer center build a system that’s essentially a flash team: cross-functional experts to speak to complex patient cases. Once you have the vision, and once you start to see how useful teams of experts are, and once you have technology to help you get people together dynamically, then you start to see it everywhere. You start to understand what’s possible across industry.

Does the future of AI involve technology suggesting better ways to work?

Our models of org charts, teams, and more have necessarily been from a more paper-based world. Every org chart from whatever year was available in a PDF. Yet there’s so much data on who’s working with whom, who’s talking with whom, and the vision involves having systems and models that are informed by that.

With flash teams, we began with the idea of temporary teams. Yet as we worked in this space, we saw that so much data is being produced on team interactions and team workflows, not just at the team level but at the organizational level. Everything that is spoken, every email sent, becomes a model of the team or a model of the organization.

The vision is that you can manage teams better and work together more effectively if you’re drawing on that kind of digital model of the team or organization.

How do people feel about this type of management?

In writing this book, I conducted a lot of research on algorithmic management. Once people are working online and using data to inform choices about who should be talking to whom and what the team structure should be, that gets into the world of algorithmic management.

There’s a lot of research on what people do and do not like about algorithmic management, and the results are unsurprising. People don’t like being exploited. They don’t like being surveyed. They like having autonomy and dignity and being empowered.

There’s one chapter in the book where we write about algorithmic ratings. Those are used a lot in flash teams because of the temporary nature of gig work and flash teams. Oftentimes, companies use ratings, such as your Uber driver five-star scale. Trying to do algorithmic ratings for knowledge workers or professional workers is complex. We have two scenarios: one where it really works, where a large network of perhaps 1,000 software engineers really loved the algorithmic rating system. They called it “Karma.” They loved giving and receiving Karma. Then, when headquarters discontinued Karma, there was a revolt. In another scenario, we studied a network of freelancers who thought that the algorithmic rating system was pretty opaque. Then it changed, and people’s ratings went down. They didn’t understand, and there was no recourse. It was not a good use of algorithms or these tools.

Culture matters. Get buy-in; be good to people. This is a tech-forward example of those basic leadership principles.

What are the biggest barriers to scaling this model?

The biggest barrier is [a lack of] awareness of the tools and the capabilities that are possible and the sense of how to apply them. A related problem is that generative AI has rightly captured a lot of public and business attention right now. For the last year or two, people have been very focused on configuring individual gen AI use cases.

Now I am observing the current conversation shift. People are thinking more about how generative AI affects teams and organizations overall. I feel lucky because our book is catching the wave at the right time. People are looking up from individual gen AI use and asking what they can do in terms of tools and organizations. It’s the right moment to publish this book. We have a point of view on different things people can do to augment their teams and their organizations as a whole.

Where do flash teams fit in the future of work?

Flash teams are one alternative to the future of work. I’m hoping this book speaks to team science and organization science.

Every day, managers make decisions about how to structure their teams or how to structure their organizations. If we’re honest, that’s an art in the sense that it’s based on intuition. We don’t necessarily know if our theories are correct. Is it actually true that a tall hierarchy is more efficient, decentralized, and innovative? Those are layperson theories that we can’t measure in a rigorous way at this point. But teams are so important. Organizations are so important and so high value.

My hope is that we were able to mature our understanding of team science and organization science. What if we had better models for the impact of the different leadership and organizational-design decisions we made? That’s the future of work I think this book can address.

There’s a conference called “Collective Intelligence,” which looks at any group that behaves intelligently. So you’ll have a faculty or any collective that studies ants and looks at all the coordination behavior of ants. Organizations are yet another example of that. We have assumptions about specialization and hierarchy that motivate the way we make decisions. What if AI and all this data could make us better informed and really improve our decision-making? That is the more important future of work that I hope happens. Whether or not teams are temporary, I would hope that people would match the team to the problem in an informed way. Team speed isn’t necessarily the most important thing. It’s strategy. It’s how we are making use of these powerful tools that are here now.

It’s very reminiscent of pre-AI network mapping.

It’s reminiscent of the early network data, where you could start to think about where the actual heat maps of influence within an organization are. That idea is exactly what’s possible here.

The open question I have now is how agents get involved in all of this. We spent a year thinking about large language models [LLM] and how agentic capabilities are at the forefront of everyone’s minds. The active research we’re currently conducting now, following the book, is about how agents fit into the whole vision.

I don’t have an answer, but that’s something I hope to keep talking to people about. When we’ve spoken with people about flash teams in the last three months, it has been a funny experience. They always assume we mean agents, though we do not mean agents.

That just speaks to the zeitgeist, which has really changed. My coauthor is actively in the agent space. He and we are not ready for teams of agents.

What should an organization that wants to try flash teams do first?

There is a fun exercise I like to do with executive audiences when I’m giving presentations on flash teams. I’ll ask the audience, “If there was a problem that you knew for sure was going to get solved in the next six weeks, what would it be?” Everyone has an answer. Everyone submits their problem, and we look at it together. Then I help them think about the idea of experts everywhere, all the time. I help them consider what the right expert is.

If you pretend that you could get access to any expertise anywhere in the world, who’s the right expert to help you? We’ll consider whether the expert is someone in an online labor market, someone in-house, or a friend. Let’s think about how you rightsize your request.

Is the ask getting Molly to come consult at your data science meeting for an hour? One topic that arises often is that people want an agent. My response is, “What are you doing for the data work?” We start to realize that there needs to be an internal flash team to “flash” the data set we need and clean up the data. That is not a fun flash team. Yet that has become the answer that arises lately when we have these conversations. It really helps people think more about expertise in a dynamic way. It helps them get used to identifying problems, matching expertise to those problems, and then structuring the expertise in a way that actually solves the problem.

How would you summarize the premise of flash teams in one sentence?

I would say, “You, dear reader, can use AI to find and manage teams more effectively and more dynamically.”

Did anything surprise you in your work on this book?

It has been surprising and mostly fun to have put this set of ideas out into the world as generative AI is getting a lot more accessible and a lot more familiar. Gen AI makes all this so much easier and accessible. We have one chapter that addresses how AI can lead you to experiment with a team structure.

It is based on a paper we wrote that covered a brilliant computer science student who built a multiarmed bandit. During that process, I learned that it is a smart method of experimentation. My coauthor realized that you can now prompt using natural language—one of the large language models—to apply that same technique for you. You don’t have to build the custom software, you don’t have to do the math, and you still achieve that functionality. We have been astonished, in a fun way, mostly, about how much easier all these techniques have gotten now that people are using off-the-shelf LLMs.

Visit Author Talks to see the full series.

Learn more