Building AI That Understands People

For Michelle Zhou (PhD ‘99), a career spanning research, entrepreneurship, and enterprise consulting has centered on a persistent question: How can AI better understand people and help them achieve their goals?

Two recent honors recognize contributions along that path. Zhou was named to Forbes’ 2026 50 Over 50 list and, with her coauthors, received the 2025 IEEE Computer Graphics and Applications Test-of-Time Award for “Context-Preserving, Dynamic Word Cloud Visualization.” Published in 2010, the paper introduced techniques for following how topics evolve across collections of text while preserving the context and relationships needed to understand those changes.

That same interest in designing technology around people has carried through Zhou’s career. After earning her PhD in Computer Science from Columbia, where she worked with Professor Steven Feiner, she spent more than 15 years at IBM Research and the Watson Group, later co-founded the AI company Juji, and today serves as Managing Director of Accenture’s Center for Advanced AI. Across those roles, her work has evolved from intelligent interfaces and conversational systems to human-centered and agentic AI.

Zhou is also the author of the forthcoming ACM book Human-Centered Agentic AI, which explores the opportunities and risks of human-AI teaming.

In this Q&A, Zhou discusses what she learned from building AI that understands people, why she moved from research to entrepreneurship, the importance of making AI broadly accessible, and what successful collaboration between humans and AI could look like.

Q: You spent more than 15 years at IBM Research and the Watson Group and helped invent IBM Watson Personality Insights. How did that work shape the way you think about AI’s ability to understand and interact with people?

A big part of my work in creating intelligent systems/AI agents has always been about using AI to better understand people so that AI can better help them. And that understanding goes far beyond modeling a user’s tasks, preferences, or interaction style. It means developing a deeper understanding of who they are: what motivates them, what they are passionate about, where their strengths and talents lie, and how they handle life’s challenges.

That perspective was shaped significantly by my work on Watson Personality Insights at IBM. It was one of the early commercial products that enabled organizations to analyze people’s communication footprints and infer their psychological characteristics at scale. While the technology was relatively rudimentary compared with today’s AI, the underlying idea was powerful. Scientific evidence demonstrated that this kind of understanding could influence people’s behavior, for example, from what they chose to read to what they chose to buy.

That experience reinforced a belief that has stayed with me: AI becomes much more useful when it deeply understands its users. It inspired much of my later work on truly human-centered AI agents that can dynamically analyze user behavior, infer characteristics and needs, and use that understanding to determine how best to help. That could mean something as simple as adapting how an agent communicates with someone, or something more meaningful, such as helping persuade someone to stay on track with their treatment journey.

In many ways, that is the thread connecting my work at IBM to what I’m doing today: building AI that doesn’t just understand the task, but reads the human behind the task.

Q: What motivated you to move from research at IBM to entrepreneurship and co-found Juji? At Juji, you were developing human-centered AI agents well before today’s surge of interest in generative and agentic AI. What did you see early on about the potential for AI agents?

After more than 15 years of doing advanced research at IBM, I wanted to take the next step: to turn my passion for creating intelligent systems into something that could have more direct, real-world impact. Research gave me the opportunity to explore fascinating possibilities, but I wanted to see those ideas translated into solutions that could directly benefit people. That desire motivated me to become an entrepreneur and co-found Juji.

The decision to focus Juji on human-centered AI agents was a natural continuation of what I had been passionate about throughout my career: creating intelligent systems that can better understand people and use that understanding to help them.

What I saw early on was that AI agents could do something particularly powerful: democratize high-touch, personalized services that have traditionally been available only to people who can afford them. Most people cannot have their own career coach, financial advisor, or wellbeing counselor. If we could create AI agents that deeply understand each individual and interact with them in a personalized way, we could make these kinds of services much more accessible.

That became the vision behind Juji. We were not simply trying to automate tasks. We wanted to create AI agents that can democratize high-touch services, directly benefiting billions in the world.

So, in many ways, Juji was not a departure from my research career. It was a natural next chapter: taking a passion I had pursued throughout my career and turning it into human-centered AI agents that could reach and benefit many more people.

Q: A recurring theme in your work has been making AI more accessible, including enabling non-technical subject-matter experts to build AI agents. Why has democratizing AI been important to you?

When a new technology is born, it naturally creates a technology divide. We saw this with the digital divide: the people and organizations with the resources, expertise, and access to adopt new technologies benefit first, while others can be left behind. And generally, the more effort and resources it takes to advance and perfect a technology, the wider that divide can become.

AI is a particularly resource-intensive technology to advance and perfect, requiring significant expertise, computing power, data, and investment. As AI advances rapidly, we face a similar challenge: an AI divide between those who have the expertise and resources to build and benefit from AI and those who do not. That is why democratizing AI has become so important to me. I want to help bridge that divide by making AI more accessible, so that people do not need to be AI experts to create and benefit from AI.

This is also deeply connected to why we founded Juji in the first place. Our vision was to democratize high-touch services, such as personalized career coaching, financial advising, and well-being counseling, so that these services could become accessible to people who otherwise could not afford them.

If the AI powering those services is itself accessible only to people with significant technical expertise or resources, there is a fundamental contradiction. That is why, for me, democratizing AI is not just about making technology easier to use. It is about making the benefits of AI more broadly accessible, hopefully to everyone.

Q: Now at Accenture, you’re helping organizations adopt AI and agentic AI at a much larger scale. What separates an AI project that is technically impressive from one that actually creates meaningful value for an organization?

The keyword is human.

Regardless of the type of AI project, humans remain at the center. AI agents are created and configured by humans, humans interact and collaborate with them, and human stakeholders ultimately evaluate AI behavior, performance, and outcomes delivered.

So, a technically impressive AI project does not necessarily create meaningful value. The real question is whether and how it augments human capabilities to help an organization grow and sustain that growth.

This is especially important with agentic AI, where agents increasingly become part of how work gets done. For me, successful AI is not simply about what the technology can do, but how effectively humans and AI work together to achieve meaningful outcomes.

Q: You’ve described a future centered on human-AI teaming and human-machine symbiosis. What does successful collaboration between a person and an AI agent look like to you?

To me, successful human-AI collaboration is fundamentally about complementarity. Humans should do what humans do best, and machines should do what machines do best.

This idea goes back to J. R. Licklider’s vision of man-machine symbiosis. Machines can process vast amounts of information, recognize patterns, operate continuously, and handle complex computations at a scale and speed that humans cannot. Humans, on the other hand, bring judgment, creativity, values, and the ability to understand broader context. The goal is not for one to replace the other, but for each to amplify the strengths of the other.

My upcoming book, Human-Centered Agentic AI, which will be published by ACM in October 2026, explores human-AI teaming in depth, including both its benefits and its risks. Ultimately, I believe successful human-AI teaming is about creating a partnership in which AI expands what people can accomplish while humans remain meaningfully in control.

Q: Looking back at your time at Columbia, what experiences or lessons have stayed with you, and what advice would you give today’s computer science students who want to build a career in AI?

Three things have stayed with me and continue to guide me today: passion, purpose, and perseverance.

Passion means finding something you genuinely love and are good at, and having the courage to pursue it. When I decided to pursue my Ph.D., I was determined to work on a problem that truly fascinated me. My advice to students is: find what excites you, AI or not AI, and don’t compromise on that.

Purpose is about connecting that passion to something bigger than yourself. My passion has always been about creating AI systems that can better understand and help people. When you connect what you love to a larger purpose, you not only create greater impact, but you also inspire others to join you.

And finally, perseverance. AI is an exciting field, but it is also full of difficult problems, unexpected setbacks, and ideas that may take years to realize. There will inevitably be moments when things don’t work. Don’t give up too easily. Stay curious, keep learning, and keep moving forward.

If I could leave today’s computer science students with just three words, they would be: find your passion, give it purpose, and persevere.