What 3 new MIT Sloan professors have learned about AI — and what they want to know

This article was adapted from the May 2026 edition of the MIT Sloan School of Management’s monthly AI at Work newsletter. Sign up for AI at Work here. 

Happy September — a time for changing leaves and students heading back to school. Here at MIT Sloan School of Management, we’re welcoming new faculty members who bring with them new questions and insights about artificial intelligence. 

I asked three of them what their research or experience has revealed about the intersection between AI and work, and what else they hope to investigate in this area. Here’s what they had to say.   

What have you learned about AI and work? 

Giannis Daras, assistant professor of operations research and statistics: My research is on how AI models are trained, rather than on AI and work directly. But building these systems, and using them daily in my own work, gives me a front-row view of how fast the technology is changing our jobs. AI is automating many of the “boring” tasks that were once part of our work. For example, during my PhD, a significant portion of my time was devoted to the cumbersome process of monitoring machine learning experiments for hardware issues or training instabilities, tasks that AI agents can now handle reliably.

Beyond research, agents are streamlining everything from calendar management to technical editing. The immediate takeaway for leaders and workers is that almost all of us are still only scratching the surface of AI’s full potential.

The second takeaway is more cautionary. AI is also capable of automating parts of our work that used to be creative, and this carries real risks: losing our personal voice when we no longer write our own documents, settling for fast but sloppy executions of complex tasks, and seeing critical cognitive skills required for top-level performance atrophy. While AI is a fundamentally transformative technology, our primary challenge is to design frameworks that ensure workers grow alongside the systems they leverage.

Ruru Hoong, assistant professor of marketing: My research finds that the value of AI depends not only on the quality of its predictions but also on how those predictions are presented to decision makers. 

A large [amount of] managerial literature points to AI’s potential to improve organizational performance; my work asks why those gains have been more difficult to realize than anticipated and offers solutions for how we might design AI assistance differently. Even accurate and informative AI predictions may be underweighted, overweighted, or combined incorrectly with a worker’s own information, or simply be cognitively difficult to use. The challenge for firms is therefore not simply to introduce a better AI model but to design a decision process [that keeps] the decision maker’s cognitive biases in mind.

In an experiment with 150 professional loan specialists, my co-authors and I found that presenting an AI risk score as a simple binary (yes/no) improved decision accuracy relative to showing the full AI probability score, as well as relative to human judgment alone. Workers also made decisions faster. Importantly, arbitrary thresholds did not improve performance. The information had to be calibrated to the decisions and costs of different errors.

The broader takeaway is that AI adoption is as much a problem of organizational and information design as it is one of technology procurement. Leaders should begin with the decision they want to improve, which errors matter, what information workers possess that the AI does not, and how the AI’s output should enter the workflow. This does not make human judgment irrelevant but changes where it is most valuable — in contributing contextual information unavailable to the model, setting objectives and trade-offs, and handling exceptional cases. 

Sadegh Shirani, assistant professor of operations management: AI is making tasks that seemed nearly impossible five years ago feel as simple as a few clicks and a few words in a prompt. For example, one can now analyze millions of unstructured records at a speed and scale that were previously impractical. That is the exciting side of the story. The more concerning side is that these capabilities tempt us to delegate an increasingly broad spectrum of tasks, from routine administrative work to complex analysis and decision support. This accelerates workflows, but it also requires people to make consequential decisions much faster, often while implicitly treating AI-generated outputs as reliable. 

Yet these systems can be highly sensitive: Changing a single word in a prompt can substantially alter an AI agent’s response. The result is a new risk: AI can increase not only the volume and speed of high-quality work but also the volume and speed of low-quality work. We need to design workflows that preserve time for verification and judgment. Workers need to become skilled not only at using AI but also at questioning its outputs, testing their robustness, and recognizing when apparent confidence is not evidence of correctness.

AI Executive Academy

In person at MIT Sloan

Register Now

What interests you most about AI and its impact on work? 

Daras: To me, the most exciting question is what skills future engineers, scientists, and leaders should develop in this new era. If we accept that AI is, or soon will be, better than humans at writing code, analyzing data, or making business decisions, then the skills that matter for success at work will change. We will quickly move from asking, “How good is someone at solving X?” to asking, “How good is someone at identifying whether X is the right problem to solve?” If we enter an era of abundant solutions, the differentiating factor will be who has the best taste in problems, research or otherwise.

As this becomes clear, several exciting, and perhaps slightly scary, questions arise: What does it really mean to have good taste in problems? As educators, how do we help students develop it? As business leaders, how do we prepare our workforce to adapt quickly to this new reality? And how do we measure success? 

Hoong: What interests me most is not only that AI impacts work productivity but the architecture of work: what information humans end up contributing, how firms allocate tasks and structure organizations, and what kinds of work people ultimately value.

One set of questions around decision-making concerns private and multidimensional information. Much of how we think about AI-assisted decisions treats both human and AI information as one-dimensional signals; this reduction becomes particularly strained in the age of LLMs. When workers possess contextual information unavailable to the AI — or when the AI is uncertain along particular dimensions — how should they be directed to acquire and communicate that information? If people must still aggregate multiple signals, can interfaces direct their attention towards the types of information they systematically underweight?

I am also interested in what happens within firms as AI changes the bundles of tasks that make up a job. Managers play a pivotal role in redesigning roles and moving workers internally but may face weak incentives to release talented employees, limited authority to restructure jobs, or incomplete information about how AI is changing tasks and where internal opportunities exist. I want to understand which organizational structures help firms overcome these frictions — and whether they predict which firms genuinely redesign work and redeploy workers rather than simply hiring or laying them off. 

A third set of questions concerns consequences for workers that conventional measures of employment, wages, and productivity miss. I am interested in how exposure to AI — and career events such as promotions, job changes, and layoffs — affects objective measures of sleep, activity, and well-being, and which workers are most affected.

Finally, I am interested in how technological change alters what people value. When machines can reproduce more kinds of work, which human contributions become more valuable precisely because they remain distinctively human? Conversely, what forms of AI-produced output will people value because of (rather than despite) their technological origins?

Shirani: I am most interested in the problem of alignment: How can we design an AI agent to accomplish what we want, adhere to our expectations, and behave reliably when those expectations are incomplete or difficult to specify? This becomes particularly important as we delegate more complex and consequential tasks to AI systems. 

I am especially interested in whether adversarial AI agents can help other AI agents behave better. An adversarial agent could challenge a proposed decision, search for failures, identify hidden assumptions, or deliberately test whether another agent’s behavior remains consistent under small changes in instructions or context. Rather than relying on a single agent’s answer, we can create a system of interacting agents that critique, monitor, and stress-test one another. 

The broader research question is whether these adversarial interactions can make AI systems more reliable, or whether they introduce new forms of manipulation, coordination failure, and misalignment. Understanding when such systems improve behavior, and how they should be designed and governed, is an important step toward using AI responsibly in decision-making.

Giannis Daras is an assistant professor at the MIT Sloan School of Management in the operations research and statistics group. He works on practical and theoretical questions around deep generative models, with a focus on training and sampling generative models in the presence of data corruption. 

Ruru Hoong is an assistant professor in marketing at MIT Sloan. Her research examines how organizations should design and integrate new digital technologies, particularly artificial intelligence, into decision-making. She asks not only whether AI improves decisions, but who benefits, on which tasks, and under what organizational conditions. 

Sadegh Shirani is an assistant professor in operations management at MIT Sloan. His research focuses on developing principled methods for modeling, learning, and decision-making in complex systems. His work spans causal reasoning from complex data, LLM-based simulation of human and human-AI interactions, AI safety in interactive and multi-agent systems, and reliable AI for decision-making.