5 things to consider when working with AI

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. 

For a while, the conversation about generative AI — and then AI agents — was largely speculative. What might it do to our day-to-day work, our workplaces, our careers, and the economy? Artificial intelligence was like a new coworker, and we were anticipating its arrival.

Now our AI coworkers have arrived in most workplaces and sectors. We have some understanding of what AI can do and where it falls short (and we’re looking for more). As Kenneth Munie, a senior managing director at Accenture, said at the MIT Initiative on the Digital Economy’s annual conference in April, humans are becoming “team leader[s] of agents and agentic technology.” 

What do we need to know about these virtual coworkers and how to get the most out of this collaboration? And how can we be effective leaders of teams of agents — or, at the very least, guide how AI is used in our work and our organizations? 

These questions are at the heart of work by the researchers at MIT IDE. The initiative helps companies adapt to new ways of doing business in the digital economy and during times of change. Its research groups are dedicated to looking at areas such as applied AI, AI in financial markets and decision-making, human-first AI, and AI and labor economics. 

At the conference, IDE researchers shared insights from their latest research and actionable advice on these topics. Here are some insights that I think could help us become better leaders in this era of AI.

Consider how your AI is designed

Research by MIT Sloan professor and MIT IDE director Sinan Aral shows the importance of AI personality pairing — teaming humans with AI agents that complement them. For example, extraverted humans paired with conscientious AI performed much worse on a task than average, while extraverted humans paired with an extraverted AI had better outcomes. For humans, gender and country of origin also influenced which AI pairings led to the best outcomes. “There is a huge opportunity in personalizing AI for each of us,” Aral said. 

And whether humans and AI share “mental models” — internal representations that people use to understand the world and make decisions — can affect how they work together. Preliminary research findings by Zezhen (Dawn) He, an MIT IDE postdoctoral associate, show that while people favor AI systems that align with their mental model, they are more likely to adopt recommendations and make better decisions when supported by an AI model that does not align with their views. When selecting models, don’t just look for the highest accuracy, He said. Design systems for the cognitive fit between humans and AI.

Consider whether AI is giving objective advice

When companies use large language models to make business decisions, the language used to prompt the AI can have a big impact on how models respond. When LLMs are prompted with language about maximizing profits, for example, they tend to dismiss risks and fail to recommend escalating information to a company’s board. Businesses tend to imbue models with motives that go under the radar but have “a big influence on what you see as a decision maker in your firm,” MIT Sloan professor Eric So said. 

Consider redesigning workflows

People tend to focus on which tasks in an occupation can be automated. But it’s better to look at how AI affects entire chains of tasks, according to Peyman Shahidi, a PhD candidate at MIT Sloan. Shahidi presented research showing that how work tasks are clustered matters as much as which tasks can be automated. It’s better to reorganize workflows by grouping tasks that can be done by AI together, even if AI might be marginally worse than humans at some of those tasks. Removing repeated oversight can outweigh the marginal performance gains earned by having humans take over tasks they are a little bit better at. And seeing ROI from this reorganization can take time: The payoff is nonlinear and arrives when you reach a reorganization tipping point. 

AI Executive Academy

In person at MIT Sloan

Register Now

Consider taking time to think before acting on AI recommendations

People tend to praise being frictionless in business — such as making AI easy to use and ensuring that AI outputs can be seamlessly incorporated into work. But friction shouldn’t be considered a pain point. In fact, making AI use too easy can erode critical thinking and introduce risks, according to MIT Sloan research scientist and senior lecturer Renée Richardson Gosline. For her ongoing research in this area, Gosline and her team worked with the Commonwealth Bank of Australia to see what happened when humans were asked to articulate why they were following AI agent recommendations before acting on those recommendations. The research team found that doing this reduced uncritical reliance on AI suggestions and also led to greater accuracy: Even just asking people whether they were sure about the results reduced their reliance on agents. And the time spent on tasks didn’t change significantly, indicating that these checkpoints can be added without reducing productivity. 

Consider what jobs LLMs will automate and what that means for workers

LLMs are disproportionally better at automating shorter tasks than longer tasks, and lower-income workers tend to do shorter tasks, according to research from MIT principal research scientist Neil Thompson. That indicates that lower-wage employment is more likely to see task loss. But Thompson also noted that task loss doesn’t necessarily mean that wages will go down. If automation removes the simpler parts of a job, the tasks that remain often demand more expertise, which can make that work better paid because fewer people are able to do it. But when specialized tasks are automated instead, the job may become easier for non-specialists to do, increasing competition and lowering wages.

To explore these and other insights about data privacy, AI search, and AI risks, take a look at this report from the MIT IDE conference.

Sinan Aral is a global authority on business analytics and is the David Austin Professor of Management, Marketing, IT and Data Science at the MIT Sloan School of Management; director of the MIT Initiative on the Digital Economy; and a founding partner at the venture capital firms Manifest Capital and Milemark Capital. He leads the Applied AI research group at MIT IDE, and his research focuses on applied AI, social media, and disinformation. 

Eric So is the Sloan Distinguished Professor of Global Economics and Behavioral Science at MIT Sloan, faculty co-director of the AI Executive Academy, faculty chair of MIT Sloan's PhD program, and lead faculty for the MIT Sloan Generative AI for Teaching and Learning hub. He leads the AI in Financial Markets and Decision-Making group at MIT IDE. His current research portfolio spans interconnected topics, including artificial intelligence, behavioral economics, human-computer interactions, and regulatory policy. His book “The Collision: What AI Does to Us” will be published in October 2026.

Renée Richardson Gosline is a research scientist and senior lecturer at MIT Sloan and head of the Human-First AI group at MIT IDE. She is an expert on the intersection between behavioral science and technology and on the implications of AI for cognitive bias in human decision-making, and a leading thinker on how AI affects human judgment and the interplay of human and AI bias.

Neil Thompson is a principal research scientist at the MIT Computer Science and Artificial Intelligence Laboratory and MIT IDE, where he is head of the Artificial Intelligence, Quantum and Beyond research group. He is also the director of the MIT FutureTech lab, which studies the economic and technical foundations of progress in computing. He studies technological innovation and firm strategy. 

Zezhen (Dawn) He is a postdoctoral associate at MIT IDE. Her research investigates how people perceive, evaluate, and interact with AI — combining behavioral experiments and methodological advances to improve human–AI collaborative outcomes.

Peyman Shahidi is a PhD candidate at MIT Sloan. He studies market design and labor economics, with a focus on the effects of AI on labor markets and online platforms.