10 levers for shaping generative AI that truly improves worker performance

Historically, workers have executed processes manually. Now, they are increasingly finding themselves in a new role: human in the loop, overseeing and analyzing the work that generative AI carries out. 

But whether that shift makes jobs better — not just faster — depends on the choices employers make.

A recent report, “Humans in the Loop: The Evolution of Work in Early Experiments With Generative AI,” presents summary findings from the MIT Working Group on Generative AI & the Work of the Future, co-led by Ben Armstrong, executive director of the MIT Industrial Performance Center, MIT Sloan School of Management professor Kate Kellogg, and MIT professor Julie Shah.

Between 2023 and 2025, the authors and contributors to the research interviewed executives rolling out generative AI, the managers overseeing it and the employees using it at more than 20 companies. They assessed firms at every stage of AI adoption — from early experimentation to measuring what worked to scaling successful applications. Then they cross-checked those findings against a large-scale 2023 survey of worker attitudes toward artificial intelligence and automation, plus additional public survey data on AI usage.

The analysis identified three ways generative AI can fail to improve workers’ performance or the quality of their work:

• Disuse — not automating where AI adds value
• Misuse — automation that delivers poor results
• Overuse — automation that works but creates new problems

To help firms avoid those pitfalls, the authors set forth 10 practical levers — three guiding principles and seven outcomes — for deploying generative AI in ways that make jobs better, not just faster.

Three guiding principles

The first three levers are operating habits the researchers found running underneath every successful case they studied.

1. Gather evidence before scaling

The researchers found that the most effective AI projects started with a business problem, set clear measures of success, gathered evidence, and scaled only once they had evidence that generative AI outperformed the alternative.

Many less-successful companies deployed AI because it appeared to shave time off a task, without asking what it meant for quality, whether it created rework, what new tasks it generated, or how it changed the team’s overall value added.

2. One size does not fit all

Don’t expect everyone to use artificial intelligence the same way. The researchers found that workers — even those in the same role — often apply AI very differently. AI’s flexibility can make its impact tougher to measure, but the researchers see it as genuinely improving job quality, because employees can use IT tools how and when they want and decline to use such tools when they don’t trust the results. What’s more, a workforce of employees who use AI differently can generate richer evidence about what works, for whom, and in which situations.

3. Learn when to trust

Using AI well depends on knowing when to trust its output and when to question it — not trusting it more or less overall, but learning to identify which outputs deserve confidence and which don’t. That’s hard because generative AI remains a black box, for the most part, and there’s disagreement over whether today’s tools can ever be made fully transparent without additional engineering. So if the technology itself stays opaque, employers should bear responsibility for building practices and systems that help workers calibrate their trust of AI, the researchers argue.

Seven AI outcomes

The remaining seven levers are outcomes the researchers see as anchoring the most promising AI applications.

4. Minimize drudgery

Workers want AI to take over the boring stuff, not the work that makes their jobs interesting. The researchers found that AI is most effective when it reduces routine work, freeing people to focus on problem-solving and creativity. Workers closest to routine tasks are often best positioned to identify them; employers can then decide which technologies should eliminate such work.

5. Promote learning

The researchers warn against “mental offloading”: when workers rely on AI to complete a task without retaining the knowledge behind it. That’s a problem because learning at work benefits both sides: It can unlock better career opportunities for workers and higher productivity for employers. AI can help employees spot gaps in their own knowledge and help close them when companies implement guardrails that nudge workers toward learning and away from simple shortcuts.

Leading the AI-Driven Organization

In person at MIT Sloan

Register Now

6. Preserve teamwork

Don’t assume that AI should replace collaboration. The researchers found that artificial intelligence can help employees complete specific tasks on their own that previously required them to pull in colleagues with different expertise. But they also warn that that self-sufficiency may come at the expense of mentoring, collective learning, and building trust between colleagues. Those trade-offs are worth weighing before letting AI turn a task into solo work.

7. Design better interfaces

Companies increasingly buy the underlying AI technology rather than build it, the researchers determined. Firms can still shape how employees experience AI, however, by designing or customizing the interface layer workers use to access it. Good design can help employees build situational awareness — an understanding of what’s happening and why — while keeping their mental workloads manageable. The researchers also recommend testing different interface designs to measure their impact on information retention and mental workloads.

8. Continue to invest in domain expertise

Fields where AI has high potential — such as medicine or computer science — may see reduced demand for entry-level roles in the short term. But that could backfire over the long term: Breakthroughs will still require experienced people to interpret and test what AI produces — and to navigate the regulatory and bureaucratic processes that require human judgment and trust. For that reason, training institutions should invest more in developing domain experts in areas with high potential for generative AI to make an impact, the researchers write.

9. Maintain accountability

The researchers warn that AI can produce work that looks convincing but masks errors or gaps in understanding. Because many organizations operate on trust between the people who analyze information and those who act on it, those mistakes may never be recognized or traced back to the person who produced them. The solution is to make people accountable for AI’s output — which builds an incentive to actually learn the material and raises the cost of making an error. They point to airport security, where random additional screenings keep people attentive and confirm that the automated system’s output is accurate and complete.

10. Create new work

Focus not only on the jobs AI removes but also on the ones it creates.

Companies that focus on using AI only to reduce or eliminate work are missing out, the researchers write. Novel technology often gives rise to entirely new roles. When AI frees up workers’ time, it creates an opportunity to redesign jobs around both business needs and the skills that employees want to develop. Workers who feel that their employer is investing in their growth are more likely to use technology effectively — and to see it as a positive force in their careers.

“Humans in the Loop: The Evolution of Work in Early Experiments With Generative AI” was co-authored by Ben Armstrong and Julie Shah, with contributions from MIT Sloan School of Management professor Kate Kellogg, along with Sabiyyah Ali, Greer Brigham, Carey Goldberg, Shakked Noy, Prerna Ravi, Azfar Sulaiman, Leana Tejedor, Felix Wang, and Whitney Zhang.

Ben Armstrong is the executive director of and a research scientist at the MIT Industrial Performance Center, where he co-leads the Work of the Future initiative. His research and teaching examine how workers, firms, and regions adapt to technological change. His current projects include a working group on generative AI and its impact on jobs, as well as a book on American manufacturing competitiveness.

Kate Kellogg is the David J. McGrath Jr. Professor of Management and Innovation at the MIT Sloan School of Management. Her research focuses on helping organizations and knowledge workers develop and implement AI systems to improve decision-making, collaboration, and learning.

Julie Shah is the H.N. Slater Professor in Aeronautics and Astronautics in the MIT Department of Aeronautics and Astronautics and director of the Interactive Robotics Group. She studies autonomous systems, human-robot collaboration, AI planning and scheduling, and interactive robotics for aerospace, medical, and manufacturing.