MIT Research Identifies 10 Levers for Deploying Generative AI That Improves Worker Performance

A new report from the MIT Working Group on Generative AI & the Work of the Future outlines a practical framework for organizations deploying generative AI — one focused on improving job quality, not just speed.

Led by Ben Armstrong, Kate Kellogg, and Julie Shah, the research team interviewed executives, managers, and employees at more than 20 companies between 2023 and 2025. They also analyzed large-scale worker attitude surveys on AI and automation.

The report identifies three failure modes companies encounter with AI:

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

To avoid these pitfalls, the researchers propose 10 levers in two categories:

Three guiding principles:

• Gather evidence before scaling: start with a business problem, measure outcomes, and scale only when AI demonstrably outperforms alternatives
• One size does not fit all: workers in the same role often apply AI differently, and that variation generates richer evidence about what works
• Learn when to trust: employers must build systems that help workers calibrate when AI output is reliable

Seven outcomes to anchor AI deployments:

• Minimize drudgery — take over routine work, not meaningful work
• Promote learning — guard against mental offloading where workers rely on AI without retaining knowledge
• Preserve teamwork — AI self-sufficiency can erode mentoring and collaboration
• Design better interfaces — good UX builds situational awareness and manages mental workload
• Continue investing in domain expertise — high-AI-potential fields still need experienced interpreters
• Maintain accountability — make people responsible for AI outputs
• Create new work — identify the roles AI creates, not just the roles it eliminates

The report emphasizes that fields with high AI potential, such as medicine and computer science, may need more domain experts, not fewer, to interpret and validate AI output over the long term.

Read the full article on MIT Sloan Management Review.