AI Workplace Capability Rising Gradually, Not in Sudden Leaps

Artificial intelligence is becoming more capable at workplace tasks through steady, incremental improvement rather than sudden breakthrough moments, according to new research from MIT FutureTech that challenges conventional assumptions about AI-driven disruption.

The study, which analyzed more than 60,000 worker evaluations of AI performance on over 6,000 text-based workplace tasks, found that AI capability is advancing broadly across different types of work rather than surging dramatically in narrow domains. Lead researcher Matthias Mertens, a research scientist at MIT FutureTech, described the pattern as a rising tide rather than crashing waves—water gradually rising around everyone instead of suddenly knocking over a few people.

How the research was conducted

Unlike previous studies that relied on coding benchmarks or standardized tests, the MIT team examined AI performance on actual workplace tasks drawn from the U.S. Department of Labor's O*NET database. Expert workers evaluated whether AI-generated outputs met minimum standards, average quality, or exceeded expectations without requiring edits.

The research revealed that current AI models can already complete 50% to 75% of text-based tasks at a minimally sufficient level. When task complexity increased tenfold, success rates dropped by only six to seven percentage points—evidence that progress spans diverse work rather than concentrating on specific task types.

Why it matters

This gradual improvement pattern gives organizations, governments, and workers a concrete window for preparation rather than facing sudden obsolescence. The finding contradicts narratives of imminent, wholesale job displacement and suggests that adaptation strategies can be developed and implemented over time rather than in crisis mode.

Improvement pace and projections

Between the second quarter of 2024 and third quarter of 2025, frontier AI models improved from 60% to over 70% success rates on tasks requiring 1.5 hours of human work. The researchers found that AI failure rates are halving every 2.2 to 2.8 years.

Projecting forward, the team estimates most text-based tasks could reach 88% to 97% success rates by 2030 at minimum quality thresholds. However, near-perfect performance remains several years beyond that horizon.

Variation across occupations

AI capability varies significantly by work type. Text-based legal tasks showed success rates below 50%, while installation, maintenance, and repair tasks exceeded 70%. This disparity means some occupations face more immediate pressure while others have longer adaptation windows.

Important caveats

The researchers emphasized their findings should not be interpreted as a roadmap for immediate automation. In the study, AI received complete information for each task—a condition rarely met in real-world settings where data may be incomplete, difficult to integrate, or subject to regulatory constraints.

The projections also assume AI continues improving at its recent pace, representing an upper-bound scenario. Many jobs include tasks beyond language model capabilities, suggesting work will more likely be reorganized between humans and AI rather than eliminated entirely.

The research was conducted by MIT FutureTech, an interdisciplinary group formed by the MIT Computer Science and Artificial Intelligence Lab and the MIT Initiative on the Digital Economy, and first reported by MIT Sloan. The study was co-authored by nine researchers including Mertens, Adam Kuzee, Brittany S. Harris, Harry Lyu, Wensu Li, Jonathan Rosenfeld, Meiri Anto, Martin Fleming, and Neil Thompson, who directs MIT FutureTech.