Will AI really work four unsupervised days by 2027?

McKinsey cites METR's benchmark finding that "the length of tasks that AI can reliably complete doubled approximately every seven months since 2019 and every four months since 2024, reaching roughly two hours as of this writing," crediting METR's "Measuring AI ability to complete long tasks" (accessed September 2025). Building on that trend, McKinsey states that "AI systems could potentially complete four days of work without supervision by 2027."

McKinsey frames this trajectory through a progression of human-role analogues: AI moving "from an intern-level employee requiring constant supervision to a mid-tenure employee who can operate independently to, perhaps, a senior executive who can shape and drive strategies." The metaphor ties a quantitative capability metric — task-horizon length, as METR measures it — to a qualitative organizational claim about how much autonomy or oversight a role needs. McKinsey hedges the projection explicitly: "The promise of this new paradigm will depend upon the continued growth of AI's capabilities, as well as other factors such as regulation."

Longer projected task horizons sharpen the tension in Does granting agents more autonomy undermine human oversight?: the longer an agent runs unsupervised, the harder "above the loop" supervision becomes to exercise meaningfully, since that position paper argues autonomy itself erodes the overseer's ability to oversee. The projection also stands at a distance from Why do AI agents fail at workplace social interaction?, whose workplace simulation finds agents still topping out around 30 percent task completion on realistic office tasks — a sizable gap between METR's benchmarked task-horizon trend and a measured autonomy ceiling in a messier simulated workplace. The organizational vision in Can AI agents truly operate above the loop at enterprise scale? (the sibling note from this same source) depends on this capability trend continuing, since "above the loop" oversight presumes exactly the kind of multi-day unsupervised operation projected here.

What this excerpt does not establish: METR's own measurement covers a task-length-doubling trend on its benchmark suite, not a verified capability to work four unsupervised days — the four-day figure is explicitly McKinsey's own extrapolation ("could potentially"), not a METR finding, and it carries no information about task difficulty, error rate, or domain scope at that horizon. The implication, held no more strongly than a trend extrapolation warrants, is that the "intern to executive" framing describes a hoped-for trajectory rather than a measured capability milestone.