Talk of an AI bubble dominates the headlines. Some analysts warn of inflated valuations and looming layoffs, while others predict that AI will soon eliminate entire job categories. Both cannot be true. The evidence points elsewhere, to steady progress, real adoption and spending on the scale of infrastructure.
Progress has shifted. Improvements in chat tools may have plateaued, but the breakthroughs are now in AI agents, systems that plan, act and work on their own for longer stretches. In a blog post, Julian Schrittwieser argues that people keep failing to grasp exponential progress. The measure that counts now is time-on-task, meaning how long an agent can work without supervision. Schrittwieser projects agents working a full day at a stretch by mid-2026, if current trends hold. The projection could miss its date, but the compounding curve behind it is measurable.
METR, an independent research group, has proposed rating agents by the length of whole tasks they can finish, measured by how long each takes a person. Early data suggests that agents’ ability to work alone is rising along a stable path, so the kinds and lengths of work they can automate keep growing.
Even so, the OECD finds employment across advanced economies at or near record highs while investment in automation grows. Full automation of whole roles remains rare, because most jobs need context, coordination and judgment that AI still lacks. Agents can now finish repeatable subtasks with little oversight, especially in high-volume work such as data review, research summaries and reporting.
The payoff lands in operating margins. Agent-driven workflows cut labor costs, shorten project timelines and reduce errors, even with people checking the work, and the gains compound as agents run longer. The real risk is being overtaken by competitors, because teams that master agent-assisted work will pull ahead of those still using AI one prompt at a time.
Adoption is already widespread. According to Stanford’s AI Index 2025, the share of businesses using AI rose from 55 percent to 78 percent in 2024. US private AI investment reached $109 billion, a level that usually marks a decade-long infrastructure build. Railways, telecommunications and the public internet were the same kind of build, and all looked overvalued in their early phases before they transformed economies.
