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 tells a different story, one that matters directly to business leaders managing growth, cost, and competitiveness.

Progress hasn’t slowed; it has shifted. While visible improvements in chat interfaces may have plateaued, real breakthroughs are emerging in AI agents: systems that plan, act, and work autonomously for longer durations. As Julian Schrittwieser argues, we consistently fail to grasp exponential progress. The key metric now is time-on-task, which refers to how long agents can operate without supervision. Schrittwieser projects day-length runs by mid-2026, if current trends hold. Treat that as directional, not deterministic, but the compounding curve is measurable and persistent.

Independent research confirms this shift. The METR team has proposed a “time horizon” benchmark based on human task duration, focusing on full-process execution instead of single-prompt accuracy. This shifts evaluation from toy benchmarks to real-world productivity. Early data suggests that agent autonomy is increasing along a stable trajectory, expanding both the types and durations of automatable workflows.

Meanwhile, enterprise adoption is accelerating. According to the Stanford AI Index 2025, business use of AI rose from 55 percent to 78 percent last year. U.S. private AI investment reached $109 billion, a figure that reflects scale, not speculation. Capital formation at this level typically signals a decade-long infrastructure build, not an imminent collapse.

The critical shift is from labor substitution to capability leverage. As the OECD’s AI and Work analysis shows, employment across advanced economies remains at or near record highs, even as automation investment intensifies. Why? Because full automation of entire roles remains rare. Most jobs require a mix of context, coordination, and judgment—qualities that AI still lacks. However, agents can now complete repeatable subtasks with minimal oversight, especially in high-volume, high-consistency functions such as data review, research synthesis, and reporting.

That matters for operating margins. Agent-driven workflows cut labor costs, shorten project timelines, and reduce errors, even with humans in the loop. These gains compound as agent duration extends and reliability improves.

The real risk is not job loss, but competitive displacement. Teams that master agent-assisted workflows will accelerate past those still stuck at the prompt level. The gap will widen with every quarter as productivity, quality, and responsiveness diverge.

What Smart Leaders Should Do Now

• Make time-on-task your core productivity metric. Measure how long an agent can complete a process before needing human input. Track outputs, failure points, and cost reduction over time. Time-on-task will soon be a key operational KPI.
• Deploy agents in repetitive, retrieval-heavy domains. Preparation, monitoring, and information gathering are ripe for multi-hour automation. Maintain human oversight for synthesis and judgment until agents reach enterprise-grade reliability.
• Train managers to design workflows, not prompts. The critical skill is orchestrating process logic, defining step-by-step execution, and embedding guardrails and acceptance criteria that map to outcomes.
• Establish governance early. The OECD finds that most European firms have adopted AI faster than they have built policy around it. Get ahead of shadow usage and data exposure by putting internal controls in place now.

This Is Infrastructure, Not Hyp

This is a story of leverage, not of speculation. Each doubling in autonomous task length increases potential margin. Each layer of integration reduces coordination overhead and shortens delivery timelines. The companies moving fastest are already reallocating talent to decision-making and high-value strategy instead of routine execution.

No, this is not a bubble. It is a capital-intensive infrastructure phase, just like the railways, telecommunications, and public internet before it. All those systems seemed overvalued in early phases before transforming economies. The true signal lies not in market chatter, but in deployment data, rising time horizons, and measurable efficiency gains.

For Leaders, the Path Is Clear

Treat agents as force multipliers, not employee replacements. Build the systems, skills, and safeguards that let your teams supervise multi-hour automated work. In an exponential environment, the cost of waiting compounds. Those who adapt first will own the next decade of productivity.

Sources

• Julian Schrittwieser, Failing to Understand the Exponential, Again — julian.ac
• METR, Measuring AI Ability to Complete Long Tasks (2025) — metr.org
• Stanford Institute for Human-Centered Artificial Intelligence, AI Index Report 2025 — hai.stanford.edu
• OECD, AI and Work — oecd.org