The Clock

METR (Model Evaluation and Threat Research) is an independent research lab that tracks the autonomous capabilities of frontier AI systems. Their methodology is straightforward: give the AI a real-world task, step back, and measure how long it can work without human oversight before it fails or needs help.

In early 2024, frontier models could handle tasks that took roughly 2 to 3 minutes of autonomous operation. Simple look-ups, basic reformatting, short summaries. By January 2026, that figure had reached approximately one hour. The growth rate between those points is consistent: a doubling every 89 days.

The critical distinction is that METR does not measure what an AI can do with a human supervising every step. It measures what an AI can do alone. This is the difference between a tool and an agent. A calculator is a tool. A system that can independently research a topic, draft a report, check its own work, and deliver a finished product is something else entirely.

Abstract doubling times are difficult to internalise. So here is what the METR trajectory looks like when projected forward from a January 2026 baseline, translated into the kind of work that fills a normal professional calendar.

Read the rightmost column again. By October 2027, the projection puts autonomous AI at roughly three weeks of continuous independent work. By April 2028, three months. These are not fantasy numbers. They are the mathematical consequence of a measured growth rate continuing at its current pace.

The obvious objection is that exponential curves do not continue forever. That is true. But the question for workforce planning is not whether the curve eventually flattens. It is whether it flattens before or after it passes through the task durations that define your job.

To understand why the METR doubling time is so consequential, you need to understand what knowledge workers actually do with their time. The answer, for most professionals, is not what they think.

Research from Asana, Microsoft, and the Anthropic Economic Index converges on a consistent finding: 60 to 70 percent of knowledge worker hours are consumed by coordination overhead. Not the work itself. The work about the work. Email, meetings, status updates, context switching, scheduling, and administrative tasks.

In a typical 50-hour week, most knowledge workers produce 12 to 18 hours of actual output. The rest is synchronisation: making sure the right information reaches the right person at the right time. This is the coordination tax, and it exists because human organisations are, fundamentally, communication bottlenecks.

AI does not have this problem. An AI agent does not need to attend a status meeting. It does not lose 23 minutes of focus every time it switches context. It does not need to be cc'd. When AI systems handle coordination tasks, they do not just do them faster. They eliminate the need for many of them entirely.

Jack Dorsey, co-founder of Block (formerly Square), estimated that 80 out of 200 employees in a typical company exist purely to manage handoffs between the other 120. If AI compresses the builder base by 50%, the coordination layer does not shrink by 50%. It shrinks by roughly 90%, because there are fewer nodes to synchronise.

Capability growth alone does not restructure economies. Cost does. And the cost trajectory of AI inference is, if anything, more dramatic than the capability curve.

In late 2022, GPT-4 equivalent capability cost approximately $20 per million tokens. By early 2026, equivalent capability costs between $0.07 and $0.15 per million tokens. That is a 280-fold reduction in 36 months.

The economic consequence is a textbook Jevons Paradox. When the cost of a resource drops dramatically, total consumption does not fall. It explodes. Despite the 280-fold drop in per-token cost, total enterprise spending on AI inference grew 320% over the same period. Organisations are not saving money on AI. They are deploying it into every process they can find.

This matters for the doubling time because cost is the practical governor on deployment. An AI system that can autonomously handle a day of work is only disruptive if it is cheap enough to deploy at scale. At $0.07 per million tokens, it is.

This is not theoretical. The restructuring has already begun.

Klarna, the Swedish fintech company, reduced its workforce from 5,527 in 2022 to 2,907 by late 2025: a 47% reduction. AI handled two-thirds of all customer service chat enquiries autonomously. Revenue increased 108% while operating costs remained flat. Revenue per employee rose 73%. The numbers are public.

In Australia, the pattern is emerging in the technology sector. Atlassian cut 1,600 roles globally, including 500 locally. Block (Afterpay) eliminated 4,000 roles globally, 700 in Australia: a 40% reduction. WiseTech Global cut 30% of its workforce. A Tech Council of Australia survey found 78% of technology leaders identified AI as the top influence on their 2026 operational strategy.

These are not struggling companies shedding staff. These are profitable, growing businesses restructuring around AI capability. The workforce reductions are happening alongside revenue growth, which is precisely the pattern the METR doubling time predicts: each quarter, AI handles more, humans handle less, and the output stays the same or improves.

The critical detail from the Klarna case is often overlooked. The company eventually walked back its total reliance on AI customer service after severe consumer backlash and quality drops. It moved to a hybrid model: AI for tier-1 enquiries, humans for complex escalations. This is the pattern that will define the transition. Not full replacement, but structural compression. Fewer people, doing different work, supported by systems that handle everything routine.

Australia is unusually exposed. The IMF assessed that almost 40% of global employment faces significant AI exposure. For Australia, the figure is approximately 60%, because our economy is disproportionately weighted toward services: roughly 63% of GDP.

The fiscal dimension is where it becomes dangerous. Australia's top 15% of earners contribute 68% of total personal income tax collected. These are the managers, professionals, and administrative workers who populate the services economy. They are also the occupational categories most exposed to AI-driven task automation.

Personal income tax is projected to reach 53% of total government revenue by 2035-36, nearly double the OECD average. The government's fiscal strategy is built on the assumption that high-earning knowledge workers will continue to earn, and continue to pay tax, at current rates. The METR doubling time suggests that assumption deserves serious scrutiny.

Part 4 of this series, "Australia's $300 Billion Blind Spot," will examine the fiscal implications in detail. For now, note the structural vulnerability: the occupations most likely to be compressed by AI are the same occupations that fund the government. If the METR clock is even approximately right, this is not a 2035 problem. It is a 2028 problem.

If the doubling time seems too aggressive, consider what the capital markets are pricing in. In late January and February 2026, software-as-a-service stocks experienced what analysts labelled the "SAASpocalypse": a repricing event that erased $1 to $2 trillion in market capitalisation in weeks.

Atlassian fell 35 to 50%, erasing approximately $11.5 billion in founder wealth. Salesforce declined 27 to 28%. ServiceNow, Workday, and DocuSign all dropped 20 to 37%. The market's thesis was straightforward: if AI agents can handle tasks that previously required specialised software platforms and the people trained to operate them, the addressable market for those platforms contracts.

The market is not always right. But when trillions of dollars move in a consistent direction over a period of weeks, it reflects a collective assessment by investors with material incentives to be accurate. The assessment: AI capability is growing fast enough to restructure white-collar work within the current planning horizon of most institutions.

The METR doubling time is a clock. It is not a destiny. The curve will flatten at some point. Physical constraints, energy limits, data exhaustion, regulatory friction, and sheer complexity will slow it.

But the question for anyone reading this is not whether the curve flattens. It is when. If the 89-day doubling time holds for even another 18 months, autonomous AI will be operating at the level of multi-day projects by early 2028. If it holds for three years, the figure reaches multi-month engagements.

Most career plans, most institutional strategies, and most government policies are built on the assumption that the current pace of change is approximately the future pace of change. The METR measurement says that assumption is wrong by a factor of two every 89 days.

The clock is ticking. The only question worth asking is: does your plan account for it?