AI task length keeps doubling: what METR's research means for business data

The short answer: the tasks AI agents can finish keep getting longer

METR, a nonprofit that evaluates what frontier AI systems can do on their own, published research in 2025 titled Measuring AI Ability to Complete Long Tasks. It introduced a simple yardstick, the time horizon: how long a task takes a skilled person, measured at the point where an AI agent succeeds about half the time. Plotted across model releases, that horizon has grown exponentially, doubling at a fairly steady interval measured in months rather than years.

This page does not reprint METR's specific figures, because METR updates its measurements as new models appear and publishes confidence intervals that matter. Read METR's own publications for the current estimate. What follows explains what the measure means, what it does not show, and why it matters to companies that hold records of real work.

How METR measures a time horizon

The method is easier to follow than the headlines suggest.

1. Assemble a suite of tasks of very different lengths, from quick lookups to multi-hour projects, mostly in software engineering and research.
2. Time skilled people completing each task to set its human duration.
3. Run AI agents, equipped with tools, on the same tasks and record which they complete.
4. Fit a curve linking the agent's chance of success to each task's human duration.
5. Read off the duration at which success falls to about one in two: that is the model's time horizon.
6. Repeat for models released over several years and plot horizons by release date to see the trend.

The output is one number per model that can be compared across generations, which is why the doubling framing spread so quickly.

What the trend shows and what it does not

The last row matters for honesty. METR measures capability, not data demand. The link from longer tasks to demand for records is reasoning, set out below, not a finding of the research.

Why longer tasks change what training data is worth

A short task fits in one exchange: a question and an answer. A long task is a chain: gather information, make a decision, hand off to a colleague, hit an exception, recover, finish. Agents that work for hours need to learn from examples of whole chains, and the public web mostly holds finished outputs such as published articles, final reports and polished code, not the path that produced them.

Companies hold the path. The table maps task length to the records that capture a complete chain.

The further down the table, the scarcer the material outside companies, and the more a long, connected archive is worth. The explainer on why AI agents fail at real business tasks shows what goes wrong when agents lack this kind of example, and why AI data demand keeps growing puts the trend in market context.

What advisors can tell clients without overselling

The useful message is a measured trend from an independent evaluator, not a prediction about the client's industry. A short version for an owner or CFO:

For a wider annual view of AI capability, investment and policy, Stanford HAI's 2026 AI Index is a common reference; quote any figure from the report itself rather than from summaries.

Before raising the topic, check whether the client's records reflect long tasks:

• Threads link across systems, such as a ticket that references an engineering issue or a CRM deal tied to its email trail.
• Timestamps show the order of steps, not just the final state.
• Outcomes are recorded: resolved, won, approved, shipped, reopened.
• History runs back several years, including systems that were archived.
• The company created the records itself rather than holding them for its clients.

Skip the conversation for now if the archives were deleted when old tools were cancelled, if the records mostly belong to the client's own customers, or if a previous AI-training license already covers the same data. A long task history only helps when someone at the company can still export it and has the authority to license it.

Limits and open questions

Treat the doubling trend as a useful signal with clear limits.

• Task mix: the measured tasks lean toward software and research, where success is easy to check. Business workflows may behave differently.
• Success threshold: a horizon at about even odds says little about the dependability a finance or operations team needs.
• Forecasting: extrapolating any exponential curve is risky; the trend could bend either way.
• Data link: whether longer horizons translate into more demand for licensed records depends on how developers choose to train, which outsiders cannot observe directly.

Rights and provenance still decide whether any record can be licensed at all; developers' provenance checks are described in what ethically sourced AI training data means.

Where referral partners come in

A referral partner's job is to notice companies whose archives show long, connected work, then make an introduction. SourceX looks for US companies with 50+ full-time employees at peak (contractors excluded), a multi-year operating history, records spread across many systems, clear rights to license them and an authorized sponsor such as the owner, CEO or CFO. The company fit checker offers a fast first screen with no commitment, and how it works lays out each stage after the introduction.

Partners earn 25% of the eligible platform fees SourceX actually collects from the referred company's licensing deals, capped at $100,000 per referred company; the reward becomes payable only after the buyer pays and SourceX receives its fee, and no reward is guaranteed.

Next step

For a client whose systems tell the full story of long projects, register as a partner and make the introduction. Owners who prefer to start on their own can apply at sourcex.si/apply through your referral link, which keeps the introduction credited to you.