Copilots didn't move the macro needle. Now comes the agent wave.

For all the noise about AI transforming knowledge work, it hasn't shown up where economists look first. US, UK, and EU output-per-worker figures have refused to budge, even as generative AI adoption across offices and dev teams has surged. The Financial Times ran the numbers this week. The numbers are not encouraging — at least not yet.

This isn't a new story. Robert Solow noticed the same thing about computers in 1987: you could see the technology age everywhere except in the productivity statistics. The observation has a name now, and it's back.

The evidence that AI actually works does exist — it's just not in the macro data. GitHub Copilot speeds up developers on specific tasks by somewhere between 20 and 55 percent, depending on the study. AI-assisted customer service agents handle calls faster. These are real, reproducible gains. The problem is they come from controlled settings with bounded tasks. What happens between those results and what national statistics eventually capture is the puzzle that economists are still circling.

The IMF, OECD, and McKinsey have all published estimates of what AI could add to total factor productivity. Their numbers don't agree with each other. The explanations for the gap cluster around three things: diffusion lags, accounting frameworks that weren't built to measure AI-augmented work, and upfront implementation costs — training, workflow redesign, organisational change — that eat into early gains.

Where this gets pointed for the agent industry is the next layer. First-generation copilots augmented individual tasks. Agents do something structurally different: they compress or replace entire process chains. Legal document review, financial analysis, professional services workflows — not individual tasks sped up, but sequences of work being automated end to end. That's a categorically larger surface area for productivity impact.

History says it still won't show up in the statistics for a while. Electrification took decades to bend the productivity curve. So did computing. The firms capturing gains now will be well ahead by the time aggregate data catches up — which is either a reason for patience or a reason to move faster than competitors who are waiting for proof.