Why AI Isn't Boosting Productivity Yet

Last week, a massive new NBER study surveyed 6,000 executives about AI, and the headlines were brutal: roughly 90% of businesses reporting no productivity gains despite $250 billion in investment. But buried in the data was a stat that reframed the whole story for me: those same executives are using AI for an average of just 1.5 hours a week.

Twelve minutes a day.

That’s not a technology problem. That’s a capacity problem. Companies are trying to bolt AI onto existing workflows that have absolutely no room for it. Every hour is accounted for. The legacy processes are still running, and the recurring meetings haven’t moved. There is nowhere for AI to go, because nobody removed anything to make space.

But there’s another layer to this.

Even among the people who are using AI seriously, the experience isn’t what anyone expected. A recent eight-month study from UC Berkeley, published in Harvard Business Review, found that AI didn’t reduce workers’ load. It intensified it. People voluntarily expanded into unfamiliar roles, dissolved their own work-life boundaries, and juggled multiple AI threads at once. Not because anyone told them to. Because the tool made it feel possible.

That study covers a specific population, but I’ve been watching the same pattern in my own work. Over the past six months, when I started using a connected sequence of AI tools for end-to-end concept development (refining ideas, running competitive analysis, drafting product specs, prototyping) I didn’t get my time back. What I got was something that looked a lot like a shadow cross-functional team. Researcher, analyst, writer, designer, all running in parallel, all producing output that required my judgment.

And that shadow team doesn’t operate in a vacuum. Its output still has to connect to your human colleagues. The AI-assisted research still has to hand off to the engineers. It still has to align with what UX is doing on their end, possibly with their own AI tool chains. You’re coordinating between a human product team and an AI-generated one now, and the interfaces between them are completely undefined.

In the book I’m writing about emergence, I keep tracking this pattern: when you add enough of a new capability to a system, the system doesn’t just speed up. It changes shape. We’re not getting faster product teams. We’re getting a different kind of product team, and nobody has figured out what that means yet.

We have a documented bias toward addition. When we see a problem, our first instinct is to add: a tool, a process, a meeting. Shopify went the other way and deleted 12,000 recurring meetings from their employees’ calendars in a single day. They didn’t add productivity software. They got out of the way.

What I’m watching for now is whether companies that do clear the space use it to rethink the work, or just refill it. Because clearing the calendar doesn’t help if you fill the space with the same tasks, just routed through AI.

In my own work, what I’ve found is that AI doesn’t behave like a faster version of the old process. It behaves like a parallel team that needs its own operating model. Subtraction makes room for it, but the harder question is what you’re actually making room for.