AI won't automatically boost productivity for everyone

The Financial Times piece by John Burn-Murdoch and Sarah O’Connor poses the right headline question about AI and productivity. The problem is the singular noun. “Productivity” sounds like a neat variable you can toggle up with a breakthrough model release, when in practice it’s closer to a blurry collage of different kinds of work, measured with tools that were built for factories, not probabilistic software.

Start with measurement, because that’s where the FT piece quite sensibly looks. Output per hour is the staple; economists love it because it’s comparable and time‑series friendly. But treating that one metric as the scoreboard for AI misses what actually changes inside firms: error rates, speed of iteration, decision quality, and where human time gets reallocated. Those are not rounding errors. They are the work.

Think about the sectors that are leaning into AI tools: law, medicine, software, marketing. A system that catches drafting mistakes before a contract goes out, or flags a likely diagnostic miss, has changed the productivity of that task even if the billed hours and revenue stay constant this quarter. The benefit shows up as fewer disasters, not necessarily more output. The math doesn't lie: if the official lens is stuck on hours and dollars, it will systematically undercount quality and risk reduction, and then we’ll call AI “disappointing” because the national accounts can’t see what’s happening on the ground.

That’s my first issue with the binary framing — either AI is lifting productivity or it isn’t. Productivity isn’t an app you download and then watch your GDP ticker jump. It’s a composite of process improvements that sometimes show up as higher output, and sometimes show up as “nothing went catastrophically wrong this time.”

Now, the FT is right to press on whether any of this is visible in the aggregates. You can’t just assert invisible gains forever. But the more interesting story is where whatever gains do exist are landing.

Right now, the shorthand answer is: with the firms that own the infrastructure and the talent. That’s not villainy; it’s just how concentrated technologies behave. Cloud providers selling AI services, large enterprises with in‑house teams, and software platforms that bake models into their products are the ones cashing in early. Workers in routine roles feel the automation heat; specialists who can steer and supervise these systems get a complement, not a substitute.

Those are three different labor markets, none of which looks like the smooth “AI raises all boats” narrative. If you’re a policymaker staring at a decent GDP print, you might think things are broadly fine while one category of workers is quietly losing bargaining power and another is suddenly overbooked.

When I was at Goldman, I watched a small edge in tooling inside a trading desk or a risk team snowball into a structural advantage over competitors who were six to nine months late. The point wasn’t just “we’re more productive.” It was that the gap compounded. Tech‑driven productivity is often path‑dependent: once a few players get meaningfully better, they can out‑invest the rest, hire the best people, and lock in their lead.

The FT piece nods at diffusion — the comforting story where early adopters blaze the trail and everyone else eventually follows. That’s the textbook version from earlier tech cycles. The less tidy version is that diffusion can stall. Not because the tools don’t work, but because smaller firms don’t have the money, time, or managerial slack to rip up workflows and rebuild them around AI systems. If your CFO already hates your IT budget, you’re not volunteering for a multi‑year transformation project just to marginally improve document search.

Look at a company like Microsoft as a case study. When it folds AI assistance into productivity software, you’ve technically expanded access. But adoption still depends on whether managers trust the tools, whether employees know how to use them, and whether the organization redesigns tasks to take advantage of the new capabilities instead of just stapling them onto the old process. The technology ships in a quarter; the real productivity shift, if it happens, takes years of dull operational work.

That’s why the firm‑versus‑economy distinction the FT hints at deserves more airtime. You can have spectacular efficiency gains inside a handful of dominant companies and still see only modest movement in national statistics. Investors cheer; wage growth and broad‑based productivity barely budge. What looks like a macro mystery is often just a distribution problem disguised as a data puzzle.

There’s also a more awkward counter‑argument to the optimism: AI might be making some tasks faster while simultaneously bloating others. Generating more content, more code, more scenarios can mean more review, more compliance gates, more oversight. If AI writes the first draft in seconds but adds two extra layers of human checking because no one fully trusts it, the net productivity gain can be much smaller than the demo implied.

So when we ask whether AI is “finally” lifting productivity, we’re bundling together at least three separate questions: how we measure improvement, how gains spread across firms and sectors, and who within those firms captures the benefit. One clean headline number was never going to give a satisfying answer.

If the FT keeps asking this question over the next few years, the more telling shift won’t be in the headline growth rate, but in who stops caring about that number because their own internal productivity dashboards are telling a much more divergent story.