<![CDATA[Operating Mode]]>https://www.operatingmode.cohttps://www.operatingmode.co/img/substack.pngOperating Modehttps://www.operatingmode.coSubstackSun, 15 Nov 2026 19:18:20 GMT<![CDATA[Is AI paying off yet?]]>https://www.operatingmode.co/p/is-ai-paying-off-yethttps://www.operatingmode.co/p/is-ai-paying-off-yetSun, 15 Nov 2026 17:54:37 GMTIn McKinsey’s latest survey, most people surveyed said AI had made them more productive at work. The gains are easy to picture: a first draft, a meeting summary, a formula nobody has to look up any more. Far fewer companies can find them in their results. In another large survey published this year, nearly nine in ten executives said AI had made no difference to their company’s productivity over the past three years.

So the work is getting faster, but the company, so far, isn’t. The usual explanation is that companies have bolted AI onto organisations built for a different way of working, and that the returns will come once they redesign around it.

Fast tasks, flat results

At the level of a single task, the gains are real. When economists at Stanford and MIT gave customer support agents an AI assistant, the agents resolved 14% more issues per hour, with the biggest gains for the newest staff.

Zoom out to the company and the gains mostly disappear. That executive survey, run by economists from central banks and universities in four countries, found most firms saw no effect on labour productivity. A bigger follow-up this summer found only a small gain.

Denmark offers a harder test. There, economists at the University of Chicago and the University of Copenhagen could link chatbot use to government records on pay and hours. Two years after ChatGPT’s launch, the effect is statistically indistinguishable from zero.

The consultancies see the same pattern. In McKinsey’s 2026 survey, just over a third of respondents said AI had added anything at all to operating profit, no more than a year earlier. BCG is more upbeat this year, but it now uses a broader definition of success than it did last year.

The organisational explanation

If the tools work and the companies don’t benefit, the obvious place to look is everything in between. That is the consultancies’ answer, and it’s what they see in their data.

In McKinsey’s survey, the companies getting the most from AI are far more likely than the rest to have fundamentally redesigned their workflows around it. BCG says “most roadblocks involve people, organization, and processes”, and advises putting most of the effort there rather than into the technology.

These are surveys of executives rating their own companies, so on their own they show that redesign and returns go together rather than which drives which. McKinsey and BCG also advise companies on exactly this work. Both are good reasons to test the conclusion against independent research, which is where it gets stronger.

What history and research suggest

The independent evidence points the same way, and some of the best of it is old. When electric motors reached factories in the 1890s, most owners kept the steam-era layout of overhead shafts and belts, and simply hung motors on it. The economic historian Paul David showed that electricity didn’t lift manufacturing productivity until the early 1920s. The gains came once factories were rebuilt around a small motor on each machine, which in turn allowed lighter, single-storey buildings and new layouts.

Computers repeated the pattern. Economists at Stanford, MIT and Wharton found that IT paid off more in companies that also widened jobs, pushed decisions down and set up self-managing teams.

The first studies of AI look similar.

  • The gains take years and come from know-how. US public companies that invested more in AI saw productivity grow faster, with the gains building gradually over several years. The researchers trace this to “organisation capital”: the know-how a company builds by doing the work.

  • It often gets worse before it gets better. A study of US manufacturers using Census Bureau data found AI first made plants less productive. In older plants, a sizeable part of the drop came from letting established management routines lapse, such as tracking performance against targets.

  • AI users were already ahead. In Canada, firms using AI are more productive, but Statistics Canada finds that most of that edge comes from two other things: they were already more productive before adopting it, and they also invest more in areas like R&D, cloud computing and data analytics.

What’s still missing

My argument is that the early evidence points towards organisation. Companies seem to get returns from AI when they change how work is done around it, not when they simply add the tools. It’s early, and the evidence is thin in places, but it keeps pointing the same way.

The research supports this in a broad form: AI pays off when companies also invest in what surrounds it, meaning know-how, management routines, skills and processes. What it can’t yet say is which changes matter most. There is no direct evidence that redrawing the structure (the org chart, the layers, the shape of teams) is the fix, even though that’s where most of the current activity is.

That’s the question I’ll keep working on: what happens to the operating model, meaning who does what, who decides, and how teams are set up and measured, when part of the work is done by AI. I’ll use evidence where there is some and say so where there isn’t.

If AI has changed how your team works, what did you have to change around it for the gains to stick? Reply and tell me.

Sources

Thanks for reading. Subscribe for free to receive new posts.

]]>