Here is a number to bring to your next steering committee. Non-tech companies in the S&P 500 that are “exploring” or “piloting” AI run net margins 2 to 3 points below companies that have not touched it.

Below. Not above.

That comes from AI Adoption in S&P 500 Firms, published in August by Yang Yu, Martin Fleming, Lucy Hampton, Christophe Combemale and Neil Thompson at MIT FutureTech and Carnegie Mellon. Chamath Palihapitiya included it in the slide deck that accompanies his September 25 deep dive on why AI is booming but productivity is not, so expect to see it quoted a lot this month. Most of those quotes will stop at the 11% headline. The margin finding is the one that should change how you run AI inside a company.

How they measured it

They did not survey anyone, which is why I trust it more than most AI studies. They read 10-K filings, the annual reports US public companies file with the SEC, the ones lawyers go through line by line because a false statement has consequences. Each filing got a score.

Level 1 means no mention of AI. Level 2, exploring. Level 3, AI inside specific products or processes with no financial impact claimed. Level 4, AI in production with expected financial results. Level 5, AI as a core part of strategy and financial performance.

4,470 scores. 510 companies. 2016 through 2025.

By 2025, 11% of the index sat at level 5, and two thirds of the companies at levels 4 and 5 were technology firms. Everyone else is further behind than the headlines suggest.

The dip, in numbers

Among non-tech firms, levels 2 and 3 carry net margins 2 to 3 percentage points below the non-adopters. Level 5 carries 12.6 points above, once size, sector and other firm differences are controlled for.

Three points of margin on a company doing $20 million a year is $600,000 that never reaches the bottom line.

Now the caveats, and I want to be straight about them. The authors say plainly that these are correlations. Stronger firms may be the ones that dared to integrate deeply. The non-tech level 5 group is small, and the confidence intervals are wide. They also found no relationship between AI adoption and revenue per employee, their productivity measure. Margins move. Productivity, not yet.

The shape is still familiar. Erik Brynjolfsson, Daniel Rock and Chad Syverson called it the productivity J-curve in 2018: every general purpose technology first drags down measured performance while the company pays for learning, integration and reorganization, and only later climbs above where it started. MIT found that J in S&P 500 net margins.

What the middle looks like from inside

Picture a company at level 2 or 3. Licenses paid. Training hours paid. A vendor pilot running somewhere. An AI chat window open on every laptop. And the process is identical: the same approvals, the same spreadsheets, the same five people touching every order.

All of AI’s cost. None of the redesign.

The pilot also feels like progress, which is what makes it dangerous. There are meetings, a board slide with screenshots, real enthusiasm. Nobody has a reason to say the thing has been running for a year without moving a number.

Where IQ Source stands on this

My read, and it is an opinion built on a correlation, is that there are two defensible positions and one that is not.

You redesign the process around AI and push toward level 4 or 5, accepting that you will pay for part of the J. Or you decide with evidence that the timing is wrong and stay at level 1 until you know where it pays.

The indefensible one is the pilot with no end date.

That is why our AI Maestro program ends in a Go or No-Go gate with a date on it and a person who signs. The first two months map how the company works today, where the flow actually gets stuck, and which AI use cases would move that constraint. If none would, the deliverable is a No-Go and the company skips the dip. If one would, the redesign is done before the next license gets bought.

I have written before about why adoption is not transformation and about finding the constraint that really limits your flow. MIT just put a price on ignoring both.

Before your next committee meeting, count how many AI pilots have a date when someone decides to scale or shut them down. Then count how many changed a step in the process, as opposed to the tool used for that step. If both counts are zero, you know which part of the curve you are standing on, and I am happy to help you map the way out.

Turn the pilot into a decision