Ninety-Five Percent

Two pieces of news arrived close together this month, and I think they belong in the same thought.

The first was GPT-5. After more than a year of anticipation, OpenAI released it, and the reception was… mixed. It’s a very capable model. But it wasn’t the leap many people had been told to expect, and some users were so attached to the previous model that OpenAI brought it back within days.

The second was a report out of MIT, which found that the overwhelming majority of corporate generative AI pilots - the headline number was 95% - have produced no measurable impact on the bottom line.

The headlines practically wrote themselves. The AI bubble is popping. The emperor has no clothes.

I don’t think that’s what’s happening. But I don’t think the optimists are reading it right either.

Technology is fast. Organizations are slow.

Here’s what I think the 95% actually means. The technology works. It works remarkably well. What doesn’t work, mostly, is the way organizations try to adopt it.

A company buys an AI tool. It runs a pilot. A few enthusiastic people use it, others ignore it. The tool doesn’t fit the existing workflow, so it sits beside the workflow instead of inside it. Nobody changes how decisions get made, who’s responsible for what, or what gets measured. Six months later, someone asks for the ROI, and there isn’t any to point to.

That’s not a failure of the technology. It’s how almost every powerful technology has been adopted. Electricity took decades to transform factories, because factories first had to be redesigned around it. Computers famously showed up everywhere except in the productivity statistics, for years. The tool arrives quickly. The reorganization of human work around it takes much longer.

The five percent

The more interesting question is: what’s different about the five percent that do get real results?

From what I’ve seen, it’s rarely the sophistication of the technology. It’s usually something human. A specific, painful problem, rather than a general wish to “use AI”. A person who deeply understands both the work and the tool, and has the authority to change how things are done. Patience to iterate. Willingness to rethink the process instead of bolting AI onto it.

In other words: judgment, ownership, and care. The same things that made projects succeed before AI. The same things I keep coming back to.

On expectations

GPT-5’s reception teaches something too. We have become so accustomed to astonishment that a very good model now feels like a disappointment. That says more about us than about the model.

I find it helpful to remember where we were three years ago. If you’d shown me GPT-5 in 2022, I would have been speechless. The progress is real and it is enormous. It’s also not infinite or instant, and the stories we’re told about it - by companies raising money, by commentators selling attention - are almost always more dramatic than reality.

Wisdom here, I think, is to hold two things at once. The technology is genuinely transformative. And transformation is slow, uneven, and mostly human work.

What I’d tell a company

If you’re one of the 95%, don’t give up, and don’t double down blindly either. Start smaller and closer. Pick one real problem that someone in your company genuinely suffers from. Give it to someone who understands that problem deeply and is fluent with the tools. Let them change the process, not just add a tool to it. Measure honestly. Then do it again.

That’s slower than a company-wide AI transformation announced in a press release. It’s also how the five percent got there.

The technology arrived. The work of absorbing it has barely begun.