Only a Third of Product Teams Using AI Are Better For It.

Melissa Perri and the team at Product Circle just published a proper survey of how AI adoption is actually going inside product organizations. These are the numbers that caught our attention. 87.7% of the product leaders surveyed say their teams use AI coding assistance. 85.4% use it for research and writing. Adoption, in other words, is nearly universal. And yet only 36% say AI is strengthening their product operating model. The product operating model is the way their organization decides what to build, learns from customers, and ships.

Melissa says it best: "the tools changed, the work did not."

There’s a reason (more than one to be honest) why there’s a 50-point gap between the adoption and those seeing value from AI. Let’s dig into some of those reasons.

What the State of AI in Product 2026 survey actually found

Three findings from the report (and Melissa’s walkthrough of it) deserve more attention than the headline numbers.

First, the bottleneck hasn't moved. Teams are faster at producing code, research summaries, specs, and prototypes while the upstream work of discovery and deciding what's worth building looks the way it did two years ago. We’re making more stuff, faster. We’re not making better decisions.

Second, AI is acting as a multiplier, not an equalizer. Organizations that already had a successful product operating model with clear outcomes, real discovery, and decision capabilities are compounding. Organizations that didn't are mostly generating more output, faster, against the same unexamined priorities. The tools amplify whatever operating model they land in, including a broken one.

Third, there's a gap between executives who believe an AI strategy exists and the product managers who say it never reaches their work. Leadership believes the org has a plan. The people doing the work are still guessing what it means for their roadmap.

AI adoption gap isn’t evenly distributed or enabled

Nearly all AI tool deployment went to coding assistance, research synthesis, and writing support. There is certainly value there and it is helping individuals produce more output, faster. Earlier this year Josh and I worked with an AI team at a large European bank, and the pattern there was the same one the survey describes: impressive individual productivity, and a roadmap process that hadn't changed at all.

The product managers, team leads, and middle managers who turn strategy into weekly decisions didn’t get anything new for the part of the job that actually determines whether the speed pays off. Which customer problems are worth this newly cheap capacity? What are we no longer going to build, now that we could build almost anything? How will we know any of it worked?

Those are judgment calls. No AI copilot makes them for you. And in most large organizations, the people expected to make them are implementing someone else's AI mandate with no new decision-making authority and no new way of working. The org bought tools for the layer that produces and changed nothing for the layer that decides (the 36% in the survey).

So what do we do with this? How does a team that's already adopted the tools get itself into the third that's successfully exploiting it?

Here is a quick question that will help in your next roadmap meeting

The 36% in Melissa’s data are better because their product operating model was already in place and strong enough for the tools to amplify. That’s not special in any real way. You and your team can build that in right now and then add in the new technology to amplify it.

Try this in your next roadmap or planning meeting. For each AI initiative or AI-assisted piece of work on the list, ask one question: which decision does this change? Not which task it speeds up. Who decides differently, about what, and how would we observe that in customer behavior?

If the answer is some version of "it makes us faster," your product operating model needs maturing. Faster toward an outcome nobody chose deliberately is just a more efficient way to be wrong.

This is the exact problem we work with in our training workshops. We hep teams reconnect the newly cheap execution capacity to the judgment layer that's supposed to steer it. But you don't need us to start. You need the question, your current roadmap, and a frank converastion.

Go read the full report — Melissa and Product Circle did the field a real service publishing it. Then run the question with your team, and let me know how it goes.