Most organizations measuring AI adoption are counting the easiest thing to see: how many people use the tools, and how often. Two pieces of research reported this week suggest that number is telling leaders less than they think.

Managers are using AI more than ever. They are also less excited about it. And a careful study of AI-assisted professional work found that better output and better skill are not the same thing.

Usage is a signal. It is not a standard.

Key takeaways

  • AI usage rates and polished AI-assisted drafts are signals. The standards leaders actually care about are better decisions, durable skill and real business results.
  • A Wharton and GBK Collective survey reported by Fortune on October 13, 2026, found daily AI use among managers rose to 62% while the share who described themselves as excited fell to 51%.
  • In the same survey, departments that measured AI by a growth metric and redesigned workflows were far more likely to report strongly positive returns.
  • An MIT-led field experiment with patent lawyers found AI raised the quality of drafts, but lasting skill gains without AI showed up only among senior lawyers.

What’s happening

On October 13, Fortune’s Nick Lichtenberg reported on the 2026 AI Adoption Report from Wharton Human-AI Research and GBK Collective, now in the fourth year of an annual study. Among managers and above, 62% said they use AI daily, up from 46% the year before. Over the same period, the share describing themselves as excited fell 8 points to 51%. A quarter called themselves skeptical, and 44% said they were cautious. The researchers described this as the first wave of the study in which usage and excitement moved in opposite directions. (The yearly waves use separate samples, not the same people tracked over time.)

Wharton’s Prasanna “Sonny” Tambe summed up the mood: “The amazement is kind of wearing off a little bit.”

The returns tell a more useful story than the mood. Eighty percent of managers reported positive financial returns from AI, but only 28% called those returns strongly positive. Departments that measured AI by a growth metric were far more likely to report strongly positive returns than those measuring efficiency alone: 35% versus 13%. Strong-return departments were also more likely to have changed their management structure (64% versus 47%) and to have significantly changed their workflows (84% versus 69%). According to Fortune, they did not report more staff cuts or hiring freezes.

Two other findings stood out to me. Seventy-two percent of managers and above said their focus has shifted to managing AI-generated outputs. And while 78% of managers said employees had received adequate AI training, only 56% of entry- and mid-level employees agreed. Junior employees were surveyed for the first time this year, and the two groups were not matched within the same companies, so treat that gap as a signal worth investigating rather than a precise measurement.

Three days earlier, on October 10, Fortune reported on MIT economist David Autor’s new study of AI and expertise. The NBER working paper, written with six co-authors, followed 133 patent lawyers at 11 U.S. intellectual property firms doing drafting work for Google. Two-thirds were randomly given a custom AI drafting assistant. After 90 days, AI access raised blind-graded draft quality by 0.38 standard deviations, about 11 percentile points. Fortune notes the gain came from less weak work, not more excellent work.

Then the researchers took the tool away. On a markup task done without AI, senior lawyers (seven or more years in practice) who had used the assistant beat their peers by 0.45 standard deviations. Junior lawyers showed no average gain. Their results spread out instead: more poor scores, fewer mediocre ones, more good ones. The authors wrote that AI “served as a springboard for some juniors and a cushion for others.” Seniors tended to use the tool as a “logic auditor.” Juniors often polished introductory text first, and sometimes noted serious flaws in comments rather than fixing them. Juniors also reported the largest effect in the study: a surge in task satisfaction.

The caveats matter. This is a working paper, not peer-reviewed. The sample was small, the study ran three months with one client, and only 91 of the 133 lawyers completed the AI-free test. Google conducted the study and paid its direct costs. Autor himself frames his recommendations as ideas, not proven fixes. But his headline point is hard to ignore: “Our results challenge the idea of AI being an automatic skill equalizer.”

What measuring AI adoption by usage misses

Read together, these two stories describe the same gap from different heights. At the department level, high usage did not guarantee strong returns. At the individual level, better AI-assisted output did not guarantee better skill. In both cases, the thing that was easy to count moved, and the thing that mattered did not always follow.

In Chapter 13 of The Next Turn, “Building a Shared Definition of Good Work,” I draw a line between standards and signals. Standards are what the organization actually needs: quality, reliability, accountability, sound judgment, follow-through. Signals are the visible cues we use to infer them. Hours at a desk used to be a signal for reliability. Polished output used to be a signal for skill.

AI adoption has created a new set of signals. Daily active users. Prompts per week. Drafts that come back clean. Training completion rates. These are not useless. But they are not the standard. A team can use AI every day and make no better decisions. A new employee can produce a polished draft and not yet know which part of it is wrong.

The Autor study puts a number on something I write about directly:

AI may reduce the production gap between an average and an exceptional employee. It does not eliminate the judgment gap.

The Next Turn

That is what the springboard-and-cushion finding looks like in practice. The tool lifted the floor on the work product. Whether it built the person depended on what they brought to it and how they used it.

The Wharton numbers connect to an argument from Chapter 6, “AI Is Not a Tool.” When leaders treat AI as a tool, they delegate it downward and measure whether people picked it up. When they treat it as infrastructure, they ask what it does to workflow, effort and judgment. The departments reporting strong returns were the ones that changed how work was structured and measured growth rather than efficiency alone. As I put it in the book:

Leaders who miss that will keep solving for adoption when they should be solving for redesign.

The Next Turn

This is the structure of work changing: the way tasks flow, who does what, and where human judgment sits. Leaders who only track adoption will see the change happening without being able to tell whether it is working.

And notice the training gap. Managers are far more likely than employees to say training is adequate. When leaders read high usage as readiness, that gap stays invisible. The goal is not faster. The goal is better.

Your next turn

If your organization is tracking AI adoption, here are five moves for this week:

  • Separate your AI signals from your AI standards. List every AI metric you report. Next to each, write the standard it is supposed to indicate. If you cannot name one, question why you are tracking it.
  • Add one growth measure. Alongside efficiency, pick one measure of what AI is helping the team do better or newly: decision quality, customer outcomes, new work taken on.
  • Check skill without the tool, occasionally. Autor suggests pairing AI rollouts with unassisted skill checks. That can be as simple as a short review conversation: “Walk me through how you would approach this without AI.”
  • Ask newer employees to frame first. One idea from the study is for less experienced people to frame the problem on their own, then use AI as a critic. Try it on one recurring task.
  • Test the training gap on your own team. Ask managers and employees the same question: “Do you have the AI training you need to do this work well?” Compare the answers.

Frequently asked questions

How should leaders measure AI adoption?

Usage rates are a starting point, not a finish line. In the Wharton and GBK Collective survey reported by Fortune, departments that measured AI by a growth metric were more likely to report strongly positive returns than those measuring efficiency alone. Pair usage data with measures of decision quality, outcomes and skill.

Does AI help junior employees build skills faster?

Not automatically. In David Autor’s patent-lawyer study, AI improved the drafts juniors produced, but on a test without AI, juniors showed no average skill gain, while seniors did. Results for juniors spread out in both directions, which suggests that how people use AI matters as much as whether they use it.

What is the difference between standards and signals at work?

Standards are what an organization truly needs, such as quality, reliability, accountability and sound judgment. Signals are the visible cues used to infer those standards, such as hours, responsiveness or polished output. AI makes many familiar signals less reliable, so leaders have to define the standards more explicitly.

The real question

I am not reading this research as a reason to slow down on AI. Tambe’s point about amazement wearing off sounds healthy to me. Excitement was never the standard either.

What matters now is whether leaders can see past the numbers that are easy to collect. Usage will keep rising. Drafts will keep getting cleaner. Neither one tells you whether your people are getting better at the work, or whether the work itself is getting better.

You can find more on the two turns reshaping work on the Next Turn blog.

So here is the question I would sit with this week: if you stopped counting how often your team uses AI, what would you count instead?

Sources