Building an AI Culture: Your Systems Count Twice as Much as Their Mindset
When AI does not take hold on a team, the explanation reached for first is almost always about the people. They are not curious enough. They are set in their ways. The younger ones will get it.
Microsoft put numbers on that assumption this year, and the numbers do not support it.
What they measured, and why it is worth more than the usual survey
Their 2026 Work Trend Index tested 29 factors against how much impact people reported getting from AI, and sorted them into organizational and individual groups. The result:
organizational factors like culture, manager support, and talent practices account for more than 2x the reported AI impact of individual factors like mindset and behavior (67% vs. 32%).
The method is better than a report like this usually bothers with. They used A way of ranking which inputs actually drive an outcome. You train a model to predict the outcome, then scramble one input at a time and see how much worse the prediction gets. Scramble something important and accuracy falls off a cliff. Scramble something irrelevant and nothing happens. It is useful here because it copes with factors that overlap, which “culture” and “manager support” obviously do. , then checked it against two other model families. All three agreed, with held-out accuracy scores of 0.680, 0.689 and 0.690, which means the model was tested on people it had not seen rather than graded on its own homework.
Two limits before anyone builds a strategy on it. Everything here is one person’s answers on one questionnaire, so somebody feeling good about their employer will tend to rate the culture and the AI impact generously, which inflates the link between them. And the survey covers 20,000 people who already use AI, so it explains what separates high-value use from low-value use. It says nothing about people who have not started.
Three levers, and you have probably pulled two
The report names its three organizational factors specifically: “a culture that supports new ways of working with AI, managers who model AI use and encourage experimentation, and talent practices that reflect AI in how people are evaluated and developed.”
The first two are well covered ground, including here.
Culture that supports new ways of working is the permission problem. Naming what AI is supposed to make faster, and saying out loud that time spent making it fit counts as work, is the argument in Your Team Is Building AI Tools After Hours.
Managers who model it is the finding How to Train Your Team on AI is built on, where Gallup found 79% frequent use when a manager actively supports it against 46% when they do not. Modeling is stronger than supporting: it means you are visibly using the thing.
The third one is the one almost nobody touches.
Nobody changes the thing that tells people what matters
The third lever is How people get hired, evaluated, promoted and developed. The performance review, the promotion criteria, the competencies in the job description, the things your skip-level asks about. It is unglamorous HR machinery, and it is also the most honest statement your organization makes about what it values, because it is the one attached to money. , and it is where the other two levers either stick or quietly come undone.
Think about what an employee is actually reading. Their manager says AI is a priority. Their performance review does not mention it. The promotion criteria do not mention it. Nobody has ever been asked in a skip-level what they automated this quarter. Meanwhile the delivery targets are unchanged and measured weekly.
That person is not resisting AI. They have read the situation correctly. Two things are telling them what the company values, and only one of them decides their next raise.
Microsoft’s own guidance to leaders lands exactly here: redesign the systems so that “metrics, incentives, and expectations reward people for changing the way they work.” Not so that they reward AI use, which is a metric that goes bad quickly. So that they reward the change in how the work gets done.
Does this mean the resistance research is wrong?
It looks that way at first, and it is worth resolving, because this site has argued both.
Why People Resist AI at Work is about an individual factor: people quietly decline AI when it threatens the thing they are good at. If organizational factors carry twice the weight, is that post overstated?
No, for two reasons.
32% is a third, not nothing. The claim is that organizational factors matter about twice as much, not that individual ones are noise. A third of the explained variation is a great deal to leave lying around.
More usefully, the two are not independent. Mindset is downstream of the system. Somebody whose manager models AI use, whose team treats a failed experiment as normal, and whose review acknowledges the work, arrives at a different mindset than the same person under a manager who does none of that. The split is not really nature against nurture. It is a statement about which end you can push on, and you can push on the system.
Which is also why the split is good news. Mindset is not something you can requisition. A review template is.
What this changes on Monday
Nothing here requires a budget, which is the point of it.
- Read your own performance review. If AI, automation, or changing how work gets done appears nowhere on it, your team already knows that.
- Ask what the metrics reward. If output targets went up and nothing else changed, you have told people to absorb AI gains as extra volume. That is a decision, and it is the one that happens by default.
- Model it visibly. Not “I encourage the team to use AI.” Show a thing you made with it, including a time it went badly.
- Make one experiment legitimate in public. Name it, name who owns it, and say what it is allowed to cost.
The manager’s toolkit is the ordered version of this if you want to run it properly rather than piecemeal.
The takeaway
The instinct when AI does not land is to work on the people. The largest measurement available says the people are the smaller half of the problem.
Culture, manager modeling and talent practices explained 67% of reported AI impact, against 32% for individual mindset and behavior. Two of those three are things most managers have at least attempted. The third, whether your evaluation and promotion criteria reference AI, is the one almost nobody has touched, and it is the one your team is actually reading.
Sources: all figures are from Microsoft’s 2026 Work Trend Index (opens in a new tab), surveyed by Edelman Data x Intelligence among 20,000 full-time knowledge workers who already use AI, across 10 countries, between 18 February and 20 April 2026. The 67% against 32% split comes from a random forest permutation importance analysis of 29 factors across 19,854 respondents, cross-validated with elastic net regression and gradient-boosted trees (held-out R² of 0.680, 0.689 and 0.690). Read the figures at Microsoft’s own page rather than the coverage of it, which repeatedly reports the report’s Frontier Professional figures as though they applied to all AI users. Two caveats stated in the body and worth repeating: this is self-reported impact regressed on self-reported conditions from a single questionnaire, so the link between them is probably flattered, and Microsoft sells the product whose value is being measured. The Gallup manager-support figures are from AI in the Workplace (opens in a new tab) (23,717 US employed adults, February 2026). The argument that talent practices are the neglected third lever, and the reconciliation with the resistance research, are mine, not claims either source makes.
Related: How to Train Your Team on AI: Why the Course Isn’t Working is the manager-support lever in full. Who Owns AI Adoption? When It’s Everyone’s Job, It’s Nobody’s is what happens when the system names no one.