The gap
Two numbers define the current moment, and the distance between them held steady across two consecutive annual surveys.
McKinsey has run its State of AI survey twice in the period this report covers: November 2025 (1,993 respondents, 105 countries) and August 2026 (1,719 respondents, 97 countries). Reading them side by side is more useful than reading either alone, because of what happened between them.
Adoption kept climbing. Eighty-eight percent of organizations reported using artificial intelligence in at least one business function in 2025; in 2026 it was nearly nine in ten. The question of whether to use AI has been settled by the market.
The share profiting from it did not move at all. Six percent qualified as high performers in 2025 — organizations attributing more than five percent of enterprise earnings before interest and taxes to their use of AI. In 2026 it was six percent again, and McKinsey says so in those words: unchanged from 2025. Widen the criterion to any measurable enterprise-level EBIT impact and the figure was thirty-nine percent in 2025 and thirty-seven percent in 2026 — still leaving a clear majority of adopters reporting no bottom-line effect they can name.
That last point deserves emphasis, because it is the premise of everything that follows. The frontier models are commercially available to every organization in the survey at broadly comparable cost. The high performers did not buy a different tool. They did something different with it once it arrived — and the survey is unusually clear about what.
Why this matters more in 2026 than it did in 2024
Two years ago, the honest answer to "why isn't this working yet" was that the technology was immature and the organization had not had time. The second survey removes that answer. Across those twelve months capability improved substantially and adoption rose further, while the share of organizations converting either into profit did not move by a single point. That is the value of having two readings rather than one: a single snapshot is compatible with "give it time," and two are not. When the input improves and the output does not, the constraint is somewhere other than the input.
For an executive team, this reframes the budget conversation entirely. The marginal dollar spent on a better model, a larger deployment, or a broader license is being spent against a constraint that is not binding. The binding constraint is organizational, and it is comparatively cheap to address — which is the good news buried in an otherwise discouraging statistic.
It is not the technology
The strongest separator between the 6% and everyone else is whether anyone changed how the work actually flows.
Both surveys asked what high performers do differently. Across the practices measured, the strongest single separator is workflow redesign — and it is the one measure that moved sharply between the two years. In 2025, fifty-five percent of high performers reported having fundamentally redesigned workflows as part of deploying AI, against roughly twenty percent of everyone else. In 2026, nearly three-quarters of high performers reported it, against just one quarter of other respondents. The separator did not merely persist. It sharpened, at a much higher base.
It is worth being precise about what "redesigning a workflow" means, because the phrase is used loosely enough to be meaningless. It does not mean inserting a model into an existing process and measuring the time saved at one step. It means changing the sequence: what gets done, in what order, by whom, with what decision rights, and with what removed. If nobody's job description changed and no approval step disappeared, the workflow was not redesigned. It was instrumented.
Why the redesign so rarely happens
Redesigning a workflow is not, at bottom, a technical act. It requires deciding who no longer approves something, which handoff is eliminated, what a role is now for, and who absorbs the work that used to justify a headcount. These are structural and political questions, and they are answered — or quietly refused — by people whose authority is defined by the current design.
This is the mechanism behind the gap. The technology arrives through a procurement process, is deployed by a technical function, and is measured on utilization. The redesign it would require sits with a different set of people who were brought in after the architecture was set, and who are asked to manage adoption rather than to change the work. Under those conditions, the predictable outcome is exactly what the survey found: a tool in wide use and a process nobody touched.
The three questions this document answers
1. Can we know in advance whether our people will carry a change of this size? — Section 3.
2. Do we know who actually influences whether they do? — Section 4.
3. Does a new way of working have anywhere to survive once it meets a quarterly target? — Section 5.
Each has an established measurement behind it, each can be run without hiring anyone, and each returns an answer in days rather than quarters. Section 6 assembles them into a thirty-day assessment.
Readiness is a number
Whether your managers will carry a difficult change is measurable before the change begins, and it is not a personality question.
Transformations rarely fail at go-live. They fail some weeks earlier, in a manager who privately concluded the thing would not work and stopped spending political capital on it. Organizations tend to treat this as a personality problem, which makes it feel unmanageable. It is better understood as a resource problem, and the resource has a name and an instrument.
Psychological capital — PsyCap — was formalized by Fred Luthans and colleagues in the mid-2000s as a higher-order construct with four components, conventionally abbreviated HERO:
- Hope. Does this person see a path to the goal, and can they generate an alternative path when the first one closes?
- Efficacy. Do they believe they have the capability to execute on it?
- Resilience. What happens to them when week three goes badly, as week three does?
- Optimism. Do they attribute setbacks to circumstances, or read them as evidence the effort is doomed?
The construct is measured with a validated questionnaire and has three properties that make it unusually useful before a large change.
It predicts
A meta-analysis by Avey and colleagues, covering a substantial body of studies, found psychological capital positively related to desirable employee attitudes, behaviors and performance, and negatively related to cynicism about change, turnover intentions, and job stress. Cynicism about change is the variable that matters here: it is the measurable form of the manager who has already decided.
It moves
Unlike personality traits, psychological capital is state-like. It responds to context and to how leaders behave. This is what separates it from a selection instrument: a low score is not a verdict on a person, it is a description of a condition you can act on. That distinction also determines how you should introduce the measurement internally — as a read on the organization's support of its managers, not an evaluation of the managers.
It is available in advance
The measurement can be taken before deployment. Engagement surveys, by contrast, tell you how people felt about something that already happened. The practical difference is a matter of months and, usually, of a great deal of money.
How to use it in an AI rollout
Measure the carriers, not the choosers. Baseline the frontline and middle managers whose teams will actually change how they work — not the executive group that selected the tool. The executive group is not where the failure occurs.
Read the variance, not the average. An organizational mean is nearly useless. What matters is whether there is a cluster of low scores concentrated in one function, because that function is where the rollout will stall and the aggregate will conceal it.
Re-measure at sixty days. The direction of travel is more informative than either single reading, and it tells you whether your communication and support are working while there is still time to change them.
One clarification worth making to any executive team that has run an engagement survey and considers the question answered. Engagement measures how people feel about the organization. Psychological capital measures whether they have the psychological resources to do something difficult. A workforce can be genuinely engaged and, at the same time, without the hope or efficacy to carry a change of this magnitude. Those are different readings, and only one of them predicts what happens in week three.
The network under the chart
Your organization chart records who reports to whom. It does not record who people actually listen to, and the second map is the one that decides adoption.
Every organization runs on two structures. One is formal, documented, and used for planning. The other is informal, undocumented, and used for getting work done. Organizational network analysis is the practice of drawing the second one, and the technique has been well established in management research since Rob Cross and Andrew Parker's work in the early 2000s.
The full method is more involved, but a usable version takes about fifteen minutes of survey time. Ask everyone in a department a single question:
Allow three names. Map the answers as a directed graph. What comes back is almost never the organization chart, and three patterns appear with enough regularity to be worth naming in advance.
The central node with no title
There is usually one person, often several levels down and frequently with no direct reports, through whom a large share of the department's advice-seeking flows. This person is a single point of failure the organization has not written down anywhere, and they are also the most efficient channel it has for changing what people believe.
The bypassed manager
There are usually one or two senior people the network routes around rather than through. This is diagnostic rather than damning — the cause is often workload or role design rather than competence — but a change plan that relies on those individuals to carry a message is relying on a channel that is not connected.
The single-thread team
There is often at least one team joined to the rest of the organization by exactly one person. That is a structural risk in ordinary operations and a guarantee of uneven adoption during a change, because everything that team learns arrives through one human bottleneck.
Why this decides AI adoption in particular
New ways of working spread through advice ties, not reporting lines. When someone is uncertain whether to trust a tool's output, they do not consult the intranet; they ask the colleague whose judgment they trust. If the people at the center of that network are unconvinced, a rollout becomes a broadcast into a system that is quietly routing around it — and the utilization dashboard will show adoption while the work stays exactly as it was.
The highest-leverage move in any rollout
Draw the map before the announcement, identify the central nodes, and bring them in early — with real access, a genuine say in the design, and the standing to say the tool is not ready. This costs almost nothing and it is the single most effective intervention available. It also fails completely if it is done cosmetically: these people are chosen by their colleagues precisely because they are hard to fool.
Where the new way lives
Most pilots do not fail. They succeed, and then have nowhere to go.
Complexity leadership theory, developed by Mary Uhl-Bien, Russ Marion and Bill McKelvey, begins from an observation that will be familiar to anyone who has run a pilot. Organizations contain two systems with genuinely incompatible logics. An operational system is built for efficiency, reliability and the delivery of committed results. An entrepreneurial or adaptive system is where novelty is generated, and it is built for exploration, which is inefficient by design.
Both are necessary. Neither can be run on the other's terms. The question that determines whether an organization adapts is what connects them — what Uhl-Bien and Arena call adaptive space: the conditions, linkages and protections that allow something new to move from where it was generated into where it must operate.
Without that space, the failure mode is not that pilots fail. It is that they succeed and are then crushed, entirely reasonably, by the operational system doing its job. A quarter closes. A target is under threat. The resources that were on loan to the pilot return to the work that is measured. Nobody made a decision to kill it; the structure simply had no place to keep it.
This explains a pattern most executives will recognize: a portfolio of successful pilots and no change in how the company operates. It is not a failure of ambition or of talent. It is a structural absence, and it can be tested for directly.
Three structural tests
Resource protection. Does the new way of working have a resource line — people, time, budget — that does not compete with a quarterly commitment? If the same manager must choose between the pilot and the number they are measured on, the choice has already been made.
Cross-boundary authority. Is there a named person with actual authority to change a process outside their own function? Redesign that stops at a functional boundary is not redesign; it is local optimization, and the survey data suggests it does not produce returns.
A real information channel. Does what is learned in the pilot reach the operational system through anything other than a status deck? Adaptive space requires people from both systems in sustained contact, not a monthly report travelling upward through a governance forum.
An organization that fails all three tests should not scale anything, regardless of how ready its people are or how healthy its network. The new way of working has nowhere to live, and every additional deployment adds cost against a structure that will return to its prior state as soon as attention moves on.
The thirty-day assessment
Three measurements, four weeks, no external help required. What you do with the result depends entirely on which combination you get.
The three diagnostics are individually useful and jointly decisive. Run in sequence, they can be completed inside a month by an internal team, and they produce a defensible answer to the only question that matters before a scaled deployment: is this organization in a position to convert it into anything.
| Week | What you do |
|---|---|
| Week 1 — Readiness | Administer a psychological capital measure to the frontline and middle managers whose teams will change how they work. Record the distribution by function, not just the mean. |
| Week 2 — Network | Ask one question of everyone in the affected departments: who they go to when stuck. Map the responses. Identify central nodes, bypassed managers, and single-thread teams. |
| Week 3 — Structure | Apply the three structural tests from Section 5. Interview two people who worked on the last pilot that did not scale, and ask them specifically what stopped it. |
| Week 4 — Read | Put the three results next to each other. The combination, not any single measure, tells you what to do. |
Reading the combination
| What you find | What it means, and what to do |
|---|---|
| Strong readiness, healthy network, no adaptive space | Your rollout will be enthusiastic and will be quietly reversed within two quarters. Do not spend more on communication. Fix the structure first — this is the most common and most expensive pattern. |
| Weak readiness, healthy network | You have the channel but not the conviction. Recruit the central nodes before the announcement, with real influence over the design. Convincing five people is cheaper than persuading five hundred. |
| Strong readiness, fragmented network | Willing people who cannot reach each other. Invest in the connective work — shared forums, deliberate cross-team pairing — before deploying anything further. |
| Weak on all three | Do not scale. Redesign one workflow end to end with one willing team, and use the result to build the case. This is slower on paper and faster in practice. |
The uncomfortable implication of the McKinsey data is that most organizations reading this will find the first pattern, and that the first pattern is the one most likely to be misdiagnosed as a communication problem. It is not. It is a structural one, and no amount of change messaging addresses it.
Method and sources
This document synthesizes three established bodies of organizational research and one contemporary survey. It is deliberately short on prescription and long on measurement, because the failure mode it describes is one where organizations act confidently on unmeasured assumptions.
The survey findings throughout are drawn from a single source, cited exactly, so that any figure here can be checked against the original. The three diagnostic constructs — psychological capital, organizational network analysis, and adaptive space within complexity leadership theory — each rest on peer-reviewed literature spanning two decades. The full bibliography is below — including the reviews that argue against one of the three constructs used here, which are listed alongside the ones that support it.
Survey data
McKinsey & Company. The State of AI in 2025. November 2025. Survey of 1,993 respondents across 105 nations. Source of the 88% adoption figure, the 2025 6% high-performer figure, the 39% any-EBIT-impact figure, and the 55% versus approximately 20% workflow-redesign comparison.
McKinsey & Company. The state of AI in 2026: On the road to ROI. 25 August 2026. Survey of 1,719 respondents across 97 nations. Source of the “nearly nine in ten” adoption figure, the 37% any-EBIT-impact figure, and the roughly three-quarters versus one-quarter workflow-redesign comparison. Source also of the statement that the 6% high-performer share is unchanged from 2025 — which is the single most important figure in this report, because it is the one that did not move.
Psychological capital
Luthans, F., Youssef, C. M., & Avolio, B. J. Psychological Capital: Developing the Human Competitive Edge. Oxford University Press, 2007.
Avey, J. B., Reichard, R. J., Luthans, F., & Mhatre, K. H. "Meta-analysis of the impact of positive psychological capital on employee attitudes, behaviors, and performance." Human Resource Development Quarterly, 22(2), 2011.
Luthans, F., Avolio, B. J., Avey, J. B., & Norman, S. M. "Positive psychological capital: Measurement and relationship with performance and satisfaction." Personnel Psychology, 60(3), 2007. Source of the PCQ instrument.
Organizational network analysis
Cross, R., & Parker, A. The Hidden Power of Social Networks: Understanding How Work Really Gets Done in Organizations. Harvard Business School Press, 2004.
Cross, R., Borgatti, S. P., & Parker, A. "Making invisible work visible: Using social network analysis to support strategic collaboration." California Management Review, 44(2), 2002.
Complexity leadership and adaptive space
Uhl-Bien, M., Marion, R., & McKelvey, B. "Complexity leadership theory: Shifting leadership from the industrial age to the knowledge era." The Leadership Quarterly, 18(4), 2007.
Uhl-Bien, M., & Arena, M. "Complexity leadership: Enabling people and organizations for adaptability." Organizational Dynamics, 46(1), 2017.
Uhl-Bien, M., & Arena, M. "Leadership for organizational adaptability: A theoretical synthesis and integrative framework." The Leadership Quarterly, 29(1), 2018.
Complexity leadership — the critical literature
Rosenhead, J., Franco, L. A., Grint, K., & Friedland, B. "Complexity theory and leadership practice: A review, a critique, and some recommendations." The Leadership Quarterly, 30(5), 2019, 101304. A structured review, published in the same journal that launched the theory, finding the literature largely conceptual and the empirical base thin. This is the source for the statement in Section 05 that adaptive space is a way of looking rather than a measure.
Tourish, D. "Is complexity leadership theory complex enough? A critical appraisal, some modifications and suggestions for further research." Organization Studies, 40(2), 2019, 219–238. Argues that the theory understates formal power and hierarchy — directly relevant to the adaptive-space test, which is a test of whether formal structure permits the behaviour the theory describes.
Servant leadership, and the author’s own research
Davis, L. The Relationship Between Servant Leadership and Psychological Capital: A Follower’s Perspective. Ed.D. dissertation, George Washington University, 2018. 402 employees at two mid-cap companies. Open access through GW ScholarSpace; the findings, and the limitations chapter, are summarized at aiorgscience.com/research.
Eva, N., Robin, M., Sendjaya, S., van Dierendonck, D., & Liden, R. C. "Servant leadership: A systematic review and call for future research." The Leadership Quarterly, 30(1), 2019, 111–132. The standard review, and the source of the caution that the construct is measured by a large number of competing instruments that do not agree with one another.
Liden, R. C., Wayne, S. J., Zhao, H., & Henderson, D. "Servant leadership: Development of a multidimensional measure and multi-level assessment." The Leadership Quarterly, 19(2), 2008, 161–177. The SL-28 instrument used in the dissertation.
Liden, R. C., Wayne, S. J., Meuser, J. D., Hu, J., Wu, J., & Liao, C. "Servant leadership: Validation of a short form of the SL-28." The Leadership Quarterly, 26(2), 2015, 254–269.
van Dierendonck, D. "Servant leadership: A review and synthesis." Journal of Management, 37(4), 2011, 1228–1261.
Greenleaf, R. K. Servant Leadership: A Journey into the Nature of Legitimate Power and Greatness. Paulist Press, 1977. The origin of the idea, and not an empirical work — worth knowing when it is cited as though it were one.
Psychological capital — measurement and critique
Luthans, F., & Youssef-Morgan, C. M. "Psychological capital: An evidence-based positive approach." Annual Review of Organizational Psychology and Organizational Behavior, 4, 2017, 339–366.
Dawkins, S., Martin, A., Scott, J., & Sanderson, K. "Building on the positives: A psychometric review and critical analysis of the construct of psychological capital." Journal of Occupational and Organizational Psychology, 86(3), 2013, 348–370. The most substantial published critique of the PCQ’s factor structure. Cited here because the Week 1 test in Section 04 depends on that instrument.
Lorenz, T., Beer, C., Pütz, J., & Heinitz, K. "Measuring psychological capital: Construction and validation of the Compound PsyCap Scale (CPC-12)." PLoS ONE, 11(4), 2016, e0152892. A freely available alternative to the licensed PCQ, for organizations that want to run the Week 1 test without buying an instrument.
Executive coaching
Theeboom, T., Beersma, B., & van Vianen, A. E. M. "Does coaching work? A meta-analysis on the effects of coaching on individual level outcomes in an organizational context." The Journal of Positive Psychology, 9(1), 2014, 1–18. Eighteen studies. Positive effects across five outcome categories, and the authors’ own observation that studies with control groups reported smaller effects than studies without.
Jones, R. J., Woods, S. A., & Guillaume, Y. R. F. "The effectiveness of workplace coaching: A meta-analysis of learning and performance outcomes from coaching." Journal of Occupational and Organizational Psychology, 89(2), 2016, 249–277. A second meta-analysis, with the finding that internal coaches outperformed external ones — which is inconvenient for anyone selling external coaching, this practice included.
Feedback and 360-degree instruments
Kluger, A. N., & DeNisi, A. "The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory." Psychological Bulletin, 119(2), 1996, 254–284. 607 effect sizes; the average effect is positive and roughly a third of interventions made performance worse.
Smither, J. W., London, M., & Reilly, R. R. "Does performance improve following multisource feedback? A theoretical model, meta-analysis, and review of empirical findings." Personnel Psychology, 58(1), 2005, 33–66. Twenty-four longitudinal studies; improvement smaller than the conventional threshold for a small effect.
Where each of these is discussed at length. The evidence file at aiorgscience.com/evidence takes the literatures above one at a time and states, for each, what it supports, what it does not, and what it would take to change the verdict — including the two entries where the verdict goes against something this practice uses. If a citation here is wrong, or a figure does not say what it is claimed to say, write and tell me: raim@aiorgscience.com. Corrections get published with the date.