Most companies now report adopting AI, and far fewer report results. Research from the 1970s by Chris Argyris and Donald Schön explains that gap better than any tool comparison: the theory you espouse is not the theory you use, and behavior only changes where the second one is reached. Here is what the original work says, and what it takes to anchor new behavior, in AI transformation as in any culture change.
The research began with a deceptively simple writing exercise. Argyris and Schön asked managers and professionals to describe a difficult conversation they had actually had, in two columns: on the right, what was said; on the left, what they thought and felt but did not say. Across hundreds of these cases, two patterns kept repeating. First, the strategy people claimed to follow rarely matched what their own transcript showed them doing. Second, while the claimed strategies varied colorfully from person to person, the behavior in the transcripts barely varied at all: stay in unilateral control, win rather than lose, avoid open embarrassment, appear rational while doing so. People were not lying about their approach. They simply could not see the difference. Half a century on, companies wondering why their AI adoption produces activity but not results are staring at the same gap.
In "Theory in Practice" (1974), Argyris and Schön gave the pattern its lasting vocabulary. Everyone operates with two theories of action: an espoused theory, the one we offer when asked how we act, and a theory-in-use, the one that actually governs our behavior. The two fall apart systematically, and the actor is usually the last to know: people are poor observers of their own action logic, and they sincerely report the espoused theory as if it were a description of their behavior.
In "Organizational Learning" (1978), they scaled the idea from individuals to organizations and drew the distinction that made the work famous. Single-loop learning detects and corrects error within the existing settings: the goals, norms, and assumptions stay fixed, and performance is brought back on track. Double-loop learning questions the settings themselves. Argyris liked a household example: a thermostat that turns the heat on below 68 degrees is a competent single-loop learner; it becomes a double-loop learner the moment it asks, in his words, "why am I set at 68 degrees?"
Their sobering organizational finding was that most organizations are very good at the first loop and structurally poor at the second. Error correction that leaves assumptions intact is rewarded and routinized. Error correction that threatens assumptions triggers defensive routines, because being confronted with your own word-deed gap threatens self-worth, and the defense itself is skilled: smooth changes of subject, diplomatic vagueness, meetings that end in agreement and change nothing.
Read through this lens, the standard machinery of culture and leadership change is aimed at the wrong theory. A new values statement, a leadership model, a freshly agreed strategy: all of it is the organization's espoused theory. The theory-in-use is written elsewhere, in this quarter's calendars, in the last ten hiring and budget decisions, in what actually gets rewarded, promoted, and quietly sanctioned. Where the two diverge, people do not learn the official model; they learn what really counts here, and that is the theory that governs. A new culture anchors only when the organization's everyday control mechanisms start teaching the same lesson the posters do.
The point is not that leaders are hypocrites. The original finding cuts deeper: the gap is invisible from the inside, which is why asking people about it, in surveys or town halls, mostly retrieves more espoused theory. You have to put the two records side by side. In our own strategy work we have started doing exactly that: keeping the choices we wrote down next to the decisions the following weeks actually produced, and treating any daylight between them as data rather than as someone's failing.
A decision rule we find useful: if you introduced a new strategy, leadership model, or set of values more than a quarter ago and you cannot name three decisions or behaviors that changed because of it, do not communicate harder. Reconstruct the theory-in-use from real decisions, place it next to the espoused one, and then work on what anchors it: what gets rewarded, what gets practiced, what gets reviewed. Producing that comparison takes a prepared session, not an afternoon of goodwill, and closing the gap is a matter of months.
The same finding explains why so much leadership development disappoints. If behavior is not governed by espoused principles, then teaching better principles, in seminars, models, and leadership frameworks, cannot by itself change behavior. Development has to reach the theory-in-use, and that is only accessible through observed behavior, feedback from others, and reflection on concrete situations, because self-description keeps returning the espoused theory no matter how sincere it is. Behavior change, in other words, is not launched; it is anchored, in routines that run on real work.
Argyris and Schön's own teaching method points the way, and a leadership team can run the original exercise almost unchanged. Take one consequential recent episode, perhaps the last time your new values met a hard trade-off. Each member writes the two-column case. Compare what was said with what was thought, then ask the double-loop question: which assumption made this the sensible way to act, and do we still want to be governed by it? The discipline lies in the repetition, not the single insight. In our team development work, the version that sticks is a standing reflection loop: act, gather feedback, reflect, adjust, on a rhythm the team keeps even in busy weeks.
Schön went on to develop this side of the argument in "The Reflective Practitioner" (1983), the book he is best known for, and the attention it still receives is deserved: skilled professionals improve not by applying theory to their work but by reflecting in and on their own action. Routines like the ones above simply put that reflection on the calendar. This is the conviction our whole approach is built on: a management practice is not something you read about, it is something you run.
On the Management Kits platform, the Working Topic "Routinize a structured approach to giving and receiving feedback" (in the Me-as-a-Leader kit) turns that loop into a working plan. Its Feedback Worksheet structures each exchange around context, observed behavior, impact, and next step, which supplies exactly the outside perspective self-report cannot; the associated milestone sets up recurring action-feedback-reflection-adjustment cycles against a personal learning agenda, about four hours of work spread over ordinary meetings. The learning agenda is where the second loop lives: most cycles adjust the action, and every so often the reflection revises the agenda itself.
AI transformation is, at bottom, a behavior change of exactly this kind, and most AI programs are single-loop by design. The workflow stays what it was; the tool makes it faster or cheaper; the governing assumptions about what the work is and who does it remain untouched. That is not wrong, but it maps onto what surveys of the last two years keep finding: a large majority of companies report adopting AI while struggling to turn it into business results. The gains that matter usually sit behind a double-loop question, whether this workflow, this handoff, this control still needs to exist at all.
There is a second, quieter connection. Learning from AI adoption depends on people surfacing failed experiments, half-working prompts, and honest confusion, precisely the embarrassing material that defensive routines are built to suppress. An organization that treats AI transformation as a tool rollout gets the thermostat. One that treats it as a learning agenda, with the reflection loops to match, gets to reset the temperature. That, rather than any particular tool, is the organizational problem to work on.
Argyris and Schön built conceptual theory out of action research and case analysis, not controlled trials, and there is no meta-analytic effect size attached to double-loop learning. The frame carries as an explanation, of why transfer designs, external feedback, and system alignment are necessary, and in that role it is broadly accepted and rarely contested; the quantitative case for the practices it points to comes from other literatures, among them the feedback-intervention research we wrote about in an earlier post. Argyris himself documented for decades how rarely organizations reach the fully open "Model II" ideal; treat it as a demanding direction, not a destination you certify. Nor is double-loop learning always the right loop: most operational work should run single-loop most of the time. And the two-column exercise asks people to show their left-hand column, which takes the kind of speak-up climate we examined in our post on Edmondson's original psychological safety study.
Which brings us back to those two columns. Fifty years on, the exercise has lost none of its bite, because the finding has lost none of its truth: the espoused culture is in the deck, and the left-hand column is where the culture actually lives. Change that anchors starts there, one honestly examined case at a time.
Want to see how your leadership team could run these loops on its live agenda? Schedule a short demo.