The Bedrock
Twenty-one foundational works, summarized for leaders
Most AI strategy is built on intuition about a domain where the science is settled. The research on how organizations think, how groups produce intelligence, how technology reshapes work, and how organizations change themselves has been accumulating for seventy years. The studies making headlines right now — the ones about AI's effects on productivity and teams — are applications of ideas that were formalized decades ago. Leaders who know the foundations will read the new findings with sharper eyes and make fewer expensive mistakes.
Below are twenty-one works that constitute the intellectual bedrock, organized by the question each one answered. Each includes a link to buy or access the original.
I. How Do Organizations Actually Think?
Four works that established organizations as cognitive systems — and showed that the quality of an organization's intelligence is a product of its structure, not its headcount.
1. Herbert Simon — Bounded Rationality (1947 / 1955)
Simon won the Nobel Prize for observing what should have been obvious: people don't optimize. They lack the information, the time, and the cognitive horsepower. They search until they find something acceptable and stop. He called this "satisficing."
The organizational implication ran deeper than the individual one. If humans satisfice, then the quality of an organization's decisions has less to do with hiring brilliant people and more to do with designing structures that channel attention, route information, and constrain search in productive directions. The organization is a cognitive architecture. Its design determines its intelligence.
What this means now: When your team adopts AI and stops questioning the output — which is what Dell'Acqua's Harvard study found happens with elite consultants — they are satisficing at a new, lower threshold. The AI delivers a "good enough" answer faster, so they stop searching sooner. The cognitive architecture hasn't improved. It has shortened.
Read it: Administrative Behavior, 4th ed. | "A Behavioral Model of Rational Choice" (PDF)
2. March & Simon — Organizations (1958)
March and Simon reframed organizations not as authority structures but as attention-allocation systems. How an organization directs its members' attention — through goals, sub-goals, roles, communication channels, and procedures — determines what problems get worked on and which get ignored.
What this means now: AI doesn't just reassign tasks. It redirects attention. When a tool surfaces information, generates drafts, or rank-orders options, it is re-engineering the attention structure of your organization. If that re-engineering is unintentional — and in most deployments, it is — you have handed the design of your cognitive architecture to the defaults of a software product.
Read it: Organizations, 2nd ed.
3. Cyert & March — A Behavioral Theory of the Firm (1963)
Cyert and March studied how organizations actually make decisions. Three findings stood out. Organizations don't have unitary goals — they are coalitions with partially conflicting interests. Organizational search is local: problems prompt people to look for solutions near what they already know. And organizational learning is shaped by the very structures that produced earlier decisions.
What this means now: AI pilot programs succeed because they are local search. A team tries a tool on a bounded problem, gets a measurable gain, and declares victory. The gain doesn't scale because the broader search routines — the handoffs, approval chains, information flows — haven't changed.
Read it: A Behavioral Theory of the Firm, 2nd ed.
4. Karl Weick — Sensemaking in Organizations (1979 / 1995)
Weick argued that organizations don't discover pre-existing meaning in their environments — they construct it. Retrospectively, socially, and through action. Sensemaking is not rational analysis but narrative construction under ambiguity.
What this means now: The "miscalibrated trust" finding in the Harvard/BCG study is a sensemaking failure. Those consultants built a narrative about AI reliability based on early experience with tasks inside the frontier. That narrative became the lens for all subsequent AI output — including output that was wrong.
Read it: Sensemaking in Organizations
II. How Does Knowledge Live in Groups?
Three works on the mechanisms by which teams and organizations hold, transfer, and create knowledge — and where those mechanisms are most exposed to disruption.
5. Daniel Wegner — Transactive Memory (1987)
Wegner studied how couples and small groups manage more knowledge than any member holds individually. His answer: they build a shared directory of who knows what. The transactive memory system has three components: specialization, credibility, and coordination.
What this means now: Introducing AI to a team inserts a new node into the transactive memory system — one the team cannot read the way it reads human colleagues. A 2024 meta-analysis found that human-AI combinations performed significantly worse on average than the best of humans or AI alone.
Read it: "Transactive Memory" (chapter)
6. Ikujiro Nonaka — Organizational Knowledge Creation (1994)
Nonaka studied how organizations create new knowledge through four types of conversion between tacit and explicit knowledge. The hardest and most valuable is articulation — getting experienced practitioners to put into words what they know but have never spelled out.
What this means now: Generative AI is strongest at combining existing explicit knowledge into new forms. What it cannot do is pass on the kind of knowledge that comes from shared experience — judgment, instinct, reading the room.
Read it: "A Dynamic Theory of Organizational Knowledge Creation"
7. Linda Argote — Organizational Learning (1999)
Argote established that organizational knowledge resides in three reservoirs: individuals, routines, and technologies. Learning happens when experience modifies any of these. But knowledge transfer between groups is systematically harder than knowledge creation within them.
What this means now: AI operates on explicit knowledge — the kind that can be encoded and transferred at scale. But the most strategically differentiating knowledge is tacit: judgment, contextual pattern recognition, the ability to read a situation.
Read it: Organizational Learning, 2nd ed.
III. How Do Groups Produce Intelligence?
Three works on the specific mechanisms by which collections of people produce cognitive output better than any member could alone.
8. Woolley et al. — Collective Intelligence (2010)
Published in Science, this study showed that groups have a measurable collective intelligence — a "c factor" — that is not correlated with the average or maximum IQ of group members. What predicts it? Social sensitivity, equality of conversational contribution, and the proportion of women.
What this means now: AI disrupts all three predictors. It has no social sensitivity. It alters turn-taking. And it compresses cognitive diversity.
Read it: "Evidence for a Collective Intelligence Factor"
9. Scott Page — The Diversity Prediction Theorem (2007)
Page proved that a group's collective accuracy depends as much on how differently its members think as on how smart they are individually. A group of diverse adequate problem-solvers will outperform a group of homogeneous experts.
What this means now: AI raises average individual quality while reducing diversity of output. By Page's math, those two effects can cancel each other — or produce a net decrease in collective accuracy.
Read it: The Difference
10. J. Richard Hackman — Conditions for Team Effectiveness (2002)
Hackman spent thirty years studying what makes teams work. His conclusion: team performance is a function of structural conditions that can be deliberately designed. Most team failures are design failures.
What this means now: Hackman's framework is the missing layer in most AI-team deployments. The AI has changed the enabling structure without any corresponding adjustment to the other four conditions.
Read it: Leading Teams
IV. How Does Technology Actually Change Work?
Two works on why identical technology produces different outcomes in different organizations.
11. Wanda Orlikowski — The Duality of Technology (1992)
Orlikowski argued that technology and organizations don't stand in a one-way relationship. The same technology deployed in two different organizations will produce different outcomes — not because the technology differs, but because the social practices around it differ.
What this means now: This is the deepest explanation for why AI benchmarks and case studies don't transfer. Copying another company's AI implementation without copying its organizational context is imitation without understanding.
Read it: "The Duality of Technology"
12. David Autor — Task-Based Analysis of Automation (2015)
Autor explained why automation consistently fails to eliminate jobs while consistently eliminating tasks. Most jobs are bundles of tasks — some routine, some not. The key move was shifting from the job to the task as the unit of analysis.
What this means now: If your AI strategy is organized by job title rather than by task topology, you're ignoring the architecture of the work itself. The returns sit in task-level redesign, not role-level rollout.
Read it: "Why Are There Still So Many Jobs?"
V. How Does Automation Actually Behave?
Four works from human factors engineering, cognitive science, and information systems on what actually happens when automated systems interact with human operators.
13. Lisanne Bainbridge — Ironies of Automation (1983)
Bainbridge identified a paradox: the more reliable you make an automated system, the worse the human operator becomes at intervening when the system fails. The designer automates because humans are unreliable, then leaves the human responsible for the situations the automation can't handle — which are by definition the hardest situations.
What this means now: Replace "industrial process operator" with "knowledge worker using AI" and the ironies hold. Dell'Acqua's finding that consultants over-relied on AI outside its frontier is Bainbridge's irony, restated for 2023.
Read it: "Ironies of Automation"
14. Parasuraman & Riley — Use, Misuse, Disuse, Abuse (1997)
They developed a taxonomy of the four ways humans relate to automated systems. Use, misuse (over-reliance), disuse (under-reliance), and abuse (deployment without regard for consequences on human performance).
What this means now: Every pattern in the current AI adoption literature maps onto this taxonomy. The taxonomy gives you a diagnostic vocabulary more precise than "change resistance" or "adoption challenges."
Read it: "Humans and Automation" (PDF)
15. Shoshana Zuboff — In the Age of the Smart Machine (1988)
Zuboff identified that every information technology simultaneously automates (replaces human effort) and informates (generates new data about the process). Most organizations capture the automation value and ignore the informating value, because the informating side requires redistributing authority.
What this means now: Most organizations are capturing the automation value and leaving the informating value on the table, for exactly the reason Zuboff predicted: the informating side requires redistributing interpretive authority, and that redistribution is politically uncomfortable.
Read it: In the Age of the Smart Machine
16. Melanie Mitchell — Artificial Intelligence: A Guide for Thinking Humans (2019)
Mitchell wrote one of the strongest available accounts of what AI systems actually do, how they learn, where they succeed, and why they fail. Her central argument: AI systems are powerful pattern-matchers that lack the conceptual understanding humans bring to the same tasks.
What this means now: This is the book that gives a non-technical leader the mental model needed to reason about AI's limitations from the technology side, not just the organizational side.
Read it: Artificial Intelligence: A Guide for Thinking Humans
VI. How Do Organizations Actually Change?
The section most AI strategies skip. These works explain how organizations revise their own operating assumptions — and what happens when they can't.
17. Eric Trist & Ken Bamforth — Sociotechnical Systems (1951)
Trist and Bamforth studied coal miners reorganized from autonomous teams into a mechanized workflow. The technology was superior. Productivity collapsed. Their explanation: you cannot optimize the technical system independently of the social system.
What this means now: Organizations are repeating this lesson with AI. They are introducing technically superior tools into work systems whose social organization was designed for a different set of technical constraints.
Read it: "Some Social and Psychological Consequences of the Longwall Method"
18. Chris Argyris & Donald Schön — Double-Loop Learning (1978)
Argyris and Schön drew a line between correcting errors within existing assumptions (single-loop) and questioning the assumptions themselves (double-loop). Most organizations are competent at the first and terrible at the second.
What this means now: Single-loop learning is what most AI deployments produce. Double-loop learning would ask: now that AI can produce this report in minutes, should the report exist at all?
Read it: Organizational Learning
19. Edgar Schein — Organizational Culture (1985 / 2010)
Schein defined culture as operating on three levels: artifacts (visible), espoused values (claimed), and underlying assumptions (actual). The assumptions are the real culture, and they're largely invisible to the people who carry them.
What this means now: Most AI change management operates at the artifacts level and the espoused values level. The underlying assumptions go untouched. Until they shift, adoption will be shallow and the organization will capture approximately none of the value.
Read it: Organizational Culture and Leadership, 5th ed.
VII. The Bridge to Now
Two works that connect the foundational research to the AI era.
20. Sebastian Raisch & Sebastian Krakowski — The Automation–Augmentation Paradox (2021)
Raisch and Krakowski argued that automation and augmentation are paradoxically interdependent. Automate too aggressively and you erode the human expertise that makes augmentation viable. Organizations that treat this as a sequence enter vicious cycles.
What this means now: This paper reframes the central question of AI strategy. Not "which tasks should we automate?" but "how do we manage the ongoing tension between automation and augmentation without losing the human capabilities we'll need when the technology shifts again?"
Read it: "The Automation–Augmentation Paradox"
21. Thomas Malone — Superminds (2018)
Malone proposed that the most consequential applications of AI will not replace individual thinking but restructure collective thinking. His contribution was reframing the design question: not "What can AI do?" but "How should we organize humans and machines to think together better than either could alone?"
What this means now: Most organizations are working the wrong question. The unit of redesign is not the task. It is the system of tasks.
Read it: Superminds
The Composite Argument
Read together, these twenty-one works converge on a single argument that most AI strategies ignore.
Organizations are cognitive architectures. Their intelligence is determined not by the capability of individual nodes — human or artificial — but by the structure of connections between them: how attention is directed, how knowledge is distributed and retrieved, how diverse perspectives get integrated, how meaning is constructed under ambiguity, how humans and automated systems interact in practice, and how the social and technical systems evolve together.
AI changes all of these at once. Deploying it without redesigning the architecture is like upgrading an engine without touching the drivetrain. Power goes up. The wheels don't turn faster.
The organizations that will pull disproportionate value from AI are not the ones with the best models. They are the ones that redesign how intelligence flows between every node in the system — human and machine — so that the output of the whole exceeds what any component could produce on its own.
That is not a technology problem. It is an organizational design problem — one that requires understanding both sides of the system. And it demands the kind of thinking that these twenty-one works, taken together, make possible.