Why Focus on AI's Institutional Impacts When Thinking Strategically?

A hospital administrator in Toronto told me she spent two years evaluating AI diagnostic tools. Feature comparisons, benchmark scores, vendor demos. Then the tool got deployed, and within six months, three senior radiologists had quietly negotiated exits, one department had restructured its credentialing requirements, and the insurance billing workflow had been redesigned from scratch - not because anyone planned it, but because the technology forced the question of who decides what.

She had studied the AI. She had not studied what the AI would do to the institution around it.

Strategic thinking about AI fails at exactly this point. We track capabilities. We track benchmarks. We treat AI as a product decision when the more consequential question is almost always structural - who gains authority, who loses it, what processes become legible or illegible, and how power redistributes inside and between organizations. The reason to focus on AI's institutional impacts when thinking strategically is simple: technology doesn't disrupt industries. It disrupts the institutional arrangements that industries run on. Get that distinction wrong and your strategy is optimizing for a reality that's already being replaced.

The Layer Beneath the Tool

Institutions are older than most of us think about. Sociologist W. Richard Scott, in his foundational work Institutions and Organizations (Scott, 2001), defines institutions as the regulative, normative, and cognitive rules that structure human behavior - the written rules, yes, but also the unwritten assumptions about who counts as an expert, what counts as a valid decision, and who bears accountability when something goes wrong.

AI doesn't just automate tasks inside that structure. It destabilizes the assumptions the structure was built on.

Take credentialing. A licensed physician carries institutional authority because the credential signals training, liability, and accountability. When an AI system outperforms that physician on specific diagnostic tasks - and in certain domains it does - the question the institution faces isn't "should we use this tool?" The question becomes "what does a physician's authority now rest on?" That's an institutional question. A landmark demonstration of this pressure came from Google's DeepMind in 2016, when researchers published results in Nature showing their AI system detected diabetic retinopathy with 94.5% accuracy against a panel of ophthalmologists (Gulshan et al., 2016). The clinical performance was striking; the institutional questions it raised - about scope of practice, liability, and credentialing - proved far harder to resolve than the technical ones. Organizations that treated that paper as a procurement question got surprised. Strategic thinking that ignores this layer is like studying a game by reading the rules and ignoring who controls the referee.

Power Doesn't Disappear. It Moves.

There's a useful framework from political scientist Langdon Winner, articulated in his essay "Do Artifacts Have Politics?" (Winner, 1980) - the argument that technologies are not neutral, that they embed particular arrangements of power and authority. Winner's canonical example was Robert Moses's bridges in New York, allegedly built too low to allow buses (and therefore poorer residents) to reach Jones Beach. Whether the specific history holds up to scrutiny is debated; the structural insight does not.

AI systems encode decisions about what data matters, whose judgment was correct in the training set, and which outcomes count as success. Those encoding decisions are institutional decisions - made by specific people, in specific organizational contexts, with specific incentive structures. When an AI system gets deployed at scale, those embedded choices about authority travel with it.

This matters strategically because the organization adopting AI doesn't get a neutral tool. It gets a tool that has already made certain institutional choices - and those choices will interact with the existing institutional arrangements in ways that can reinforce, or radically unsettle, the existing power structure.

In practice, this means middle management is often the most disrupted layer - not because their tasks are most automated, but because AI systems often bypass the coordination and gatekeeping functions that gave middle management its authority. A 2023 working paper from the National Bureau of Economic Research by Brynjolfsson, Li, and Raymond, studying a customer service AI deployment, found that the workers who benefited most were junior employees, while experienced senior workers saw smaller (and sometimes negative) productivity effects (Brynjolfsson, Li & Raymond, 2023). The institution's internal knowledge hierarchy got partially inverted. That inversion doesn't show up in task-level automation analysis. It only becomes visible when you're asking institutional questions.

Why Individual-Level Analysis Misses This

Most AI strategy advice centers on the individual. Which skills to build. Which tasks will be automated. How to stay relevant. And that advice isn't wrong, exactly - it's just operating at the wrong level of abstraction for most strategic questions.

Institutions have inertia. They have coordination mechanisms, enforcement structures, and legitimacy requirements that individual actors can't simply bypass. A lawyer who becomes extremely capable at using AI legal research tools still practices inside a bar association that sets standards, a firm that sets billing structures, a court system that sets procedural rules, and a client relationship that carries liability arrangements. Each of those layers will adapt to AI on its own timeline, driven by its own institutional logic.

Strategic planning that doesn't account for this ends up in a particular failure mode - what I'd call capability-structure mismatch. You have a highly capable individual or team, but the institutional structure around them either doesn't support the new workflow or actively works against it. The technology works. The organization can't absorb it.

This is, incidentally, why so many AI pilots succeed and so few AI deployments scale. Pilots can be run around institutional constraints. Deployment runs into them.

The sociologist Andrew Abbott mapped out the logic of this kind of jurisdictional conflict in The System of Professions (Abbott, 1988), arguing that professions don't compete on capability alone - they compete for legitimate jurisdiction, the recognized right to perform certain kinds of work. AI doesn't just improve performance inside existing jurisdictions. It forces renegotiation of which jurisdictions exist at all. That's a slow, political process, and it happens at the institutional level, not the individual level.

The Regulatory Lag Problem - and When Institutions Strike Back

Regulatory institutions are a special case. They typically operate on decade-scale timescales, reacting to harms after they've accumulated enough political salience to force legislative attention. The EU AI Act, which began formal drafting in 2021 and faced successive revision cycles before implementation, illustrates this - the regulatory structure was designed partly around systems that were already being superseded as the text was finalized.

This creates a genuine strategic opportunity for organizations willing to move during the lag window. But it also creates a trap. Organizations that build strategies around that lag - assuming they'll have permanent regulatory freedom - tend to get caught when the institutional correction eventually arrives. And it does arrive, usually faster than expected once it starts.

The edge case worth sitting with: some institutions are too sclerotic to adapt even when AI genuinely improves on what they do. Academic publishing is a possible example - the peer review process has known structural failures, and AI assistance could plausibly improve both scale and consistency of review. But the institutional logic of academic prestige, which depends on scarcity and exclusivity, creates resistance that has nothing to do with the quality argument. In those cases, the institutional inertia may be permanent - or near enough that it changes the strategic calculus entirely.

Who this analysis doesn't apply to, clearly, is anyone operating in a context without institutional structure - independent consultants, early-stage startups in unregulated spaces, individual creators. For those actors, the individual-level AI strategy advice is mostly right, because the institutional layer is either absent or not yet formed. The institutional focus kicks in once you're operating inside structures that carry their own authority, accountability, and inertia.

Limitations

This framing doesn't solve everything, and I want to be honest about that. Focusing on institutional impacts gives you a better map of where disruption will actually land - but it doesn't give you a decision procedure. Two organizations could correctly identify that AI will restructure their credentialing hierarchy and make opposite strategic bets about how to respond, both reasonably.

The evidence on how specific institutions will adapt is also genuinely thin. We have useful case studies, theoretical frameworks, and early empirical data from deployments like the Brynjolfsson et al. customer service study. We do not have rigorous longitudinal data on how institutions across different sectors have actually reconfigured authority and accountability in response to AI adoption at scale. Most of that data doesn't exist yet - we are observing the process in real time, which means the frameworks here are working hypotheses, not established findings.

The institutional lens gives you the right questions. It cannot give you certain answers about timing, severity, or direction of the institutional changes you're navigating.