2026 Reports

AI Power at Tempo: Conversion, Harnesses, and Anticipatory Influence

Tumin, Zachary; Edelmann, Rasmus

As artificial intelligence moves from tool to infrastructure, the locus of organizational power shifts. Power no longer resides primarily in decisions made at the moment of choice. It resides in the systems that determine what is seen, what is prioritized, where work flows, and what happens by default before any formal decision is made. This paper develops a framework for understanding and governing that shift.

The framework advances four propositions. First, AI becomes power only when converted into outcomes. Technical capability alone is insufficient to produce organizational value. Conversion requires a clear vision of value and objectives, the essential capacities to execute, and the authority and legitimacy to act. Where these are absent, AI systems falter even at rollout. Second, AI's distinctive mechanism of influence is anticipatory. AI systems shape the decision environment upstream, through ranking, routing, defaults, thresholds, and queues, often designed and deployed outside any formal policy review, yet forcefully constraining the choices available to people downstream. This is anticipatory influence: the structuring of action before choice, framing decisions before deliberation takes place and frequently outside visibility. Third, conversion is achieved through harnesses (the surrounding architecture through which institutional purpose becomes machine-executable action), spanning data pipelines, workflows, decision rights, human oversight, and feedback loops. Models generate signals; harnesses determine what happens. Fourth, management becomes infrastructural and essential. As AI systems act upstream, managerial work shifts from supervising tasks and evaluating outputs to designing and governing the upstream conditions under which systems act, before deliberation begins and often before consequences are visible.

The paper develops these propositions through cases drawn from enterprise AI, healthcare, policing, social services, and critical infrastructure. Cases of failure (including Klarna's miscalibrated customer service automation, the Optum health algorithm's encoding of racial bias, the NYPD's Patternizr deployment, and the 2023 East Palestine rail derailment) illustrate how weak harnesses produce conversion failure: systems that sense or predict accurately but cannot translate signal into authorized, appropriate action. Cases of success (including Spotify's harness-centered engineering management and the Viz.ai stroke detection deployment at UC Davis Health) illustrate how well-designed harnesses enable reliable, scalable performance. Causal evidence from a 2026 randomized controlled trial of 515 firms confirms that the binding constraint on AI value is not access to the technology but the managerial capacity to discover where across the production process conversion is possible.

The paper draws on classic theories of power (Dahl, Bachrach and Baratz, and Lukes) to situate AI as a new face of power, and on management theory (including Moore, Davenport, Simons, Meadows, and Christensen) to identify the institutional conditions under which anticipatory influence becomes governable. This paper asks what kind of power is being exercised when AI systems act, and what governing that power actually requires, questions the practitioner literature on AI deployment planning has largely left unaddressed. It identifies eight conditions of governable AI power, spanning epistemic, organizational, economic, normative, cognitive, technical, legal, and orchestrative dimensions. It concludes with five managerial disciplines (specify, instrument, assign, contest, and learn) that define the practice of harness-centered management. The central claim is this: harnesses create impact, governance creates trust, and values create legitimacy. The model has become table stakes. What surrounds it has become the prize. Managers, as leaders, are responsible for them all.

This issues paper is the second in a series developed in connection with research for the book, "AI: The New Face of Power," under contract with Columbia University Press (forthcoming 2027).

Keywords: artificial intelligence, anticipatory influence, AI governance, organizational governance, harness, conversion, agentic AI, algorithmic power, infrastructure, managerial authority