Does AI adoption require rethinking how organizations coordinate?

Agrawal, Gans and Goldfarb's 2021 NBER working paper (published in the Journal of Economics & Management Strategy, 2023) argues that "analyses of AI adoption focus on its adoption at the individual task level," while "what has received significantly less attention is how AI adoption is shaped by the fact that organisations are composed of many interacting tasks." Their formal model treats AI as "a prediction tool" layered onto one of two interdependent decisions inside a firm, and finds that "reliance on AI... increases decision variation which, in turn, raises challenges if decisions across the organisation interact."

The mechanism runs through modularity. In the model a "focal" decision D1 earns a stand-alone payoff α when it matches an external state, a "supporting" decision D2 earns an alignment payoff γ when it matches D1, and the two together earn a synergistic payoff β > α + γ if both are correct and aligned — a premium only a non-modular organization can capture. Without AI, both decision-makers default to the same predictable "focal point" action, so alignment is trivial. AI lets the first decision-maker follow a prediction and deviate from that default more often, raising its stand-alone return by m(1−ρ)α but making its action harder to anticipate, which lowers the second decision-maker's alignment return by m(1−ρ)γ. "Modularity... softens that impact" because a modular organization only risks losing γ when alignment fails, not the larger β; a non-modular one risks the bigger prize. The paper's resolution is that communication functions as "a type of organisational change": once the two decision-makers can coordinate, AI becomes "especially valuable to non-modular organisations," which can then capture β. Hence the paper's own framing: AI adoption is a task-level story only "when examining modular organisations that do not change," and a systems-level story once "the organisation can change how it operates."

This supplies a mechanism behind a gap that later, LLM-era surveys report without explaining. Why do individual AI gains not translate to business results? documents exactly the individual/organizational split this model predicts: task-level gains (the analogue of one decision-maker's α) do not automatically aggregate into firm-level return, because the interacting decisions around that task are never recoordinated. Can AI boost how teams work together? makes the same individual-to-collective move from the tooling side, arguing teams need "new tools and new norms" — a 2026 echo of this 2021 model's claim that communication, not better prediction alone, is what lets coordinated organizations benefit. Does easier tool-building actually solve enterprise adoption problems? names a related bottleneck from a different angle: ease of building a tool does not solve "getting a whole company to adopt it," the organizational-redesign problem this paper formalizes mathematically.

The model is stylized: two decision-makers, one prediction, a reduced-form notion of "correctness," and no treatment of generative or agentic AI — this is pre-LLM prediction technology, built to analyze automated forecasting tools, not today's AI agents. The paper does not test the model against real firm data or claim its parameters (α, γ, β, ρ) are measurable in practice; it is a proof, not a measurement. What it establishes is a reason, not an empirical finding: if an organization is non-modular and uncoordinated, task-level AI adoption can make outcomes worse before communication infrastructure exists to capture the larger synergistic payoff. The implication for later task-versus-system debates is that an individual/organizational productivity gap is not necessarily a change-management failure to be fixed with training — it can be the predicted equilibrium of adopting a prediction tool into one task of an otherwise uncoordinated system.