The Institutionalization Stress Test: When Successful AI Agents Fail to Become Durable Organizational Capabilities

An artificial-intelligence agent may complete a realistic enterprise task, satisfy an operational reliability target under human oversight, and receive a statistically defensible deployment qualification without establishing a durable organizational capability. The missing outcome is not another property of the model alone. It is whether an organization can repeatedly perform, govern, reconstruct, improve, and legitimately transfer the work after the personnel, vendor conditions, oversight intensity, workflow configuration, and assumptions supporting the original success have changed. This paper proposes the Institutionalization Stress Test (IST), a perturbation-based qualification protocol that distinguishes successful execution, qualified deployment, and institutionalized capability. It defines an organizational capability episode, a counterfactual capability boundary, a time-indexed capability trajectory, a structured Capability Creation Receipt, noncompensatory survivability gates, and a recovery-provenance trial separating governed restoration from exceptional rescue and capability recreation, and an integrated prospective decision procedure with a temporal evidence firewall. The proposal is positioned alongside enterprise-agent benchmarks, oversight-aware qualification, evaluation-design reliability, organizational routines, knowledge retention, fault-injection testing, absorptive capacity, dynamic capabilities, and AI management systems. Rather than treating an agent's task score as evidence that the enterprise has acquired a new ability, IST asks which essential dependencies are carried by organizational arrangements and which remain concentrated in exceptional people, hidden configuration, uncontracted external services, or temporary interventions. A falsifiable study design compares the incremental predictive value of IST against strong technical, operational, and organizational-readiness baselines under matched resources. No deployment study, experimental outcome, causal benefit, validated index, or new general theorem is claimed. The contribution is a bounded research protocol for investigating when an apparently successful AI deployment becomes a capability that the institution can legitimately retain.