From Autonomy to Initiative: Enterprise AI's Real Endgame

Abstract

Current agentic AI discourse often treats autonomous execution as the apex of maturity — agents that operate independently within assigned parameters. We argue this is the wrong endpoint for enterprise AI in high-consequence environments. Autonomy is independence in how to work while remaining dependent on humans for what to work on. Initiative — the capacity to identify and prioritize action opportunities aligned with organisational purpose through immersion rather than instruction — is a categorically different capacity, and the one that determines whether enterprise AI generates value or merely generates output.

We define organizational intelligence as a system property: the capacity to perceive institutionally relevant patterns, infer their consequences, and select action opportunities under changing constraints. Governed knowledge is not intelligence itself — it is the substrate on which organizational intelligence operates. On this basis, we propose three conditions for governed initiative — initiative that is auditable, correctable, and institutionally legible: fertile pattern density above a perceptual threshold (Condition 1), a constraint-legible action space where governance rationale is causally transparent (Condition 2), and progressive governance relocation as the institutional world model deepens (Condition 3). We examine the professional consulting pyramid as a mature institutional analogue — one of several graduated immersion systems, including medical residency and military officer development, that satisfy all three conditions through accumulated exposure, legible constraints, and graduated relaxation of control as judgment develops.

For AI agents, each condition transforms in implementation but the functional requirements are identical. The central engineering claim is a three-step governance relocation mechanism: as the institutional world model deepens, governance enforcement migrates from explicit pre-action constraint checking (early lifecycle) through substrate-resident causal structure (mid lifecycle) to risk-based monitoring and audit sampling (mature lifecycle). We identify a substrate gap in the current literature — world models without governance and governance without world models each satisfy only a subset of the conditions — and propose six architectural enrichments to the Governed Intelligence Lifecycle (Paper D) that close it. We introduce the domain graph as the missing middle layer between the foundation model and application context: a governed, reusable, compounding substrate for domain-specific organizational intelligence.

Active inference (Friston, 2010) provides a normative model for describing why agents operating on deep, governed institutional world models may move from autonomous task execution toward institutionally legible initiative. We address the strongest counterargument — that scale alone may produce initiative without governed architecture — by scoping the claim: in regulated industries and enterprise deployments where agent actions carry material consequences, initiative must be auditable, correctable, and governable. Scale may produce initiative-like behaviour in low-consequence settings; governed architecture is necessary where initiative must be institutionally legible.

This is the capstone of a five-paper programme: Paper A (Reichhart and Gelas, 2026a) diagnoses dynamics blindness as the failure mechanism; Paper B (Gelas and Reichhart, 2026b) specifies the architectural resolution; Paper C (Reichhart and Gelas, 2026c) provides theoretical foundations from ten independent traditions; Paper D (Gelas and Reichhart, 2026d) operationalizes the governed intelligence lifecycle with epistemic immunity; this paper identifies the conditions under which that architecture produces governed initiative — agents that perceive what matters through immersion, not instruction.

Keywords: initiative, autonomy, enterprise AI, active inference, governed world models, knowledge graphs, organisational intelligence, governance relocation, domain graph, regulated industries