Data provenance asks where information came from. Autonomous AI introduces a related problem: action provenance. When an agent takes a consequential action, organizations need to be able to understand not only what happened, but what caused it to happen.
An autonomous action may emerge from a complicated chain. A user establishes an objective; the system develops a plan; retrieved information changes its interpretation of the problem; memory affects its priorities; another agent provides advice; a tool returns new information; and the system eventually acts. By the time the action occurs, its origin may be distributed across multiple components and interactions.
Traditional logs may tell us that an action happened without explaining why the system considered that action appropriate. Meaningful action provenance therefore requires preserving the relationship between goals, instructions, information, reasoning, permissions, and execution. This matters for auditing and incident response, but also for accountability. An organization cannot meaningfully govern autonomous systems if it cannot reconstruct the chain through which authority became action.
As AI systems become more complex, provenance must move beyond tracking data and model versions. We will increasingly need to answer a more practical question: Why did this system do this, and what gave it the authority to do so?