Who Is Responsible When AI Agents Make Decisions

Summary

This Tomorrow Today interview has Shekhar Natarajan question Matt Mills of SailPoint about governing human and non-human access as AI agents become autonomous enterprise actors. Mills argues that discovered machine identities should have accountable owners, that authorization should respond to context in real time, and that transparent records and demonstrated performance are prerequisites for trust. The episode also exposes a limit in this model: assigning a human owner creates an accountability anchor without guaranteeing that the owner can predict or explain a probabilistic agent’s behavior.

Key Claims

• Agent Identity Governance extends employee account management to bots, robotic process automation, service identities, and AI agents that can reach applications, entitlements, and data.
• SailPoint’s stated approach is to discover non-human identities, categorize them, and attach each one to an accountable human owner; Entro is presented as the acquired discovery layer supporting that work.
• Orphaned machine identities are a governance risk because authority can persist without a person clearly responsible for reviewing, narrowing, or revoking it.
• Contextual Agent Authorization evaluates time, location, device, requested resource, and immediate risk rather than relying only on static account permissions.
• Mills distinguishes access governance from control over an AI system’s output quality: authorization can constrain what an agent may reach or do without making probabilistic behavior deterministic.
• Human Accountability For AI Agents is presented as the current practical anchor because an agent cannot meaningfully be punished, while the person authorizing or certifying it can be held responsible.
• The host’s challenge qualifies that anchor: formal ownership can exceed practical control when behavior emerges across models, data, developers, deployers, authorizers, and spawned agents.
• Policy, risk, and intent engines are expected to support real-time authorization as enterprises manage thousands of agents and manual review becomes infeasible.
• Trust is described as an accumulated outcome: rules and transparency establish a basis for evaluation, repeated successful action builds credibility, and consistent performance over time earns trust.
• AI-supported governance decisions require auditability and consistency, but adding AI to the policy layer also places another consequential probabilistic system inside the control chain.

Key Quotes

“every agent should have a human in the loop” - Mills’s current accountability position as summarized in the episode.

“an agent itself cannot meaningfully be punished” - the episode’s rationale for retaining a human authorization anchor.

Connections

• Matt Mills, SailPoint, and Entro - guest, identity-governance company, and non-human identity discovery acquisition.
• Tomorrow Today and Shekhar Natarajan - show and host context.
• Agent Identity Governance, Agent Identity And Authentication, and Agent Permission Boundaries - identity, attribution, and scoped-authority layer.
• Contextual Agent Authorization and Enterprise Agent Governance - dynamic access and production-governance layer.
• Human Accountability For AI Agents, Human Judgment Under AI, and Probabilistic Software - responsibility and control-limit branch.
• AI Governance And Compliance - regulation, transparency, ethics, and audit context.

Contradictions

• No settled factual contradiction was found.
• The episode creates a productive tension inside Human Accountability For AI Agents: assigning an owner makes responsibility legible, but it does not prove that the owner had enough information, authority, time, or technical control to prevent an emergent action.
• SailPoint product capabilities, customer scale, revenue, breach percentages, and suite outcomes are vendor claims in a structured episode summary, not independently audited evidence.