OECD Releases Report on Agentic AI Adoption Status

The OECD has published a report titled 'Agentic AI in organisations' based on interviews with practitioners from 25 organizations. The study found that no organization granted unrestricted autonomy; instead, all applied a phased oversight system requiring human approval for high-risk tasks.

Key Use Cases and Deployment Lessons

• Key Areas: Software development (code, testing) and cybersecurity (vulnerability identification) are central. In customer support, a human-on-the-loop approach is common, where agents draft responses and human agents perform final reviews.
• Deployment Strategy: The 'Customer Zero' strategy, involving internal deployment before external release, and the 'minimum viable intelligence' approach, using cheaper and faster models for repetitive tasks, are preferred.
• Autonomy Management: Irreversible actions such as payments or deletions must undergo human approval, and autonomy levels are managed across five tiers.

Adoption Barriers and Risk Factors

• Technical Limitations: Maintaining consistency in long interactions, misaction, and error propagation between multi-agents were identified as major risks.
• Cost and Evaluation: Predicting computing costs due to repetitive triggers is difficult, and some organizations have experienced cost overruns. Unlike standalone LLM evaluations, standardized evaluation metrics for agent systems remain immature.
• Language Barriers: Performance gaps in non-English languages and increased computational costs due to the 'tokenization tax' hinder regulatory compliance and adoption.

Mitigation Strategies and Governance

• Technical Safeguards: Layered controls are implemented through sandbox isolation, the principle of least privilege, agent registries, and programmatic tool calling.
• Governance: While leveraging existing frameworks such as the OECD AI Principles and the NIST AI RMF, a shift from static controls to dynamic/distributed oversight is necessary. Establishing separate identities for agents and clear accountability systems remains a future challenge.