The Organisation for Economic Co-operation and Development published a working paper finding that agentic AI is moving beyond experimentation into enterprise, consumer and public sector workflows, although deployment maturity varies with technical readiness and risk tolerance. Based on interviews with 25 developers, adopters and government agencies, the paper identifies applications in enterprise productivity, software development, cybersecurity, industrial operations, infrastructure planning and scientific research. Near-term adoption remains concentrated in structured tasks with verifiable outcomes and bounded or reversible risks. Organisations are generally limiting autonomy to defined tasks and requiring human approval for consequential or irreversible actions. Emerging controls include incremental increases in autonomy, scoped permissions, sandboxing, agent authentication, audit logging, continuous monitoring and grounding in trusted data. However, evaluation and assurance remain immature, while cybersecurity, unpredictable actions, traceability and accountability across organisational and agent-to-agent boundaries present continuing challenges. The OECD also finds that existing identity, access and data governance frameworks are not well adapted to autonomous systems operating continuously across multiple tools and environments. The findings are illustrative rather than representative because they are qualitative and cover a limited range of organisations, countries and sectors. The paper identifies scope for broader research into workplace effects, interoperability, security and accountability as agentic systems become more autonomous.

Original source