Beyond Chatbots: Agentic AI Takes Action as Businesses and Governments Face New Risks

Agentic AI is moving from experimentation into real business and government workflows, offering potential productivity gains, workforce support and new opportunities in public services, industry and research. The OECD warns that scaling these systems will require stronger cybersecurity, human oversight, accountability, local-language capabilities and governance before organisations grant AI greater autonomy.

Agentic artificial intelligence is beginning to move beyond chatbots and content generation into systems that can plan tasks, use digital tools and take actions with limited human supervision. A study by the Organisation for Economic Co-operation and Development (OECD), its AI and Emerging Digital Technologies Division, and the Tokyo Centre of the Global Partnership on Artificial Intelligence (GPAI) Expert Community finds that organisations are already testing these systems across business operations, public administration, telecommunications, cybersecurity, software development and scientific research.

The findings are based on interviews with 25 organisations, selected after researchers approached 74 organisations. Participants covered 11 countries, Brazil, Canada, France, Germany, India, Japan, Kenya, Mexico, Singapore, the United Kingdom and the United States, along with a global industry association. The sample included technology developers, businesses, governments, academic institutions and industry bodies. The OECD stresses that the study is qualitative and does not measure economy-wide adoption or productivity gains.

From AI Assistants to Systems That Can Act

Agentic AI represents an important change because these systems can do more than recommend what humans should do. They can potentially execute parts of a workflow themselves. Multiple AI agents can also divide complicated assignments, communicate with each other and pursue objectives over time.

Early applications include meeting summaries, proposal preparation, knowledge management, data processing, customer support, compliance and legal research. More advanced systems are being tested for supply-chain optimisation, industrial operations, telecommunications networks, software development and cybersecurity.

Software agents, for example, can write, test, review and debug code. Telecommunications companies are exploring agents for monitoring networks, forecasting capacity requirements and addressing operational problems. Researchers are also examining AI agents for scientific literature reviews, hypothesis generation, experiment design, drug and materials discovery and automated laboratories.

For private companies, the opportunity lies in moving from automating individual tasks to redesigning complete business processes. This could reduce administrative workloads and allow employees to concentrate on activities requiring judgement, creativity and human interaction.

Productivity Gains Meet Workforce and Development Challenges

Labour shortages provide another reason for adoption. Organisations in Japan highlighted agentic AI as a possible response to demographic pressures and workforce constraints. Similar technologies could eventually help governments and businesses maintain services when qualified workers are difficult to recruit.

For developing economies, multilingual AI could potentially expand access to public information, technical expertise and digital services. Smaller companies could also gain access to sophisticated analytical and operational capabilities that were previously affordable mainly to larger enterprises.

But language remains a significant weakness. Organisations reported that performance can decline outside English, particularly when agents must understand domestic laws, administrative processes and cultural contexts. Some languages also require more processing tokens, increasing computing costs.

International development partners therefore have an important potential role in supporting local-language datasets, evaluation systems, digital infrastructure, workforce skills and institutional capacity. Without such investments, agentic AI could widen rather than reduce existing digital divides.

Governments Face a New Accountability Test

Public institutions are taking a more cautious approach. Government applications identified in the research include coding assistance, email processing, document review and administrative workflows. Agents can check whether applications are complete, identify relevant legal provisions and prepare draft assessments, but final decisions generally remain with officials.

Significantly, none of the 25 interviewed organisations reported unrestricted AI autonomy. Human approval remains particularly important for consequential or irreversible actions such as payments and data deletion.

This provides an important policy signal. Governments may need procurement rules that examine not only AI model performance but also what information an agent can access, which tools it can operate, what permissions it receives and how every important action can be traced.

Cybersecurity adds urgency. Risks identified by organisations include prompt injection, agent hijacking, credential theft, sensitive-data extraction and unauthorised transactions. Multi-agent systems create another challenge because an error originating in one agent can spread through interconnected systems.

Building Guardrails Before Scaling Autonomy

Organisations are responding with sandbox environments, restricted permissions, human approval checkpoints, agent authentication, audit logs, trusted data sources and continuous monitoring. Existing frameworks including the OECD AI Principles, NIST AI Risk Management Framework, ISO 42001 and the EU AI Act are also influencing governance approaches.

Cost controls deserve greater attention. Two interviewed organisations reported cost overruns, showing how continuously operating agents and repeated computational loops can create unexpected expenditure. Companies also need to address automation bias and "cognitive debt," where excessive dependence on AI gradually weakens employees' expertise.

For policymakers, the priority should be creating clear rules for accountability, cybersecurity, identity management, interoperability and traceability. Development partners can support governments in building regulatory capacity, local-language AI infrastructure and independent evaluation systems. Businesses should introduce autonomy gradually, beginning with structured and reversible tasks before expanding into higher-risk operations.

The OECD does not provide numerical estimates for GDP, employment or productivity gains, meaning such benefits should not yet be assumed. Instead, its findings offer an early picture of a technology moving rapidly toward operational use. The development opportunity is substantial, but the central challenge will be ensuring that governance, skills and security advance alongside the machines being given greater freedom to act.