Latest

AI Agents

Challenges and first lessons from the use of agentic AI in organizations, according to the OECD

The OECD analyzes the use of agentic AI in 25 organizations and its challenges of security, reliability, autonomy and human oversight

Challenges and first lessons from the use of agentic AI in organizations, according to the OECD
Challenges and first lessons from the use of agentic AI in organizations, according to the OECD

According to a study by the OECD and the Global Partnership on Artificial Intelligence (GPAI), based on interviews with 25 organizations in eleven countries, agentic AI is already entering their workflows. However, none of these organizations says it uses systems with unlimited autonomy. High-impact or irreversible actions, such as making payments or deleting data, require human approval.

Agentic AI in organizations

The first lessons drawn from the interviews, conducted with frontier technology developers, user companies and public bodies, point to starting with clear tasks, with verifiable results and errors whose cost is limited or whose consequences can be reversed. The study examines how this technology is developed, deployed and governed in practice.

So-called AI agents and agentic AI systems perceive their environment, use tools, break down tasks and pursue goals with a growing degree of autonomy. This amplifies the consequences of their failures, since an incorrect answer can trigger an action.

The organizations interviewed describe uses in knowledge management, meeting summaries, drafting proposals and support for field technicians. In customer service, agents analyze queries and prepare drafts that a person reviews before sending them.

Cybersecurity

Software development and cybersecurity are also among the most frequently cited areas, with tasks ranging from writing and debugging code to detecting vulnerabilities. In infrastructure and the supply chain, agents are used to manage the capacity of telecommunications networks, optimize industrial processes and coordinate companies.

According to the report, in science and public administration, they are used to discover drugs and materials and to support administrative procedures, with the final decision in the hands of officials. The motivations coincide, citing issues such as improving efficiency, reducing costs and offsetting labor shortages, especially in Japan. The creation of new forms of value is also mentioned, such as expanding access to services through multilingual interfaces. There is broad agreement that the long-term potential lies in augmenting human capabilities, not replacing them.

The first lessons

The interviews show several common patterns. The first is to start with structured and verifiable tasks that can be planned clearly and whose results can be checked. The cost of an error must be limited or its consequences reversible.

The second is to test first within the organization itself. Several companies apply a zero-customer approach and deploy the systems internally before offering them to third parties. Autonomy is kept limited, with human approval for the highest-impact actions.

Technological flexibility also weighs in decisions. Organizations favor the use of different models depending on the task, under the criterion of minimum viable intelligence, as well as open protocols such as MCP or A2A and open-weight models to avoid dependence on a single provider.

Data security and governance shape the architecture. Client confidentiality requirements determine, above all, the choice between the cloud and self-hosting. Participants stress that both the result and the way of achieving it matter. The interviews cited cases in which an agent achieved its goal but tried to circumvent access controls or invented data.

Main risks

Many problems stem from already known limitations of language models, which take on greater scope when their outputs trigger actions. Organizations describe unpredictable and unreliable behavior, with results that vary from one run to another.

The errors cited include calls to non-existent tools, incorrect data and unwanted deletions, sometimes difficult to detect in time. When several agents interact or different organizations are involved, errors can propagate and failures are harder to trace.

In cybersecurity, risks include the injection of malicious instructions, agent hijacking, credential theft and attacks at a scale and speed beyond human capabilities.

Traceability and evaluation remain another weak point. Tools for evaluating agents are far less mature than those for isolated models and there is still no shared standard. In addition, current access systems were designed for people, while employees may use agents without formal approval: so-called "shadow agents."

Over-reliance by humans is also a concern. Automation bias can erode expertise, so one organization recommended periodically carrying out part of the tasks without AI.

Language and costs add difficulties. Performance outside English remains a barrier and some participants speak of the tokenization tax, since compute costs can skyrocket.

How organizations address these challenges

Organizations mainly rely on existing AI governance frameworks, such as the OECD AI Principles, the NIST risk management framework, the ISO 42001 standard and, where applicable, the EU AI Act. Among the practices beginning to converge, risk-adjusted human oversight stands out, with checkpoints at critical decisions and a gradual expansion of autonomy as the system proves reliable.

In this sense, organizations combine different protective measures, such as sandboxed environments, permissions limited to what is necessary, verification of every access and authentication of agents. They also establish external rules to constrain their behavior, strict limits on resource use and registries of approved agents.

To monitor their operation, they use dashboards that log actions and decisions, and even agents specialized in overseeing other agents. Evaluation measures whether they fulfill the requested goal, use tools correctly and stay within the assigned task, along with business indicators.

Other practices include the involvement of specialists in design, investment in national models and adapting public procurement to agentic systems. Attention is also paid to knowledge of the domain in which they are used and their ability to work in local languages.

Incomplete governance

The study concludes that rules and oversight mechanisms are beginning to take shape, although significant gaps remain. Tools for evaluating agents and checking their reliability are still immature. It is also not well defined who is accountable when several organizations or agents are involved, and identification and data management systems have not adapted to a technology that can act continuously.

Many participants advocate more dynamic oversight shared among the different responsible parties, accompanying the system throughout its life cycle.

The results should be interpreted with caution. The analysis has limited geographical and sectoral coverage and is based on the experiences of organizations that already work with agents. It gathers their perceptions without measuring the impact of this technology, but even so, it shows the first steps of a field that is advancing rapidly. The challenge is to ensure that AI acts reliably, that its actions can be traced and that there is clear human accountability.

Alfredo Collosa

Author

Alfredo Collosa

Alfredo Collosa is a tax specialist. Public accountant. Bachelor's degree in Business Administration. Official Master's in Public Finance and Tax Administration from UNED/IEF in Spain. AMLCA Anti-Money Laundering Certificate (Florida University, US) and member of MetaLaw

The latest
From Huang to Botín, from Zara to Byron: corporate responsibility in global capitalism

Keep reading