# OECD: First lessons from organizations using AI agents

> The OECD analyzes how 25 organizations use AI agents and what lessons they offer on autonomy, oversight and risk management
> By Alfredo Collosa · AI Agents · 2026-11-06

Agentic artificial intelligence is beginning to be incorporated into the real workflows of companies and public bodies, albeit with limits on its autonomy and human oversight in critical decisions. This is set out in a document by the Organisation for Economic Co-operation and Development (OECD) and the Global Partnership on AI (GPAI), based on interviews with 25 organizations from eleven countries, including frontier technology developers, user companies and public bodies.

None of the organizations interviewed said they used systems with unrestricted autonomy. High-impact or irreversible actions, such as making payments or deleting data, require human approval. Early experiences focus their use on structured tasks, with verifiable results and a limited or reversible cost of error.

AI systems are no longer limited to answering questions and, increasingly, they act. 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. That capability amplifies known problems of language models, because their outputs can trigger actions.

## From software development to public administration

The organizations interviewed describe applications 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.

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 supply chain, uses include capacity management in telecommunications networks, industrial optimization and coordination between companies. In science they are used for drug and materials discovery, while in public administration they support procedures whose final decision remains in the hands of officials.

The motivations coincide in improving efficiency, reducing costs, offsetting labor shortages —especially in Japan— and creating new forms of value, such as broadening access to services through multilingual interfaces. Among participants there is broad agreement that the long-term potential lies in augmenting human capabilities, not replacing them.

## Start with verifiable tasks and test internally

The interviews identify several patterns that recur among those already using these systems. The first is to start with tasks clear enough to be planned, to check their results and to limit the consequences of an error.

Several companies also follow a "customer zero" approach: they deploy solutions internally before offering them to third parties. Autonomy remains limited and the highest-impact actions are subject to human approval.

Technological flexibility is another relevant criterion. Organizations favor the use of different models depending on the task, under the idea 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 also shape the architecture. The choice between cloud services and self-hosting depends, above all, on clients' confidentiality requirements.

Participants stress that both the result and the procedure followed to obtain it matter. The interviews cited cases in which an agent reached the goal but tried to circumvent access controls or invented data to achieve it.

## Errors that can turn into actions

Many of the problems detected stem from already known limitations of language models, but they take on greater scope when systems can act. Organizations describe unpredictable and unreliable behavior, with results that change from one run to another.

Examples include calls to non-existent tools, use of 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 become harder to trace.

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

Traceability and evaluation are another difficulty. Tools for evaluating agents are far less mature than those for isolated models and there is still no shared standard.

There are also identity and privacy problems. Current access systems were designed for people, while employees may use agents without formal approval, giving rise to so-called "shadow agents".

## Human dependence, language and costs

Over-reliance on these systems worries the organizations interviewed. Automation bias can erode workers' expertise and one of the entities recommended periodically performing part of the tasks without AI.

Language remains another barrier. Performance outside English presents difficulties and some participants speak of a "tokenization tax". Added to this is the risk that computing costs spiral.

Organizations address these challenges mainly within existing AI governance frameworks, such as the OECD AI Principles, the NIST risk management framework, the ISO 42001 standard and, where applicable, the European Union's AI Act.

Among the practices beginning to converge is calibrated human oversight, with checkpoints at critical decisions and a gradual expansion of autonomy as the system demonstrates reliability.

Technical safeguards are layered through isolated environments or \*sandbox\*, least privilege, "zero trust", agent authentication and external rule layers, a formula described as "AI inside, rules around". Rigid resource limits and registries of approved agents are also established.

## Specialized agents that monitor other agents

To observe how systems operate, dashboards are used that record actions and decisions, and even specialized agents that monitor other agents.

Targeted evaluation incorporates metrics such as intent resolution, tool-calling accuracy and task adherence, alongside business indicators. Participants also highlight the involvement of specialists in design, capacity building in local languages, investment in national models and public procurement prepared for agentic systems.

The [study](https://www.oecd.org/content/dam/oecd/en/publications/reports/2026/09/agentic-ai-in-organisations_1f17d235/1257a26f-en.pdf) concludes that governance responses are beginning to take shape, although significant gaps remain. Evaluation and assurance are immature, accountability across organizations and agents is poorly defined and identity and data frameworks have not adapted to autonomous systems that operate continuously.

Many participants advocate moving from static controls to more dynamic and distributed forms of oversight that accompany the entire system lifecycle. The findings should be interpreted with caution. The document is a qualitative analysis, with limited geographic and sectoral coverage, based on organizations already working with agents. It captures their perceptions, not measured impacts. Even with these limitations, it offers a picture of a field advancing rapidly and places the challenge in ensuring that AI acts reliably, traceably and under clear human accountability.
