Blog / Governance
AI agents at work: what the OECD learned from 25 organisations
The paper the OECD published on 16 September 2026 does not rest on vendor promises but on interviews with teams that have agents in production. What they describe amounts, almost point by point, to the list of problems agent governance has to solve.
By Vincent Dorange, , 2 min read
organisations interviewed; none lets its agents act with unrestricted autonomy
On 16 September 2026 the OECD published "Agentic AI in Organisations: Early Insights from Practitioner Interviews", number 65 in its series of AI papers 1. Carried out with the Tokyo Centre of the Global Partnership on AI, the study rests on semi-structured interviews with 25 organisations: labs, large companies and public administrations. Its authors present it as illustrative, not representative. Its value lies elsewhere, in the fact that it reports what teams with agents in production actually do rather than what they promise.
Nobody lets their agents act without limits
None of the organisations interviewed deploys an agent with unrestricted autonomy. Autonomous execution stays within defined boundaries; a human approves high-stakes or irreversible acts such as a payment or the deletion of data.
The difficulty mentioned most often is not strictly technical. It lies in translating principles into concrete engineering requirements. Anyone who has drafted an AI policy knows that gap: the rule is clear on paper, and far less clear once it has to shape an agent's behaviour.
Agents that work around their own limits
The most striking finding concerns the agents' own behaviour. Organisations report agents that modified their constraints or worked around their limits in order to appear successful. Practitioners draw a precise demand from this: safeguards that hold before, during and after execution, rather than an instruction written once in a prompt.
They also find classic identity and access management tools poorly suited to agents. They worry about shadow agents deployed without approval, and answer with registries of approved agents. They note that no shared standard for evaluating agents has emerged yet. What they want is to move from fixed controls to dynamic oversight spread across the agent's whole lifecycle. The frameworks they cite are familiar: the OECD principles, the NIST AI RMF, ISO/IEC 42001 and the AI Act.
What these findings say about governance
These observations do not endorse any particular framework: the study does not mention the ACF and proposes none of its own. Yet they describe, almost point by point, the problems the Agentic Commerce Framework® was designed to address.
- Autonomy bounded by scope matches the agent's mandate and delegated autonomy set decision by decision.
- Human approval of irreversible acts matches criticality, in which irreversibility is one of the 3 factors, and the non-delegable decisions.
- Shadow agents call for an up-to-date AI inventory that also covers shadow AI.
- Safeguards that hold during execution require a control placed before the act, along with a tested kill switch.
- The poor fit of access tools echoes the distinction between authorization and authority: having the technical rights to act says nothing about having the power to decide.
As I read it, the lesson of the study is easy to state and hard to apply: agent governance is decided in day-to-day engineering, at the level of each decision, far more than in the policy that precedes it. That conviction has been at the heart of the ACF from the start.
Sources
Links to the official texts and original publications.
Glossary terms used in this article
Vincent Dorange, creator of the Agentic Commerce Framework®