OECD interviews show how early AI agents are being kept within human checkpoints
Practitioners in 25 organisations across 11 countries describe agents moving into real workflows, with autonomy limited around consequential actions. The interview study is a window into practice, not a global adoption rate.
The short version
- What happened
- The OECD/GPAI project interviewed practitioners in 25 organisations across 11 countries about agent applications and governance.
- Who it affects
- People affected by an agent's decision need a named responsible organisation, an understandable record of actions and a route to correct a consequential error.
- What remains unproven
- Interviews with 25 engaged organisations cannot establish an adoption rate, a causal productivity gain or a general failure frequency.
Keep in mind
Interviews with 25 engaged organisations cannot establish an adoption rate, a causal productivity gain or a general failure frequency.
The 24 September OECD.AI article explains the same project as the 16 September working paper; two links do not amount to independent corroboration.
Source analysis · 2 linked sources. Primary analysed source date: 16 September 2026. Latest linked supporting source: 24 September 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
What the interviews reveal
The OECD published Agentic AI in organisations on 16 September; an OECD.AI/GPAI explainer followed on 24 September. The project spoke with practitioners at 25 organisations in 11 countries, including frontier developers, enterprises, public bodies and academic institutions. Its question is practical: where are agents being used, what benefits and frictions do practitioners see, and how are they deciding when a system may act? This newsroom analysis appears on 28 September and keeps those source dates distinct. The 36-page working paper is an early qualitative evidence base, not a statistical survey of all employers.
The accompanying explanation describes uses in software development, cybersecurity, network planning, scientific work and public administration. Interviewees describe systems that can perform multi-step tasks and interact with tools, but the authors say no participating organisation reported granting unrestricted autonomy. Many use a staged approach: an agent may work until a defined checkpoint, then seek human review before a consequential or irreversible action. That is a meaningful operational detail. It does not mean all such checkpoints are effective, or that every organisation outside the interview group uses them.[1][2]
Why the guardrails matter for people
When software only drafts a recommendation, a person may still decide what to do. An agent with access to records, messages, code or payment systems can take an action that changes the world beyond its answer box. The OECD discussion points to hallucinations, incorrect tool use and behaviour that varies across runs. Multi-agent and cross-organisation workflows add another problem: responsibility for a bad outcome can become harder to trace. Staff therefore need to know which actions require approval, which can be reversed, whose credentials an agent uses and where the action log can be inspected.
Participants described adapting existing governance frameworks and combining controls rather than relying on a single safety promise. Examples include restricted access, sandbox testing, approved-agent registries, monitoring, domain expertise in design and calibrated human oversight. These are approaches reported by practitioners, not an OECD certification that any one deployment is safe. A nominal human checkpoint can also fail if a reviewer lacks context, time or authority to stop an action. A useful workplace assessment would test the checkpoint with realistic error cases, record the rate of unnecessary escalations and ask affected people how they can seek correction.[1][2]
The evidence still missing
The interview group is selected for engagement with agentic systems. It cannot supply a denominator for how many organisations in a country deploy agents, how often agents make mistakes, or whether agent users are more productive than comparable non-users. The OECD.AI authors themselves call for broader sectoral and geographic coverage and identify system-level evaluation, traceability, accountability and cybersecurity as unresolved challenges. Their discussion also notes that there is no widely accepted standard for evaluating long sequences of actions, including when an agent should request human input.
The next phase of evidence should combine interviews with measured workflow outcomes and documented incidents. Researchers would need to define an agent consistently, separate pilots from sustained deployment, observe the actions taken rather than just managers' expectations, and measure both value and harm. Public agencies, firms and workers have different stakes in the same workflow: shorter processing times may help applicants, while a wrong eligibility decision can impose a serious cost on one person. The study's most useful contribution is a concrete agenda for scrutiny—bounded authority, visible handoffs and traceable action—while the scale and net effect of agent adoption remain open questions.[1][2]