OECD interviews show how organizations are governing the first wave of agentic AI deployment

The OECD published a working paper on September 16 examining agentic artificial intelligence inside organizations.

The study draws on semi structured interviews with 25 organizations across several countries and sectors.

Participants included technology developers, enterprise adopters, government agencies, and research organizations.

ChatGPT is insanely powerful.

But most people waste 90% of its potential by using it like Google.

These 1,000+ proven ChatGPT prompts fix that and help you work 10X faster.

Sign up for Superhuman AI and get:

Claim your free prompts

The report does not provide a representative survey of organizational adoption.

It offers qualitative evidence about how early practitioners are building and governing these systems.

The interviews cover uses in enterprise productivity, supply chains, industrial operations, cybersecurity, software development, infrastructure planning, and scientific research.

The report says organizational maturity varies widely across the participants.

Some teams are still testing limited assistants, while others are connecting agents to operational workflows.

The paper records motivations reported by interviewees rather than measuring realized business outcomes.

Those motivations include operational efficiency, cost considerations, workforce constraints, and possible new forms of value creation.

The findings place governance beside deployment rather than after it.

Organizations are using incremental autonomy to limit what an agent can do without review.

They are also adding human checkpoints for actions with high consequences.

Subject matter experts participate in system design when domain context affects safe operation.

These controls recognize that an agent’s authority depends on its tools and environment.

A model may produce a reasonable answer while a connected tool performs an unsuitable action.

This guide distills 10 AI strategies from industry leaders that are transforming marketing.

Outpace your competitors by mastering AI.

Get Your Free Guide

The report identifies evaluation and assurance as continuing challenges.

Traditional model tests may not capture failures across a long task sequence.

An agent can make a series of individually plausible choices that produce a poor final outcome.

Organizations therefore need tests for planning, tool selection, recovery, and state changes.

Those tests must reflect the actual work environment rather than only an isolated model interface.

Cybersecurity is another unresolved area.

Agents may access internal systems, external services, repositories, records, or other tools.

Each connection creates a permission boundary and a possible path for misuse.

The report also raises accountability questions across organizational and agent to agent boundaries.

Responsibility becomes harder to assign when one agent delegates work to another system.

Logs need to show which system acted, what information it received, and which authority allowed the action.

Traceability matters when a result must be reviewed after an incident.

It also matters when a person needs to correct a decision without discarding the entire workflow.

The OECD describes oversight as increasingly distributed and adaptive.

That description follows from systems that change tools, inputs, and task paths during execution.

A fixed approval step may not cover every action created by a changing workflow.

Organizations may need controls that adjust to task risk and system state.

The paper’s limitations are explicit.

The sample includes organizations already engaged with agentic systems.

Its geographic and sector coverage is limited.

The interviews also reflect different levels of technical maturity.

Rapid system development means some observations may change as deployments expand.

The research does not provide quantitative performance measures or adoption rates.

Those limits narrow what can be concluded from the report.

It supports an account of early practice, not a forecast of universal organizational behavior.

For technology leaders, the practical issue is control design.

Permissions, evaluation, human review, logging, and recovery need to be considered together.

For researchers, the report offers a grounded list of questions for future measurement.

The first wave of agentic deployment is therefore a systems governance problem.

Model capability matters, but organizational authority determines what capability can change.

This research check records early evidence and leaves adoption claims open where the sample cannot support them.

The useful follow-up is to compare these early controls with evidence from systems operating under real organizational constraints.

You don't have to scroll every AI thread, track every new tool, or watch every demo. 

The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning. 

Trusted by 2M+ professionals at Apple, Google, and NASA.

Sign up to start learning.