On 25 September 2026 OECD.AI published an article on AI agents in government that discusses how systems are moving from answering questions to carrying out steps towards a goal. It is an analytical piece reflecting its authors' views, not binding regulatory guidance. What interests me in this development is defining the relationship between a tool's ability to act and its authority to do so within a public institution.
An application assistant with clear limits
Imagine an assistant that reads an application and its attachments, searches the procedure manual, and then prepares a preliminary list of missing items. That is a different task from sending the applicant a formal request or changing the application's status. The two can be combined in a single user experience, but their boundaries should stay clear in the system, in the permissions and in the event log.
My proposal is to begin with a task limited to reading and preparing a draft. The assistant shows the document it relied on, the fields it could not find and the questions that remain open. A member of staff then reviews the result before any action that affects the case. This way, the quality of the analysis can be evaluated separately from the effect of execution.
Approval needs details
A general phrase such as approval to proceed is not enough if the next step will modify a record. The reviewer should see the application concerned, the proposed change, the reason for it, and the party that will receive any notification. The approval is tied to that specific operation so that it does not turn into an open-ended mandate for later steps.
- Read only the data available to the job role.
- Present the evidence and missing information alongside the draft.
- Request a clear review before changing a status or sending a message.
- Log the outcome of every attempt, including failures and cancellations.
How do I test it before widening its role?
I choose cases that reflect reality: a complete file, an illegible document, a conflict between two sources, and a request outside the department's remit. I then measure the assistant's ability to tell these cases apart, and to stop when the information is not enough. Testing must also cover dropped connections and repeated requests, so that a retry does not produce a duplicate action.
The practical path starts with a small function that can be measured, and widens once there is evidence of its usefulness and of the quality of its review. Speed in carrying out steps is not enough on its own; what matters more is that every change remains understandable and can be traced back to its source and its approval.