Uğur Deniz İlhan
  1. Home
  2. Insights
  3. AI Agent Manager
  4. How is agent governance set up? Inventory first, policy second
AI Agent Manager · Analysis · 16 November 2026 · 7 min read

How is agent governance set up? Inventory first, policy second

The first thing governance brings to mind is writing a policy document. A working arrangement is built from a five-line record for every agent and a short review held once a month.

AI Agent ManagerWhere does governance start?
Short answerAgent governance is set up by having five things in writing for every agent in the company: its owner, its brief, its permissions, the steps that need human approval and its success measure. These records are kept in a single agent inventory and reviewed once a month. A comprehensive policy document is not a precondition; it is what emerges over time.

Why should governance not start with a policy document?

Because a policy speaks in general terms, while an agent does a specific job. The sentence “AI is used responsibly” does not say which data the agent that prepares Monday’s report can access, or who will notice its mistake.

Writing a policy takes time and goes through approval. Meanwhile teams keep setting up agents. By the time the document is signed, the company may have a set of agents nobody has a full list of.

Agent governance, meaning the arrangement that sets who manages the agents, under which rules and with which records, works when it is built from the bottom: first a record for each agent, then the shared rules that come out of those records.

What do the standards say?

The main references are these. The US National Institute of Standards and Technology (NIST), in the AI Risk Management Framework it released on 26 January 2023, defines four functions: govern, map, measure and manage. The framework positions govern as a cross-cutting function that informs the other three. NIST notes that it is updating the framework; the source should be checked for the current text.

ISO published ISO/IEC 42001 in December 2023, the international standard for AI management systems.

In the European Union’s AI Act (Regulation (EU) 2024/1689), Article 14 covers human oversight: high-risk systems must be designed so that people can oversee them while in use, and the person overseeing must be able to override the output and stop the system.

These texts are not specific to agents; they were written for AI systems in general. In our view what they share is this: responsibility is tied to a person, and oversight is treated as part of the design. The arrangement below is the small-scale counterpart of those two ideas.

What is an agent inventory and what goes in it?

An agent inventory is the list on which every agent working in the company appears on one page. Each agent is a row, and each row answers the same five questions.

RecordQuestion it answersIf it is empty
OwnerWhich person is accountable for the result?When an error appears, nobody takes it on.
BriefWhat does the agent do and not do?The agent fills the gap with its own guess.
PermissionsWhich data does it read, where does it make changes?Access stays wider than the task requires.
Approval stepsWhich step is tied to human approval?An action that cannot be undone passes unchecked.
MeasureHow is it known the task was done well?The agent’s benefit is described as “seems good”.

One more column is enough: the date of the last review. How to write the brief is covered in the guide with a template, and how to choose approval steps in the article on human approval.

What does the monthly review look at?

Four questions. The meeting runs down the inventory and takes a few minutes per agent.

  • Did the agent do anything unexpected last month? If so, was the brief updated?
  • Where does the measure stand against the baseline?
  • Are any permissions unused? If so, they are removed.
  • Do the approval steps need changing: is there a step to relax or to tighten?

A new agent also enters the inventory at this meeting. The rule that an agent without a record does not run is the one rule that holds the whole arrangement up.

Who does what?

Governance is the work of four parties, and each asks a different question. The agent’s owner asks “is this task being done correctly?” and keeps the brief and the measure. The AI Agent Manager looks after the inventory as a whole, runs the monthly meeting and makes sure no agent is left without an owner.

IT and security grant and withdraw permissions; which system is open to which agent is in their records. Leadership decides which risk is acceptable. Leaving that decision to the technical team means the company’s risk limit is set by whoever does the setup.

What happens when an agent makes a mistake?

The sequence is known in advance and has four steps: the agent is stopped, what happened is recorded, the cause is found, a decision is made. Where to look for the cause is also known: a gap in the brief, a permission that was wider than needed, or a missing approval step.

The options for the decision are written down beforehand: continue with the same authority, continue with narrower authority, or stop. The agent’s owner makes the decision and notes it in the inventory. A record that does not hide errors makes the next agent safer to set up.

Assessment

In our view what matures governance is not how thick the document is but how current the record is. A company with two agents does not need a certification process; it needs the owner and the limits of those two agents in writing. The standards become a guide after that point.

  1. This week, write every agent and automation running in the company into one table. If nobody knows the full list, that is the first finding.
  2. Put an owner’s name on every row. Stop any agent that cannot be given a name until an owner is found.
  3. Have the owners fill in the missing columns; give them a month.
  4. Put the date of the first review meeting in the calendar now.
  5. Write the company policy after the third meeting. By then the inventory will show which rules are actually needed.

To see where your organisation stands today, take the eight-question agent readiness test.

Frequently asked questions

Does agent governance need separate software?

Not at the start. As long as the inventory fits in a spreadsheet, a spreadsheet is enough. A tool is worth discussing when the number of agents grows and keeping records current by hand becomes difficult; by then it is also clear what needs to be tracked.

Who runs governance?

Every agent has its own owner; one person looks after the inventory as a whole and the monthly review. This role is called the AI Agent Manager. IT and security take the access side, and leadership takes the risk decisions.

Does the EU AI Act apply to every agent?

No. Article 14 of the Act, on human oversight, is written for AI systems classed as high-risk. Whether a given agent falls within that scope depends on what it is used for and requires a legal assessment.

Sources

Read next

All insights

For a conversation, message on WhatsApp.

A short introduction is enough: your company and what you would like to discuss.

Open WhatsApp

+90 530 242 14 55