AI agents on the project, and the question of who governs them

An AI agent, in the description PMI uses, is a system that works towards defined goals, uses tools and acts across multiple steps. The announcement it published on 22 October 2026 sums it up by saying agents are moving from assistants to actors, systems that take steps, make decisions and execute work.

That last word is the one that matters on a project. Whatever executes work enters the scope, consumes resources and produces deliverables, and all of that has an owner in any plan you have ever made. The uncomfortable question is who the owner is when whoever executes is not a person.

What the figures say

The same announcement cites the SAP LeanIX Agentic AI Survey 2026. According to that survey, 98 per cent of companies had deployed or planned to deploy AI agents. Only 17 per cent had visibility into those agents’ performance or conformance, and 48 per cent lacked clear roles and responsibilities for managing and governing them.

What strikes me most is the distance between the first figure and the other two. Almost every company already has them or will have them, and fewer than one in five knows how they are performing.

PMI adds its own data on the use of AI in the profession, published in the same announcement. Project professionals use it for research, analysis and information gathering (77 per cent), for creating content, documentation and deliverables (69 per cent) and for identifying risks, opportunities and emerging issues (53 per cent). And only 19 per cent are concerned their employer will replace them with AI. The announcement does not give the name or date of that research, so we cite it as PMI does.

The bronze giant who guarded Crete

There is a myth that helps think about this. Talos was a bronze giant who, in the most told version, was forged by Hephaestus to protect Crete. He walked around the island three times a day and hurled stones at the ships that came near. He never tired, never got distracted and did exactly what he had been made for.

He had a single vein, which ran through his body and was sealed at the ankle with a bronze nail. When the Argonauts reached the coast, it was Medea who found that spot, and the giant came crashing down.

The myth does not tell of a failure of Talos. Talos worked perfectly. It tells something else, that what held him up was a detail that, it seems, nobody else on the island was watching. An agent is similar. It can make its round flawlessly a thousand times, and the risk is not in the round, it is in nobody knowing where the nail is.

Four questions that require no coding

To govern an agent you do not need to know how to build one. You need to answer four questions you already know how to ask about any other resource on the project.

The four questions are our own. The figures are from the SAP LeanIX Agentic AI Survey 2026, as cited in PMI’s announcement of 22 October 2026.

The first is what it does. It is the age-old scope question, applied to something that does not read the project charter. What problem it solves, with what tools and, above all, what is left out. An agent without written limits does not have a broad scope, it has an undefined one, and what is undefined is the first thing to slip away.

The second is who oversees it. Saying there is human oversight is not enough. You have to decide at which steps a person steps in, whether they review before the agent acts or after, and what they do when they do not like what they see.

The third is who answers for it, and it is the one 48 per cent had not resolved. An agent can carry out a task, but it cannot be accountable for it in your responsibility assignment matrix, because nobody can hold it to account. If there is no person’s name in the accountable column, the agent is not governed, it is loose.

The fourth is how it is measured, and it is the one only 17 per cent had covered. Measuring an agent means setting its objectives before you set it running and reviewing its performance against them, just as you would with a supplier. If you do not know how it performs, you do not know whether it is helping or producing mistakes very quickly.

If this sounds like ordinary risk management, that is because it is. In the post on the risk breakdown structure with AI we saw how to sort those risks by category, and PMI’s standard on AI gives the general framework for project work with these tools.

What PMI offers

In the same announcement, PMI presented the PMI Foundations Agentic AI for Project Professionals course. According to the announcement, it gives project professionals the literacy and judgement to scope, oversee and govern AI agents responsibly, without requiring a coding or data science background. Its page says you finish it with a one-page brief that defines the problem, the agent’s role, the tools, the oversight design and the key risks.

If what you are after is specialising in AI projects, and not only governing agents inside a project, the path is a different one, and we covered it in the post on going from the PMP to the PMI-CPMAI.

Where the nail is

Talos did not fall because of bad design. He fell because a stranger knew his weak point better than the island that had him as its guardian.

Governing an agent is, at bottom, not letting that happen to you. Making sure that someone on your project knows where the nail is, and that this someone has a name.

Join the waitlist

At The PM Architect we write the PMBOK study guides in five languages, with real classroom cases and exercises that come with an explanation. They will be free. Leave your email on the home page and we’ll let you know the moment they’re ready.