Summary
Who sets limits for AI agents? IT, the business, or leadership? The short answer is all three, each at its own level. An agent can be responsible, but a human is always accountable. My thesis is that this shifts our roles in projects toward commissioning, scoping, and overseeing. And that exactly these competences matter more now.
How Did This Post Come About?
On October 29, I asked on LinkedIn who in an organization decides what an AI agent is allowed to do. The trigger was a series of news items in which, each time, the limit was drawn by whoever feared the damage. A shared set of rules was nowhere in sight.
A week earlier, the Project Management Institute (PMI) had introduced a new course at its Global Summit in Detroit. It is meant to give project professionals the literacy and judgment to scope, oversee, and govern AI agents responsibly. I touched on it briefly in my post AI and the Future of PM Certification. Here I want to go one step further. What does this actually mean for our roles and competences?
What Is Actually New About Agents?
PMI puts it this way. AI agents are moving from assistants to actors, systems that take steps, make decisions, and execute work. A chat assistant gives you a text. An agent sends it, opens the ticket, or changes the plan.
That turns a tool question into a delegation question. And delegation is something we project managers know well.
A survey quoted by PMI shows where practice stands. In the SAP LeanIX Agentic AI Survey 2026, 98 percent of companies said they had deployed AI agents or planned to. Only 17 percent had visibility into how well and how compliantly their agents perform. 48 percent lacked clear roles and responsibilities for governing them. The respondents were just over 200 IT professionals, and it’s a software vendor’s survey. Still, the direction is clear.
I’ve seen this on a small scale myself. When I introduced my AI assistant Max this spring, the biggest challenge wasn’t the technology. It was writing down how he should behave. For every action with external impact, the rule has been the same ever since: show first, then wait, then act. At its core, it was onboarding a new team member. With a role description, rules, and escalation paths.
Who Sets Limits for AI Agents?
Back to the question from the LinkedIn post. IT? The business? Leadership?
In my opinion, each of the three answers falls short on its own. The limits sit on three levels.
- IT sets the technical limits. Access, permissions, logging, a kill switch.
- The business and the project set the content limits. Which task, which data, which result, and when a human checks.
- Leadership sets the frame. How much risk is acceptable, who answers for it, and what never happens without approval.
The relevant frameworks see it much the same way. The NIST AI Risk Management Framework calls for roles, responsibilities, and lines of communication around AI risks to be documented and clear to everyone (GOVERN 2.1). The accompanying Playbook explicitly lists project management among those roles. And under GOVERN 2.3, executive leadership takes responsibility for decisions about AI risks.
The EU AI Act gets more specific, but only for high-risk systems. Under Article 26, deployers must assign human oversight to people who have the necessary competence, training, and authority. Article 14 describes what those people should be able to do.
- Understand the system’s capabilities and limitations.
- Stay aware of the tendency to over-rely on its output (automation bias).
- Disregard, override, or reverse its output.
- Stop the system.
Most agents in everyday project work are unlikely to be high-risk systems. In my opinion, the four points still make a good yardstick.
A participant in my workshop in March summed it up: “Accountable is always the human. Responsible can be human or AI – depending on the task and its complexity.” In RACI terms, an agent can get an R. Never an A.
Which Roles Are Shifting?
The project manager becomes the one who commissions and oversees agents. An agent needs an assignment with a goal, boundaries, quality criteria, and an escalation path. In other words, exactly what we should be defining for every work package anyway. What’s new is the pace. An agent delivers in minutes, and the review has to keep up.
According to the Scrum Guide, the Product Owner is accountable for maximizing the value of the product. And the Product Owner is “one person, not a committee.” An agent can propose, refine, and pre-sort backlog items. The decision on the order stays with the Product Owner. That, too, is a limit worth stating explicitly.
I see the biggest shift in the PMO. Whoever owns templates, methods, and reporting today will, in my estimation, also own the standards for agents tomorrow. A register of which agents run in which projects with which permissions. Consistent rules for approvals. A review of where agents got it wrong.
Steering committees and portfolio boards should decide how much autonomy an agent gets. That’s a risk decision, just like releasing a budget. So it belongs in the governance bodies and not with IT alone.
Which Competences Matter More?
In short, less creating things ourselves, more commissioning and checking. Four competences stand out.
- Scoping. Cutting an assignment so an agent can actually do it. And saying clearly what it must not do.
- Overseeing. Taking samples, reading logs, spotting warning signs, stepping in on time.
- Judging. Telling a plausible result from a correct one.
- Communicating. Explaining to stakeholders what an agent did and who answers for it.
Very little of this is truly new. ICB4, the basis for today’s certification, describes 28 competence elements for projects. Many fit directly, such as “Governance, structures and processes,” “Scope,” “Risk and opportunity,” “Leadership,” “Personal communication,” and “Stakeholders.”
ICB5 keeps the “Eye of Competence.” Artificial intelligence is one of its five cross-cutting themes, and Digital Enablement is a new competence element. PMI’s agent course also explicitly requires no coding background.
I firmly believe that leading agents mostly takes familiar competences applied in a new way. If you’re good at delegating, setting boundaries, and asking follow-up questions, you have a head start. If you’ve always done everything yourself, you’ll have to relearn.
How Do You Get Started?
A few things I recommend for any project that uses agents.
- Create a short profile for each agent. Assignment, data, permissions, limits, responsible human.
- Name one person as accountable. By name, not by department.
- Start with small permissions. Read before write, write before send.
- Allow actions with external impact only after approval.
- Define a kill switch. Who can stop the agent, and how fast?
- Collect and review errors. Where did the agent get it wrong, and what did we change?
- Add agents to the risk register and report on them in the steering committee.
Interestingly, PMI’s course also ends with a one-page “Stakeholder Brief” for your own agent. That’s essentially the same profile.
And if you want to be more thorough, run a pre-mortem. Imagine an agent has done something expensive in your project. Who gave it which permissions? And why did nobody notice?
What Does This Mean for the Coming Years?
From here on, this is my assessment, not an established development.
PMI’s interim CEO Johannes Heinlein expects project professionals to become essential to governing autonomous systems. That’s an expectation, not a fact. I share it with one caveat. It only holds if we actively take on the role. Otherwise IT will handle it alone. Or nobody will, as the 48 percent from the LeanIX survey suggest.
For certification, this means it will have to test whether someone can commission and oversee agents sensibly. Which brings us back to the question from my ICB5 post. Document or conversation? Judgment tends to show in conversation.
Delegating has always been part of our job. What’s new is that the delegate never gets tired and often only asks questions when we explicitly tell it to.
Who in your project is accountable for the first agent? And does that person know yet?
Feel free to subscribe to my free newsletter. And follow along on Mastodon, Threads, Bluesky, LinkedIn and Instagram.
Sources
- PMI: AI Fluency Plus Human Judgment – Project Management Institute Helps a Growing Profession Deliver Greater Value. Press release, October 22, 2026 – https://www.pmi.org/about/press-media/2026/project-management-institute-helps-a-growing-profession-deliver-greater-value
- SAP LeanIX: Agentic AI Survey 2026 – https://www.leanix.net/en/download/agentic-ai-survey-2026
- NIST: Artificial Intelligence Risk Management Framework (AI RMF 1.0), GOVERN function, January 2023 – https://airc.nist.gov/airmf-resources/airmf/5-sec-core/ · Playbook GOVERN 2.1 – https://airc.nist.gov/airmf-resources/playbook/govern/
- Regulation (EU) 2024/1689 (AI Act), Articles 14 and 26 – https://artificialintelligenceact.eu/article/14/ · https://artificialintelligenceact.eu/article/26/
- Ken Schwaber, Jeff Sutherland: The Scrum Guide, November 2020 – https://scrumguides.org/scrum-guide.html
- IPMA: Council of Delegates Communiqué (ICB5), October 4, 2026 – https://ipma.world/news/ipma-cod-communique-3/
- IPMA: Individual Competence Baseline for Project Management, Version 4.0 (2019 edition, hosted by IPMA Switzerland) – https://www.vzpm.ch/fileadmin/dokumente/downloads/English/IPMA_ICB4_PM_2019_EN.pdf





