TL;DR
- Most AI tools free-wheel: you type a request, the AI decides everything, and you become the quality control on every single output, forever.
- The tools you can actually trust use defined activities instead. The steps are fixed, and the AI works only at the points where judgment is needed.
- The part almost every tool skips is what happens when the AI is not sure.
- In Akigo! Pulse, each of its 100+ activities is a defined loop. It pre-fills what it can infer, and a low-confidence guess surfaces as a Question card rather than getting quietly written into your register as fact.
- That mirrors how you already work with a good analyst: the confidence is part of the deliverable, and the uncertainty is flagged, not hidden.
- AI is an amplifier. It makes a good PM faster and a careless one more confidently wrong. The craft is knowing when to trust it, and building tools that know when not to.
Ask a project manager what they want from an AI tool and you will not hear “creativity.” You will hear something closer to “I want it to do the boring parts without making me double-check every line.” Those are not the same request, and most AI tools quietly deliver the first while pretending it is the second.
The reason comes down to something the industry has started calling the agentic loop, and once you see the distinction it explains why some AI tools feel like a trustworthy junior analyst and others feel like a confident intern you cannot leave alone.
Two ways an AI tool can work
An agentic loop is just the cycle an AI runs to finish a task: look at the situation, decide what to do, do it, check the result, repeat. There are two ways to build that cycle.
The common one is what I would call the free-wheeling loop. You give the AI a goal and it decides everything itself, top to bottom. What to do first, whether to check its work, when to stop. This is how a chat box works, and in a chat box it is fine, because you are sitting right there to catch it. The catch is the problem. It means you are the quality control, on every single output, forever. For a one-off question that is fine. For a status report you produce every week, it is a tax you pay in perpetuity.
The other kind is a defined loop. The steps are fixed in advance. The AI is still doing real work, but only at the specific points where judgment is needed, and the tool has an opinion about what to do when the AI is not sure. That second half is the part that matters, and it is the part almost every tool skips.
What a defined AI activity looks like
I want to make this concrete, so let me use the tool we built for exactly this problem. Akigo! Pulse is a desktop hub for project, change, and improvement work, and it ships with more than a hundred defined activities: status reports, risk registers, stakeholder analyses, process assessments, the documents that make up the actual week of a working PM.
Each of those is a defined loop, not a free-wheeling one. Here is what that means in practice when you open, say, a risk register.
The activity is a form with a known shape. Pulse reads what it already knows about the project and pre-fills the fields it can reasonably infer. That is the AI doing the tedious first-draft work, the part that eats your afternoon. So far this sounds like every other tool.
The difference is what happens when the AI is not confident. It does not paper over the gap. A low-confidence proposal does not get written into your register as if it were fact. It surfaces as a Question card, a plain ask for you to resolve, sitting apart from the fields it was sure about. The tool draws a visible line between “I am confident enough to fill this in” and “you need to decide this one.” You are reviewing a draft with its uncertainty labeled, not proofreading a wall of text that is all delivered in the same calm, certain voice.
That is the whole trick, and it is a small one, but it changes the relationship. A free-wheeling tool makes you the safety net for everything it produces. A defined activity with a confidence gate does the safe parts on its own and hands you exactly the decisions that were yours to make anyway.
Why this maps onto how the job actually works
If you have run projects for any length of time, this will feel familiar, because it is how you already work with people.
A good analyst does not hand you a finished risk register and swear every line is right. They hand you a draft and say “these three I am sure about, this one I need your call on, and I could not find anything on that dependency.” The confidence is part of the deliverable. The uncertainty is flagged, not hidden. You trust that person precisely because they tell you where the edges of their knowledge are.
A language model, left to free-wheel, does the opposite by default. It will state a date, a number, or a dependency with total confidence and be flatly wrong, in the exact same tone it uses for the things it has right. That is not a bug you can prompt your way out of. It is the nature of the thing. So the tool around it has to supply the missing judgment: it has to know when to stop and ask. Building that in is more work than letting the model run free, which is exactly why most tools do not bother.
The part the hype skips
Here is what I keep coming back to. The PMs who get real leverage from AI are not the ones with the cleverest prompts. They are the ones who can look at an output and know instantly whether it is any good. That judgment does not come from a tool. It comes from having run enough real projects to feel when something is off.
A well-built AI activity respects that. It does not try to replace your judgment with confidence. It does the volume work, marks its own uncertainty honestly, and routes the real decisions back to the person who is supposed to make them. That is the difference between a tool that makes a good PM faster and one that makes a careless one more confidently wrong.
Which is the whole reason the apps we build start from thirty years of practice rather than from the latest model. The model is an amplifier. The discipline of knowing when to trust it, and building tools that know when not to, is the actual craft.
If you want the deeper version of this, the workflows, the worked examples, and where AI earns its keep in a real project week, it is the heart of the AI-for-PMs eBook and video series. Drop your email on the waitlist and you will be first to get it.