Back in 2019, Gartner published a forecast that felt like a provocation at the time: by 2030, artificial intelligence would take over roughly 80% of project management tasks.
Seven years later, we're somewhere in the middle of that forecast — and it's now safe to say what it actually meant. Not "a robot will replace the project manager."
But "80% of the administrative routine a manager does by hand today will stop being human work."
That difference matters. And it's exactly the part people miss when they talk about AI in project management.
First, let's look at where a manager's time actually goes
There's one piece of research that brings everyone back down to earth. Asana once calculated that office workers spend about 60% of their working time on "work about work": chasing statuses, clarifying things, rewriting the same thing in different words for different people.
Wellingtone's annual report adds another detail: half of all companies spend at least one full working day per month just assembling an up-to-date project status. Not analysing it. Just assembling it.
Now picture a project manager running five active projects.
How much of their week is decisions, prioritisation, negotiation — the things they're actually valued for?
And how much of it is translation? The client describes something in business language — it has to be translated for developers. A developer replies in technical language — it has to be translated back. A meeting ends — an hour of conversation has to become five agreed points. A month ends — completed tasks have to become an act of work and an invoice.
This is exactly where AI turned out to be useful. Not as "intelligence", but as a very fast and very patient translator between formats.
What has changed over the past two years
Here's the curious part: the tools arrived for everyone at roughly the same moment, but systematic use is still almost nowhere to be found.
According to PMI, only about a third of organisations have adopted AI in their project processes to any moderate degree, and just 12% substantially. In tech-forward companies the number is higher, around 34% — still far from "everyone".
APM research adds an even more telling detail: 68% of project managers say they haven't been properly trained to use these tools.
So the picture looks like this: everyone has access, half have the skills, and almost nobody has shared rules.
It's a bit like a company that bought every employee a car but never built any roads. Some drove off fast. Some are parked. Some drove in the opposite direction, and everyone found out a month later.
We walked exactly that path. Which is why the rest of this article is about processes, not technology.
How it works at Molfar
We don't treat AI as a separate "innovation project". We apply it in specific places — wherever there's repetitive work with text and structure. Right now that's six areas.
Writing tasks for developers. A client describes a problem the way they see it: "the cart behaves strangely on mobile." A developer's task has to look different — reproduction steps, expected result, acceptance criteria. The quality of that translation used to depend on how attentive the manager was that day and how much of a rush they were in. Now the draft is generated to a single structure, and the manager reviews it and adds the context only they have.
The effect here is easy to measure. A properly written task used to take about an hour. In that same hour, a manager now produces five tasks, review included.
Technical documentation. Most of the time documentation doesn't get written not because nobody needs it, but because there's never time for it. When a draft description of a module or an integration appears immediately after implementation, the odds of that document surviving to a final version go up sharply.
Functional requirements deserve a separate mention. Describing a new block of functionality used to take a week — and that was before client approval even started. That stage is now roughly three times shorter. More importantly, approval got shorter too: when a client sees a structured document rather than a stream of thoughts, they come back with specific edits instead of a vague feeling that "something's off here."
Acts of work and invoices. At the end of the month, everything delivered has to be collected and turned into a list of work the client can actually read. It's pure mechanics — the kind of work where a human gets tired and makes mistakes, and a machine doesn't.
Recording what was agreed in meetings. The most expensive misunderstandings on a project don't come from conflict. They come from the phrase "but we agreed on this." Transcribing a meeting and structuring it into a short list of decisions and next steps closes that gap almost entirely. The client gets the summary the same day and can immediately say: no, what I meant here was something else. That is far cheaper than finding out three sprints later.
Written client reporting. A report shouldn't be a list of tickets. It should answer the question "what changed for my business this period?" Those are two different texts — and the second one used to take a full working day, which is exactly why there was always a temptation to postpone it or send whatever was quicker. The draft is now generated automatically and the manager spends an hour checking and completing it. Reporting stopped being a feat of endurance, so it simply started arriving on time.
Keeping task statuses current. The most mundane item on the list — and, oddly enough, the most useful one. An accurate status in the tracker is no longer the result of a manager walking around on Friday evening asking everyone "so, how's it going?"
The hardest part wasn't the technology
Once all six areas were running, we hit a problem we hadn't anticipated. Every manager was doing it their own way.
One described tasks in detail, another kept them short. One manager's act of work looked like this, another's looked like that. One recorded agreements immediately, another when they got round to it. Formally, everyone was using AI. In practice, we had four different schools of work inside one company.
Which raises an uncomfortable question: what does the client see?
The client sees that the quality of communication depends on a specific person rather than on the company. That is precisely what an agency is supposed to protect its clients from.
So we're now building a single scheme — a master template that fixes what a task looks like, what a report looks like, what a meeting summary looks like, where each piece of data comes from and who is responsible for what.
The logic is simple. The tool gives you speed. The template gives you predictability. Without the second, the first just accelerates chaos.
Where we deliberately don't use AI
This is probably the most important part.
- Estimating timelines and budgets. AI is excellent at producing plausible numbers. The problem is that a plausible estimate and a realistic estimate are not the same thing. Our estimates come from a human being who will then be accountable for them.
- Difficult conversations with clients. A missed deadline, an overrun, a production bug. What's needed there isn't a perfect text — it's a person taking responsibility. A flawlessly polished email in that situation only makes things worse.
- Final calls on priorities. A machine doesn't know that this client has an investor presentation next week, and that one has just survived a difficult release.
- Confidential data. Access credentials, personal data, commercial terms — none of it goes near any external service. The rule here is simpler than any methodology: if you're not sure, don't send it.
And one more observation worth stating on its own. The most dangerous property of these tools isn't errors. Everyone makes errors. What's dangerous is the confident tone the errors arrive in.
A text that has mixed up two integrations looks exactly as convincing as a text that has everything right. So we have one unbreakable rule: no document reaches a client without a human who has read it and put their name to it. You can automate preparation. You can't automate accountability.
What the client gets out of it
In short — not "magic", but things that are easy to verify. A task a developer understands the first time, rather than after the third round of clarification. A functional description in two days instead of a week. A meeting summary the same day. A report written in human language rather than tracker language. An act of work that shows exactly what was paid for.
And a manager who has time left for the thing they were hired to do: think about the project, instead of retyping text from one format into another.
To be straight about it: the time saved doesn't disappear from the project, and it doesn't turn into a discount. It goes into the things there was never enough time for — closer review, more questions asked before work starts, faster reaction when priorities change. That's where the difference between "built to spec" and "built the way the business needed" actually lives.
Conclusion
Over the past few years we've arrived at a fairly down-to-earth conclusion.
Artificial intelligence doesn't make project management smarter. It makes it more disciplined — but only when the discipline already exists inside the company.
If your processes are chaotic, AI will accelerate the chaos. If your processes are defined, it will take the routine off people and give them back time for decisions.
Which is why the question "do you use AI?" tells you almost nothing about an agency.
The far more interesting questions are: where exactly do you use it, where do you deliberately not use it — and who is accountable for the result.
Molfar Insight
We don't believe artificial intelligence will replace the project manager.
We believe it will take away the part of the job clients never wanted to pay for — and leave the part they hire a manager for.