Project and Task Management with AI

First, the bad news

A peer-reviewed, controlled study measuring specifically the quality of a project status report, a risk register or a next-step recommendation does not exist. I went looking and did not find one.

That does not mean it does not work. It means that anyone quoting you a specific percentage for this is not citing a measurement — they are citing a vendor case study or their own impression.

The largest available survey on the state of the profession asked 2,314 project management professionals across 129 countries. The data was collected in the summer of 2023, though, so it is three years old — which is how its numbers should be read:

• only 35% said their organisation had a live or pilot AI project;
• 60% put their organisation's AI maturity at four or below on a ten-point scale;
• 65% had no AI knowledge at all, or only a basic level of it;
• the biggest expected impact was marked for data collection and reporting (72–76% depending on the region), the smallest for stakeholder management, communication and budgeting.

That last breakdown still holds up as a guide today: the machine takes the administrative layer, not the relationship or the financial judgement.

What we do know: two government rollouts

Two large public-sector deployments published task-level numbers. Both are self-reported, not stopwatch data — but at least the samples are large and the methodology is published.

The publisher of the second rollout added a note to its own figures saying that these are approximations, and probably quote the upper bound. That sentence is worth reading together with the table, otherwise the numbers mislead.

One further, rarely quoted figure from the same source: in the spreadsheet app and the presentation app take-up was 23% and 24% respectively. Which means the numerical part of project work is weakest exactly where it would help most.

The three documented failure modes

These are concrete findings, not fears.

1. You almost always have to edit. One of the rollouts states in so many words that editing was almost always necessary to fit the content to the audience or the context. 60% of participants reported a moderate or significant amount of editing.

2. For some people the work grew. Up to 7% of respondents said the tool added time to the task, because of the need to check it. This is the most concrete documented negative-return figure I found — and it is valuable precisely because it comes from an organisation's own report evaluating its own rollout.

3. Search brings up outdated documents. According to the focus groups, the built-in search frequently surfaced outdated or irrelevant documents. In a project setting that is not a detail: a status generated from a six-month-old specification gives a false picture of progress and dependencies, and it looks exactly like the correct one.

Why the status report is where it is most dangerous

There is one measurement that speaks to project management indirectly but forcefully.

In a controlled experiment, experienced professionals working on their own familiar work came out 19% slower with AI — while beforehand they had expected a 24% speed-up, and after finishing the task they still believed in a 20% speed-up. Thirty-nine percentage points between the feeling and the stopwatch.

For a project lead, one thing follows from this: if a specialist is that wrong about how long their own work takes, then estimates and progress reports cannot be verified by self-reporting. Not by human self-reporting, and not by what the machine kneads together out of human reports.

A status does not become true by being nicely worded. It becomes true because there is a closed, checkable item behind it.

What changed in 2026: credits instead of price

This is the most concrete recent development in the area, and worth knowing about for cost planning.

In project management tools, AI is no longer an add-on you buy separately; it is part of the paid plan. Several large tools advertise explicitly that it is included at every paid tier.

The limit, though, has moved out of the price and into the credits. At one large tool, custom agents came with the business plan without any limit until 3 May 2026; from 4 May they switched to usage-based billing, at $10 per 1,000 credits. At another, the credits consumed depend on the complexity of the interaction and the number of API calls, and the allowance differs by plan. In several places there is a clause saying that access to features may be temporarily reduced depending on usage.

So the question is no longer "is it worth buying" but how much of it gets used up in a month — and that is hard to estimate before a rollout. The same shift has happened in office suites, which the built-in office AI post covers with current prices.

What you can hand over, and what you cannot

Four things where both the state of the evidence and the nature of the task argue for the machine:

SAFE TO HAND OVER
1. Structure out of raw material
   A task list from meeting notes, a breakdown from
   an outline, a summary from a long thread.

2. Rewriting for an audience
   The same status for the board and for the team.
   The strongest measured area in the rollouts.

3. Consistency checking
   "Which task has no owner or no deadline?"
   That is a search, not a judgement.

4. First draft of a risk register
   As a seed for ideas, not as analysis.

And three where it is not:

NOT SAFE TO HAND OVER
1. Progress and estimates
   Human self-reporting is off by 39 points. What the
   machine kneads out of it will be no more accurate.

2. Establishing dependencies
   The built-in search demonstrably surfaces stale
   material. A false dependency is worse than none.

3. Stakeholder decisions
   The survey puts this in the lowest-impact area,
   and not by accident: it is relational judgement.

The practical working order that comes out of this: the machine writes the report, the system supplies the data. If the source of a status is a closed task, a commit, a signed handover — then you can safely hand over the wording. If the source is what somebody said in a meeting, the report will be exactly as accurate as that sentence was.

Common misconceptions

• "They measured that it saves X% on project administration." They did not. What exists is task-level self-reporting, and its publisher writes on it that it is an upper estimate.
• "The machine will spot the slippage." Spotting slippage requires knowing what the reality is. The built-in search demonstrably surfaces outdated material — and out of that comes a confident but false picture.
• "If it is in the plan, it is free." In 2026 the limit moved into the credits. The price is not in the licence, it is in the usage.
• "I only hand over the report writing, that is harmless." The report is what the decision gets made from. The wording can be handed over — the progress figure inside it cannot.
• "It saves me time." It may. But in that same rollout, up to 7% of respondents said their work grew because of it, and 60% of them had to do substantial editing.

What to take away

In project management, AI today is an administrative layer, not decision support. That is not a put-down: administration is a genuine burden, and that is exactly where the measurements show a gain.

The boundary is sharp, though, and a single question draws it: where does the number come from? If it comes from a system, you can hand over the wording. If it comes from a human estimate, the report does not become truer because a machine wrote it down — it will only sound more confident.

How a meeting turns into a usable task list is walked through step by step in the meeting notes post; for recurring, scheduled reports a recipe is what the automated reporting post provides.