You are a project manager or a PMO who is preparing the next project review and, in the meantime, every week a new announcement arrives on artificial intelligence: tools that "rewrite" project management, reports that generate themselves, automatic forecasts, assistants that promise to save you hours. The question you ask yourself is legitimate: of all this, what really stands up to the test of facts?
This article tries to answer honestly. You won't find the promise that AI will do your job for you: it's not true, and it's not even the point. You will find a practical map: what AI can do well today, what it can't do, how to evaluate a tool and what risks to keep an eye on. The underlying criterion is only one: first the verifiable data, then the interpretation.
What AI can do well today
Three activities, if well delimited, really work.
Summarize.A language model can read minutes, mail threads and project documents and return an orderly summary: decisions made, actions assigned, deadlines. The summary must always be reread — the models can lose details or flatten nuances — but it eliminates the most tedious work: the first pass over dozens of pages.
Structuring.From scattered notes, raw data and messy logs, AI can produce a first preliminary version report, a risk checklist, a topic index for review. Structure is useful when it is treated as a starting point: refinement, i.e. deciding what is truly relevant, remains a human job.
Report anomalies on structured data.On real and well-defined numbers - budget, progress, deadlines, commitments - the most useful work is not creative but statistical: comparing the values with thresholds, trends and previous periods and reporting the deviations to be verified. It is the only type of analysis that can be checked item by item, because each report refers to a specific piece of data.
The scientific literature, moreover, is as cautious as necessary: onesurvey on linguistic models applied to planningrecords notable results on some tasks, but concludes that broader application still requires systematic testing. In a word: promising, not established.
What AI can't do
He can't decide for you.A model does not bear the responsibility of the decision, does not suffer the consequences and does not know the constraints that are not written anywhere: the relationship with the customer, the culture of the company, the promises made orally, the real availability of people. The decision remains, entirely, with the project manager and the sponsor.
It cannot replace judgment.Distinguishing an acceptable delay from a dangerous one, understanding whether a risk is truly mitigated, evaluating whether a person is overloaded: these are judgments that are formed on context and experience, not just on data. AI can prepare material for judging; judgment is not.
You don't have to invent data.Here lies the most important boundary. Generative models can produce plausible but untrue content: so-called hallucinations, a phenomenon widely documented in research (a dedicated surveyreconstructs types, causes and countermeasures). In a project report, an invented but plausible number is worse than missing data: it looks like data, and is used as such in a decision.
Hence the principle that, in my opinion, should guide every use of AI in project management: verifiable data first, then interpretation. The AI can only work on the interpretation after the data has been checked; never the opposite. If the data does not exist or is not verified, the interpretation does not start.
| Activities | Can AI help? | What's left for the project manager? |
|---|---|---|
| Summarize minutes and documents | Yes, with rereading | Validate contents and decisions |
| Structure reports and checklists | Yes, as a preliminary version | Choose what is relevant |
| Report anomalies on structured data | Yes, verifiably | Check the signals and decide on actions |
| Produce data and estimates | No: he can invent them | Produce and validate each number |
How to evaluate an AI project management tool
Before adopting a tool, five checks, in order.
- What is deterministic and what is generative.Rules, thresholds and calculations must be deterministic and reproducible: same data, same results. The generative part must be declared, separate and optional.
- Where does the data come from.Each report number must trace back to a precise and verifiable source. If a number has no source, it is not data: it is a decoration.
- How uncertainty signals.A good tool tells you when data is missing or old, rather than filling in the gaps with plausible values.
- Who controls the outcome.There must always be a human review step between the generated content and the decision maker. If the tool doesn't provide it, you provide it.
- How it processes personal data.Projects contain data about people, costs and contracts: the use of AI tools must comply with the GDPR, and theGuarantor for the protection of personal data collects specific indications on the topic of artificial intelligence.
Risks to watch out for
The first risk is thehallucinations: Plausible but false numbers or facts, presented as confidently as true data. In a project review, a single made-up number can taint a decision worth months of work. Basically: if a generated report contains a number that you can't verify in thirty seconds, treat it as an opinion.
The second is theaddiction. If the team stops checking because "the tool takes care of it", quality control collapses without anyone noticing. In my opinion this is the most underestimated risk: the damage does not come from AI that makes mistakes, but from human control that slowly shuts down.
The third is theopacity: accepting a result without being able to trace how it was produced. A report that doesn't show the evidence isn't a report: it's an opinion in a better format.
None of these risks are a reason not to use AI. This is a reason to place it at a precise point in the chain: after the verified data, before the human decision.
How ControlRoom fits in
In ControlRoom the default engine is deterministic: indicators, thresholds and calculations on project data are reproducible and verifiable. AI-assisted reports are an optional level still under validation: activated, but distinct from the main reporting and with evidence cited next to each statement. Today you can see how it works with onefree guided demo. The Free plan is in preparation and will include up to 2 AI reviews valid per workspace per week and 3 one-off PDF/PPTX exports.
If you want to delve deeper into the reasoning, you can start fromprocess analysis for PMO in research and development, fromproject analytics guideor from the page onAI for process intelligence.
The other two articles of this path also derive from this principle:the ten questions of a project review that decidesethe risk of missing data in reports.
The next step
Before choosing or discarding a tool, do the simplest test: ask to see where each number in the report comes from. If the answer is not a verifiable fact, you already have your assessment. If you want to see a setting that keeps deterministic and AI separate, start with oneControlRoom free guided demo. And if the question is "in a project review, where is AI?", the answer is the same as data: it helps prepare, not decide.