Imagine a project office that prepares, each month, the portfolio committee of an IT department. Its three members copy reports in different formats into a consolidated view, write the summary note and search old records for what the committee had decided. Their director offers them an artificial intelligence assistant and asks what they intend to entrust to it. Between the PMO and AI, the split follows the nature of the tasks.
An assistant can prepare those that transform texts and data already produced: consolidating, summarising, flagging a gap, finding a passage in the documentation. Those that commit a person remain with the PMO and those it serves: proposing an arbitration, raising the alert, maintaining the relationship with sponsors. This split holds under conditions: well-kept project data, human review, rules on what can be entered into the tool.
What a language model does with a project report
These assistants rely on a large language model. Such a model produces text from text: it rephrases, summarises, classifies, extracts, compares two versions. It only knows of a project what is provided to it at the time of the request. What no one has written escapes it, starting with the reservation a project manager expresses out loud.
It can also produce a plausible and false sentence: in a note intended for a committee, an invented milestone date looks the same as an accurate one. Everything that follows therefore describes assisted tasks: the assistant prepares, a person from the PMO checks and signs. This check takes time, to be weighed against the time that the preparation saves.
For it to answer from a document collection, a search engine first passes the relevant passages to the model, following the principle of RAG. Our page on generative AI in the enterprise describes the design and evaluation of these assistants.
PMO and AI on consolidation, synthesis, gaps and search
| Task | What the assistant can prepare | What the data must allow | What a person keeps |
|---|---|---|---|
| Consolidation | Bringing reports in different formats into a common record, spotting empty headings | Headings with the same meaning from one project to another | Checking each date and each amount against its source |
| Synthesis | A first draft of the note for the committee, from the records and decision logs | Records that are up to date, dated, with a known author | Choosing what deserves the committee’s attention, and the signature |
| Gap detection | Flagging a date that differs between the schedule and the report, a risk with no owner | A written baseline: approved schedule, baseline budget | Sorting data entry errors from signals, then the conversation with the project manager |
| Document search | Finding a past decision or a clause, with the source passage in support | A sorted collection, outdated versions removed, access rights carried over | Reading the cited passage before any use of the answer |
Comparing two dates or checking that a field is filled in is a matter of a written rule, which a spreadsheet or the portfolio management tool applies predictably. The language model brings the reading of free text, for example a report announcing “slight pressure on acceptance testing” while the schedule remains unchanged. Calculating indicators stays on the side of rules: its conventions are covered in our article on PMO KPIs.
Arbitration, alerts and sponsors remain with people
An arbitration is formulated by weighing what each option costs, and to whom. The assistant can draw up the list of options from the files. It does not know what a postponement means for the team waiting for the delivery. The proposal submitted to the committee bears the name of a person, who will answer for it in the meeting.
An alert requires an author. Telling a sponsor that a project’s declared status no longer matches the facts commits the person who says it, and presupposes having first heard the project manager. A signal produced by the assistant is a clue to be investigated: passed on as it is to senior management, it exposes the PMO to raising a false alarm.
The relationship with sponsors rests partly on what appears in no document: obtaining honest information from a project manager in difficulty, choosing the moment for bad news. If assistance frees up time on consolidation, that time can be devoted to this.
Project data kept in order before being entrusted to a model
An assistant gives back what it receives. If a project manager copies from one month to the next a status that is no longer accurate, the synthesis will be fluent and false. If a milestone reached means delivered for one project and accepted for another, consolidation will add up different things under a uniform presentation. Common definitions, the single project record and the update frequency come before the tool: these are the prerequisites described on our page on setting up a PMO.
Confidentiality is settled before the first trial. Project data contain budgets, supplier names, and personal data as soon as effort is charged to a named person; the GDPR, Regulation (EU) 2016/679, then applies. With an external service, the points to read in the contract are the reuse of the data entered, the place of processing and the retention period. In the document collection, access rights must be reflected in the answers: a project manager should not obtain through the assistant the arbitration file of another project. Usage rules (authorised tools, data not to be entered) fall under AI governance.
Project tracking must not slide into evaluating people
From flagging a late milestone to ranking project managers according to their delays, the technical step is short, and it changes the purpose of the system.
Regulation (EU) 2024/1689, known as the AI Act, is read in its consolidated version of 27 July 2026. Its Annex III, to which Article 6(2) refers to designate high-risk systems, covers in particular, in its point 4(b), AI systems intended to be used to “monitor and evaluate the performance and behaviour of persons” in contractual work-related relationships. Article 6(3) provides for derogations, which it rules out in the case of profiling, and the classification of a system is a matter for the organisation and its legal advisers. The application timetable appears in our article on the AI Act’s obligations for businesses.
For a PMO, the practical consequence is to write down the intended use of the tool: it concerns projects, its outputs serve to prepare committees, and they do not enter into the assessment of people. This boundary also serves the PMO: project managers who suspect that their reports feed into their evaluation may write them accordingly.
A first trial run in parallel, over one committee cycle
We suggest a cautious method: choose a single task and run it in parallel for one cycle. The PMO prepares its file as usual. The assistant produces its own from the same documents. A person compares the two and notes the cause of each difference: missing data, vague definition, sentence invented by the model, omission by the PMO itself.
Document search and consolidation lend themselves to this trial, because their result can be checked against a source. Synthesis can come next.
The record of these differences points to what comes next: the headings whose meaning varies from one project to another need to be redefined with the project managers before extending assistance to a second task.