This week I have been mulling over the Project Management Office (PMO) and where AI could make a useful difference. Years ago when running a portfolio, I had an excellent Power BI dashboard driven by data from Microsoft Project Server, it gave us a consolidated view across the project portfolio, including owners and their portfolios of projects, status and risk information - it was really useful.

However, the dashboard could only report what the projects were telling it. If a project was green, it was green on the dashboard. If the underlying project was actually struggling, with milestones moving, dependencies becoming problematic and resources disappearing, the reporting process might not expose that until somebody changed the status (the PMO and Project Managers worked hard to surface these issues early).

Consider the classic watermelon project though: green on the outside, red on the inside and this is where I think AI starts to make the PMO discussion much more interesting. The opportunity is not simply to use AI to write status reports faster but to use AI to independently examine the evidence behind those reports.

When I consider three important areas of PMO work, they could change with AI: assessing project quality, understanding risk across the portfolio, and supporting Project Managers in the day-to-day project delivery.

What is it?

There is already plenty of fairly obvious AI functionality appearing in project and portfolio management tools, AI can summarize a project, draft a status report, extract actions, answer questions about a schedule and prepare an executive update. Useful, but increasingly fairly standard.

The more interesting capabilities go a step further. ServiceNow ‘s Now Assist for Strategic Portfolio Management, for example, can predict project health independently across schedule, cost, resources and scope, generate an executive summary and provide the rationale behind its assessment.

Planview is taking a similar approach. Its Anvi capability continuously looks for hidden risks and potential roadblocks across organizational data. More interestingly for the PMO, Planview Viz specifically describes an early-warning capability that goes beyond subjective status reporting, analyzing large numbers of data points in delivery pipelines to identify slowdowns and delivery risk.

monday.com’s Portfolio Risk Insights takes data from individual project boards, including dates, status values, updates and activity, and produces a daily portfolio-level assessment of potential risks. It can then generate an executive portfolio report highlighting health, metrics, trends and risk.

This starts to create three different levels of AI involvement.

• Reporting assistance: Tell me what the project says.
• Project analysis: Tell me what the evidence suggests about the project.
• Agentic support: Continuously watch the project, identify something that needs attention, recommend what should happen and, where appropriate, take a bounded action.

That last category is also becoming real. monday.com seems to have agents that can monitor work, make decisions within defined rules and permissions, and perform actions such as updating items, assigning owners and triggering follow-ups. Planview supports pre-built and custom agents. Planisware’s newer Prisma approach can reason over live portfolio data and prepare changes to costs and schedules, but keeps those changes behind an approval gate. Let’s not forget Microsoft in this space with all the AI focused releases from them and the changes across MS Project and Planner.

For me, though, the biggest opportunity is still: Can AI provide a second opinion on whether a project is actually healthy? That is potentially much more valuable than having it write another project update.

What does it mean from a business perspective?

The watermelon problem has always existed because project reporting contains a significant amount of judgement. A Project Manager may have good reasons for calling something green - perhaps recovery action is underway, perhaps a delayed activity has float, perhaps a resource problem will be resolved next week.

But optimism, pressure and simple human bias can also creep in (my weakness in this instance is optimism - that’s why I plan, to temper the optimism. To be honest though, optimism often plays a huge part in projects - otherwise why would we do them).

Traditionally, a strong PMO challenges that through experience. It asks questions, examines the schedule, looks at risk, checks financials and spots inconsistencies, AI potentially gives the PMO another set of eyes.

Imagine a project reporting Green while the underlying evidence says:

• three important milestones have moved in six weeks;
• unresolved dependencies are increasing;
• several critical tasks repeatedly roll into the next reporting period;
• resource availability has fallen;
• risk exposure is increasing;
• scope is still changing;
• delivery throughput is declining.

The AI does not have to declare the Project Manager wrong, much more useful output might simply be: “The reported Green status appears inconsistent with several delivery indicators. Review recommended.”

This is important. I would be uncomfortable with an AI system autonomously deciding the official status of a project. I would be very interested in one that continuously challenges the reporting and tells the PMO where to look, that’s not really an LLM problem - it is a project evidence problem.

A good watermelon detector needs access to more than the status report. Depending on the project, that could include schedule movement, critical path and baselines, RAID information, budget and forecast data, resource assignments, scope changes, delivery-system information from Jira or Azure DevOps, change activity, defects and perhaps even stakeholder sentiment (something we used to survey for as in internal project indicator).

The second opportunity is at the portfolio level - traditional portfolio reporting is very good at aggregation:

• Twenty projects. Four red. Six amber. Ten green.
• Ten high risks. Twelve medium risks.

AI potentially allows the PMO to ask a different type of question - What is happening across these projects that we should care about? Perhaps seven projects have different risks logged, but underneath them is actually the same resource constraint (something we used to look for - the sources of risk, not just the impact of risks). Perhaps several apparently unrelated schedule risks trace back to one shared dependency or a particular vendor appears across four deteriorating projects.

Now the PMO is not simply consolidating project reports. It is looking for patterns across the portfolio which feels like a fairly natural use of GenAI because the task involves synthesizing a lot of semi-structured information and finding relationships that may not use identical language.

Portfolio-level AI is already relatively mature compared with some of the more ambitious project-quality use cases. monday.com’s portfolio capability, for example, creates AI-generated portfolio summaries and risk analysis, while Planview supports cross-portfolio insights and executive summaries.

The third area is Project Manager support. There is a lot of administrative overhead in running a well-controlled project, status reports need producing, risks need reviewing, actions need following up, dependencies need checking., steering material needs preparing and schedules need challenging.

AI can increasingly act like a lightweight project analyst alongside the PM. A Project Manager might ask:

• “Review my schedule and identify dependencies that look weak.”
• “Which risks haven’t been updated recently?”
• “What changed since last week’s steering committee?”
• “Draft my status report, but highlight anything in the project data that contradicts my assessment.”
• “What questions is the steering committee likely to ask me?”
• “Which actions from the last three meetings remain unresolved?”

This is where agent capabilities become particularly interesting. Rather than waiting to be prompted, an agent could run every Thursday, review the project, identify anomalies and prepare a briefing for the PM before the weekly status cycle.

Don’t start by allowing the agent modify your schedule and re-baseline your project - there is a useful progression to consider here:

Observe → Analyze → Recommend → Prepare → Act

The governance requirement increases as we move to the right. Preparing a draft risk for a Project Manager to approve is quite different from changing the official project forecast. Planisware’s use of approval-gated changes is a sensible example of that distinction.

The technology underneath this is changing quickly as well. MCP support is appearing in project and work-management platforms and gives enterprise AI agents a more standardized way of accessing project data and tools. Agent-to-agent interoperability is earlier in its development. For PMOs, though, I would treat these as architectural enablers rather than reasons to buy a particular product.

The business question is still: what useful PMO capability does this create?

What do I do with it?

I would start with project quality rather than trying to create an autonomous AI PMO.

• Define what project health actually means. Many organizations use Red, Amber and Green without being particularly precise about the evidence behind them. Establish the indicators that should cause a PMO to question reported health.
• Build a project evidence model. Identify the information that tells you how a project is really behaving. Schedule variance, milestone movement, dependency status, resource availability, budget variance, scope movement, unresolved risks and delivery performance are obvious candidates. The exact measures will vary by organization and project type.
• Compare reported health with evidence-based health. This, to me, is the interesting experiment, don’t initially ask AI to determine project status - ask it to identify discrepancies which becomes a PMO review trigger rather than an automated management decision:

Reported status: Green. Evidence-based assessment: Amber. Reason: repeated milestone movement, increasing dependency risk and reduced resource coverage.

• Test it against historical projects. This is important - a slick demonstration using today’s portfolio proves very little. A historical test tells you whether the capability might actually be useful.. Take projects where you know the eventual outcome and go back through the data. and ask:

Could the AI have identified deterioration earlier? How much earlier? What did it miss? How many healthy projects did it incorrectly flag?

• Then move up to portfolio analysis. Give the AI access to risk and status information across projects and ask it to identify common themes, systemic risks, dependencies and contradictions. I would still want every significant conclusion to be traceable back to source information. An executive summary saying “resource capacity is the portfolio’s largest delivery risk” is considerably more useful if I can click through and see exactly which projects and resource constraints led to that conclusion.

• Use AI to make the Project Manager stronger. Automating report preparation, follow-ups, analysis and meeting preparation could remove quite a lot of administration. That should give the PM more time for the things that remain very human: stakeholder management, negotiation, leadership, judgement and making difficult trade-offs.

• Keep consequential actions controlled. Let an agent monitor continuously and identify problems. Let it draft reports and remediation actions. Be a lot more careful if you let it start changing baselines, financial forecasts, resource commitments or official project status.

The PMO software marketplace has moved quite quickly. Summarization and report generation are already becoming routine. Portfolio-level analysis is increasingly credible. Proactive project-health analysis is emerging and, for me, is one of the most interesting areas to watch. Fully autonomous project and portfolio governance is considerably less mature.

The point here is not that AI is going to run the PMO. It is that for the first time we can realistically give the PMO a system that does more than faithfully reproduce the information projects submit - it can question it, and perhaps that is where some of the real value sits.

The next generation of PMO dashboard may not simply tell us that a project is Green. It may be able to answer: “Are you sure?”