Originally published in November 2022 · updated in October 2026
Uffective has used machine learning to steer project portfolios since 2019. Along the way, our models have produced hundreds of thousands of predictions. Some models surprised even us, others failed spectacularly. We share both here, because both show what AI in project portfolio management can do today and what it cannot.
What AI can do for the project portfolio
In practice, AI in project portfolio management means one thing above all: making problems visible early, before they show up in a status report. Machine learning models can calculate probabilities for every ongoing project, for example for:
- a budget overrun,
- late delivery,
- cancellation of the project on the current roadmap,
- a value contribution far below the business case – such as less than 10% of the expected value.
The missing piece since Markowitz: risk
Ever since Nobel laureate Harry Markowitz published modern portfolio theory, one rule has applied: a good portfolio maximises return for a given level of risk. Markowitz showed this through the efficient frontier, which maps expected return against risk. The principle was carried over from shares to projects. In project portfolio management, however, the efficient frontier is hardly known today.
The reason lies in the second input: risk. Many organisations struggle to quantify the risk of a project investment, or have given up. More than 50 years after Markowitz, machine learning finally makes this risk predictable. That makes the efficient frontier usable again in the project portfolio.
How to recognise a good model
A model that predicts, say, the likelihood of a budget overrun produces a value between 0 (no overrun) and 1 (overrun will occur) for each project. The key question is: can I trust this value? If a project scores 0.8, will it really need more budget?
How well a model discriminates is described by the ROC (receiver operating characteristic). It usually lies between 0.5 and 1:
- 0.5 is like tossing a coin,
- from 0.8 a model counts as good,
- above 0.9 as very good.
The models we built in Uffective ranged from 0.6 to 0.9.
One failure and one success
The failure (ROC 0.6). An agile Dutch software company with development in four countries releases new software almost daily. After optimising its regression testing with Uffective, it asked for a model that predicts which new releases will cause customer-facing issues. The model only used metadata: who specified the user story, who wrote the code and how long it took. With an ROC of 0.6, it did not help the customer focus on the riskiest releases. The lesson: metadata alone is not always enough. Better candidates are features such as interruptions while programming and properties of the code actually written.
The success (ROC above 0.9). For a large telecom operator, we built a model that predicts the likelihood of project cancellation. After tuning, it reached an ROC above 0.9. Management and project leaders use its output for decisions and scenario analysis.
How to feed a model
Project risk has several dimensions. The most basic one is cancellation, because a cancelled project never delivers the expected value. Budget overruns and delays also reduce value. The often neglected, perhaps most important risk: finished projects that do not deliver the value that was promised.
We usually start with a model for cancellation risk. These models use 20 to 40 input features, such as budget size, expected value and time to market, but also the roles involved. They are normally trained with more than 1,000 rows of data, although we have had successful models with as few as 600 projects.
Three lessons from practice:
- The start month matters. In some companies, projects launched during the annual budget cycle are cancelled more often. Including the start month as a feature improves the model.
- One model per maturity stage. The ROC improves when separate models are trained per project phase, for example one model for all projects before the first gate, one before the second gate, or before and after the first sprint.
- The sentiment of the status report. For more mature projects, features such as days on hold and the time needed to resolve a red signal are added. In advanced environments, we also analyse the sentiment of the project leader’s weekly status text.
What surprised us most: the models with an ROC above 0.9 did not assess the content of the projects at all. The metadata alone had remarkably strong predictive power.
How organisations use the predictions
Scenarios for project leaders. Project leaders see the prediction for their project and can run scenarios by changing input features: staffing the team differently, or rerunning the model once open red risks are closed. A feature importance view shows which factors lower or raise the cancellation likelihood the most. A side effect: data quality improves, because cancelled projects usually have poor data too.
A second yardstick for decisions. Boards that decide on projects and resources get a second dimension next to expected value: the predicted risk. This makes prioritisation and gate decisions more rational. More on this in our podcast episode Improving Gate Decision Making Rationality with Machine Learning.
The moment of truth in management. It gets interesting when the list of projects most likely to be cancelled is presented to senior management. For the first projects on the list, management usually nods: we see it the same way. That raises the obvious question of why not release the team earlier and avoid further sunk cost. Further down, however, pet projects appear too. The typical reaction: leadership joins the steering committee or intervenes in other ways. For the model, this is a new, unknown variable, and its accuracy drops. Our customers are fine with that, because their goal is not the highest possible ROC, but better project results.
What this means for utilities and grid operators
The examples above come from software and telecom portfolios. The principle transfers directly to investment portfolios in critical infrastructure. Grid operators and utilities now steer hundreds to thousands of measures at once: grid expansion, asset replacement, smart meter rollout. The scarce resource is rarely money, but delivery capacity: planning engineers, field crews, outage windows.
Spotting cancellation, delay and budget risks early means putting capacity where measures will actually be delivered. Those who also capture the condition and risk of their assets systematically, for example with enterprise asset management, combine asset risk and project risk in one portfolio. How investments can be steered fundamentally differently than through annual project-by-project approval is described by Uffective CEO Dr. Mark van der Pas in his article “A New Paradigm for IT Budgeting” in The CIO Magazine. What it means for utilities: CAPEX Planning: The Most Expensive Habit in Critical Infrastructure.
Frequently asked questions
What does AI in project portfolio management mean?
Usually machine learning: models that learn from completed projects and predict probabilities for ongoing ones, for example for cancellation, delay or budget overrun.
How many projects do you need for a reliable model?
Typically more than 1,000 rows of data, although we have had successful models with around 600 completed projects.
How can you tell whether a model is good?
By its ROC. 0.5 is like tossing a coin, from 0.8 a model is good, above 0.9 very good.
Does AI replace the steering committee’s decision?
No. The prediction is an additional, objective yardstick next to expected value. People still decide, just on a better basis.
Read more in the series
- What is Machine Learning?
- What are Similarity Predictions?
- Watermelon Projects and Machine Learning
- How Machine Learning Can Solve the Biggest Problems in Your PMO
- Why is our AI called Sandra?
- Use Machine Learning with Scrum to Plan Your Resources Better
- Why You Should Give Machine Learning a Seat at the Table
See your portfolio through a different lens
What would cancellation and budget risks look like in your portfolio? We will show you in a demo based on your own questions. Request a demo
