AI in Portfolio Management: What 300 PMO Leaders Say About AI Agents
AI in portfolio management has left the pilot phase, at least on paper. A new survey of 300 senior project and PMO leaders finds that 91% are already piloting or using AI in project delivery. That sounds like a revolution. A closer look, however, shows a quieter picture: lots of tools, little clarity, and a surprising number of invoices nobody can read.
For IT project managers and PMO leads, this matters now. AI agents are moving from demos into real delivery work, and your portfolio will feel it first. In this article, we go through what the data says, where the friction sits, and what you can do in 30 days. Spoiler: the biggest problem is not the model.
01 – What the survey says about AI in portfolio management
The numbers come from the 2026 State of AI in Portfolio Management report by Tempo Software. The fieldwork was run by Potloc in June 2026. It covers 300 senior project, portfolio and PMO leaders at enterprises in North America and Western Europe. Many highly regulated industries are not included.
Note one more detail. Tempo sells portfolio software, so the survey also serves as a product argument. That does not make the numbers wrong. Still, read them as a snapshot from one vendor, not as a law of nature.
The headline figures are easy to remember:
- 91% are piloting or actively using AI in project delivery. Another 7% plan to start soon, and 2% have no plans.
- 47% actively use AI tools, but only 33% hand real delivery work to AI agents.
- 33% use AI to identify risks and flag issues, while 26% use it to prioritise or reprioritise work.
In other words, almost everyone is trying AI. Far fewer let it do anything that matters. The word “adoption” covers both, which is a generous use of the word.
02 – Using AI is not the same as delegating to it
The gap between 47% and 33% is the interesting part. A tool that drafts a status text is a helper. An agent that updates tasks, flags risks and reprioritises work is closer to a team member. Therefore, it needs the same things a team member needs: a role, boundaries and someone who checks its output.
The respondents agree that this is hard. The top challenges they named were:
- Quality and oversight of AI output (43%)
- Integrating AI into project management processes (38%)
- Governance, compliance and risk (36%)
None of these is a model problem. They are process problems, and process problems belong to us. If you want a deeper look at the governance side, read our article on why governance in AI projects must start earlier than in classic IT. The pattern is the same here: late rules are expensive rules.
Interestingly, the report links agents in production to better delivery. Teams that run agents finish more projects in under six months, a nine-point advantage. However, this is a correlation from a survey. It does not prove that agents cause faster delivery. Good teams may simply adopt tools earlier.
03 – The costs nobody can see
Here the survey gets uncomfortable. According to the report, 42% struggle to tie AI spend to ROI, and 39% cannot see their AI costs at all. In addition, 46% say teams duplicate work with AI. Put differently, many organisations pay for a tool they cannot measure, to do work somebody else already did.
Visibility also depends on your role. The report says 54% of individual project managers cannot see AI costs. Among portfolio managers it is 36%, and among PMO leaders 28%. The closer you sit to the budget, the better you see the bill. That is not a surprise, but it is a useful hint on where to fix the problem.
Attribution is a second blind spot. Of the teams actively using AI, 55% struggle to tell AI work from human work. For teams with agents in production, the figure falls to 16%. Apparently, running agents seriously forces you to count them. Meanwhile, teams that just “use AI” keep guessing.
An agent that nobody can see, count or bill is not a digital colleague. It is a very quiet cost centre.
04 – Where agents run, and where they do not
Adoption is far from even. The report finds that 40% of North American firms have AI agents in production, against 21% in Western Europe. By industry, 47% of software companies run agents, while only 5% of construction and engineering companies do.
What does this mean for you? First, do not copy a software company’s playbook if you manage physical projects. Second, do not assume you are late. In many sectors, nobody has a long head start yet. Only 18 of the 300 respondents have all seven surveyed AI capabilities live and agents in production. That is a small club, and it is not accepting applications on a large scale.
Those 18 report fewer pain points. For example, 17% cannot tie AI spend to ROI, compared with 39% of the other group. Still, with 18 people the sample is tiny. Treat it as a direction, not as proof.
Another number deserves attention. Only 25% of respondents get early warnings on drift, and only 27% can manage human-AI teams. If your portfolio already struggles with reporting, our article on why IT project status reports drift from reality shows how small gaps grow over time.
05 – Data comes first
Agents are only as good as the data they read. A guide for PMO leaders by Cora Systems cites Gartner research on this point. According to that guide, 57% of CIOs say data preparedness is a significant issue or a dealbreaker. It also reports that only 14% of IT leaders are highly confident that their data is properly governed for AI.
Both figures are second-hand, so check Gartner’s original report if you plan to quote them. The message is still plausible. Project data is often messy: outdated charters, inconsistent budget formats and risk logs that last saw a review in spring. An agent will read all of it with great enthusiasm and no scepticism.
If you want to start with tools, our overview of AI tools every IT project manager should use gives a first shortlist. Just clean the data before you connect anything.
06 – A practical plan for the next 30 days
You do not need a transformation programme. Instead, you need a few clear steps with owners. The following plan fits most PMOs:
- List every AI tool and agent that touches your portfolio, including free trials and features inside existing software. Add an owner and a purpose.
- Make costs visible. Collect licence, usage and API costs per tool. Report them next to project budgets, not in a separate corner.
- Tag AI work. Mark which tasks, reports or decisions an agent produced. Without this, attribution stays a guess.
- Define oversight. For each agent, state who reviews its output, how often, and what happens when it is wrong.
- Clean the data. Fix charters, budget formats and risk logs for one pilot portfolio before you scale.
- Set one success measure. Pick something simple, such as the time to produce a status report, and measure it before and after.
Each step is small. Together, they turn AI from a vague promise into something your steering committee can discuss with numbers.
The survey’s message is simple. Adoption is easy, but control is hard. Therefore, start with the boring parts: inventory, costs and ownership. Next week, put one question on your PMO agenda: “Which AI tools run in our portfolio, and what do they cost?” If nobody can answer it in one meeting, you have found your first project.
