Scaling AI Responsibly: Why CIOs Must Look Upstream to Secure Project Success

AI's Failure Problem Starts Before the Project Does

Most AI projects don't fail because of the model. They fail because the foundational capabilities were not in place. Three separate research papers published in the last six months back this up.

Gartner found that only 28% of AI use cases in IT infrastructure and operations met ROI expectations, with 20% failing outright.¹ A separate Gartner survey reported that 39% of technology leaders are confident of their current AI investment improving financial performance.² And BCG saw 60% of companies generating no material AI value at all.³

Three sources using three different methodologies all point in one direction. The first two look at infrastructure and analytics functions; the third covers the enterprise broadly. AI project failure isn't unpredictable, nor is it established by the model you choose. It's about what happens upstream of the project.

What Separates a Project that Ships from One that Stalls

The model, the vendor, and change management are usually blamed for AI projects failing to meet expectations. But in our experience, the root cause lies in the foundations laid before the project begins. We call this unified portfolio management (UPM): the six portfolio functions stand on — process, tools, budgeting, governance, decision criteria, and policies. Projects fail when two of them, process and tools, remain manual while the project needs them to be standardized. (New to the UPM series? Here's a refresher on our pillars and maturity levels.)

In all likelihood, the model was fine; the strategic framework lacked coherence. The project's master data contained duplicates and orphans that no one had cleaned for two years. There was no accountable human owner with decision-making authority - only a sponsor and a vendor. By the time the steering committee reviewed the output, the underlying data had drifted, the recommendations were outdated, and the visible hallucinations had already set off alarms.

That's what a weak foundation looks like without a framework as a guardrail. This is a process and tools failure, not an AI failure. Each one points to a specific gap that can be measured and solved without AI expertise.

Most AI project failures trace back to weak processes and tooling, not the model. Resolving them needs stronger governance, not AI expertise.

The Cracks that Show Up Before Launch

Three diagnostic signs show up repeatedly. Together, they predict the failure rate of the next project more reliably than the choice of vendor or model:

• The taxonomy isn't stable throughout the project. Strategy changes midway through a portfolio re-segmentation, or cost-center mappings are revised for next year's budget. The project trains on one taxonomy but reports against another.
• The data (the project needs) isn't deduplicated or up to date. Project records aren't single-sourced. Master data is updated quarterly, but it should be updated continuously. The signal the project produces is also not actionable.
• The success metric is an artifact, not a decision. A dashboard. A report. A summary deck for the steering committee. Nothing in the project specification says that this succeeds when the portfolio committee makes decision X by date Y, and that decision wouldn't have been made without AI.

None of this is unusual. Gartner found that successful AI initiatives invest up to four times more (as a percentage of revenue) in foundational areas such as data quality, governance, AI-ready people, and change management than those that see poor outcomes². The weak foundation isn't the exception. It's the median.

The Pre-Funding Checklist

In the first blog of this series, we shared three questions for the standing portfolio function. These are the ones to run on a Monday morning before you fund a specific project: three yes-or-no questions, each mapped to a different pillar. If the answer to any is no, you'd be better off getting the foundation fixed.

• Process: Is your strategic taxonomy stable for at least the duration of the project? If not, the project will produce a signal that drifts before the steering committee can act on it.
• Tools: Will the project read from systems of record with deduplicated, current data? And is master-data ownership clearly defined? If the answer is no, the project is producing the average of your noise.
• Decision Criteria: Does the success metric name a real decision with an owner, a date, and what the committee would have decided without the AI? If it doesn’t, you've funded a demo, not a project.

Together, they take about five minutes per project proposal and map directly to three of the six pillars. They fail to identify every failure mode. But they expose the ones that repeatedly show up in our consulting practice.

If you've funded a dashboard instead of a decision, you've funded a demo, not a project.

The Fix Isn't to Slow Down

In June 2025, Gartner forecasted that more than 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value, and weak risk controls.⁴ That number gets read as a warning about AI. It's a UPM diagnostic problem.

Enterprises are funding advanced ambitions on manual groundwork — Level 4 goals on Level 2 plumbing. The answer isn't to slow down or spend less. It's to score the foundations honestly, lift the weakest one, and redirect the budget at the bottleneck instead of the surface.

The next piece in this series turns to the tools pillar and reads the strategic portfolio management magic quadrant as a CIO would, not as a buying guide. Same evidence-led approach, same UPM lens.