Why Most AI Programs Still Struggle to Achieve ROI

Mitch Kwiatkowski, Techstra Solutions’ Head of Data & AI, has a proven track record of delivering measurable ROI from data and AI initiatives. Here he answers a question many organizations are still trying to answer:

Why is AI ROI so difficult to achieve?

Over the last few years, global companies have poured trillions of dollars into AI investments. Gartner forecasts more than $2.5 trillion will be spent on AI in 2026 (a 47% increase over 2025) and $3.3 trillion in 2027. Despite that investment and a near-universal use of AI (88% of companies state they use some form), only 39% can track any sort of financial value. Worse yet, one study reports that 22% of agent deployments result in negative ROI. (PwC 2026 CEO Survey)

As we’ve seen with so many technology advancements over the decades, failure to transform is rarely because of the technology itself. AI projects often fail to get out of pilot or POC and into production because of people, processes, or unclear goals. When we fail to understand value, the issue may simply be the way we’re measuring the return on our investment.

We Measure the Wrong Things

If you home in on the measurement component of value, you are likely to find that most AI programs are looking at model accuracy, adoption rates, or the number of use cases in flight. These represent activities, not outcomes or value. I’ve lost count of how many times I’ve heard a data scientist tout a model’s 95% accuracy while it fails to demonstrate any business impact because it’s not tied to a process that moves revenue, cost, or risk.

ROI conversations stall because the “R” was never defined in business terms.

We Never Established a Baseline

You can’t validate an improvement if you never measured the starting point. Teams will launch a pilot, and, if they’re lucky, declare success based on qualitative feedback. Six months later, a leader is asked to justify the spending, and they struggle because there wasn’t a measurable starting point from which to compare or a successful method of measuring. According to a recent survey, “only 15% can calculate AI ROI without significant bottlenecks”. (Hubley 2026)

Without a baseline and validated method to measure, a ROI claim is little more than an opinion.

We Chase the Exciting Use Cases Rather than the Valuable Ones

Leaders love flashy, high visibility use cases because they get attention. Teams love them because they get to play with new tools on the bleeding edge. Executives no longer feel weighed down by FOMO. Unfortunately, these use cases don’t often come with a transaction volume, cost structure, utilization, or impact that generates meaningful returns. The better candidates are usually more mundane targets like high-frequency, high-cost, repeatable processes where even modest improvements can compound into real dollars.

AI programs that lead with excitement over economics will wind up with a portfolio of interesting solutions stuck in pilot and no P&L impact.

We Underestimate the True Cost

The development cost of a model is typically a fraction of what it costs to run it. Data collection, preparation, integration, governance, monitoring, retraining, and change management don’t usually make it into a business case. When those costs surface later, they erode the ROI story. AI may not have underperformed, but the total cost of ownership exceeded what was anticipated because those other factors were never fully scoped.

This is evident in a trend that has emerged. Over the last 12 months, 79% of enterprises experienced AI cost overruns and 80–85% of enterprises miss their AI infrastructure forecasts by more than 25%. Even organizations who consider themselves mature in FinOps have overshot budgets by an average of 30.9%. (Hubley 2026)

Total cost of ownership that treats data and run-and-operate as an afterthought will overestimate the return of an AI initiative from day one.

The Fix Starts Before Building

There are different approaches to identify, estimate, and achieve value in an AI program. No matter what, every use case should be tied to a defined metric and a dollar figure before any code is written. This can be a challenge for organizations who are not used to financial analysis and project justification, but given the high investment in AI, doing anything less could be considered financially irresponsible.

ROI tracking starts at the beginning and should never be treated as a retrospective exercise. Ideally, it is an IT or AI governance requirement that is reviewed in the same cadence as other key project roadmaps. Programs that can answer the ROI question are the ones that treat value measurement as a core component of an initiative rather than an afterthought.

Fix that and the returns will follow.

About the Author

Techstra Solutions’ Head of Data & AI, Mitch Kwiatkowski, is a recognized leader in enterprise transformation. Over the course of his 20+ year career, Mitch has led data and AI initiatives that have delivered $250 million in measurable value. His expertise includes data and AI governance programs, enterprise data strategy and architecture, data product development and commercialization, operating model and portfolio optimization, and digital ethics and risk management. Mitch is passionate about responsible AI adoption and helping organizations move beyond pilots and proofs of concept to delivering sustainable value at scale.