By now, almost every enterprise has an AI story.

Years into the AI boom, disappointment has become a familiar refrain.

Only 28% of enterprise AI projects meet ROI expectations, with more than 90% of AI pilots never making it into production.

AI projects stall, returns fail to materialize, and executives quietly conclude that the technology “wasn’t ready”.

That narrative is convenient, but often wrong. So, where does ROI come from?

The problem with half-hearted AI

Most AI failures come from businesses unwilling to change how they work. Many organizations fund pilots, rally up innovation teams and deploy smart tools, but stop short of changing the systems those tools need to operate within.

Teams are given access to AI without being upskilled to use it effectively, and processes designed for human-speed decision making are left untouched, even as machine-speed systems are layered on top.

The disappointment soon follows... isolated productivity wins, no enterprise-level impact and increased skepticism from the boardroom. And yet, walking away from AI isn’t the answer either.

Businesses lose up to 30% of revenue annually due to inefficiencies, precisely the kind of structural waste AI is designed to eliminate. The European Parliament has even explicitly warned that underuse of AI tools could cost the EU its competitive edge and stall economic growth, with estimates showing AI could add up to $4.4 trillion annually to the global economy.

For organizations sitting on the fence, failing to innovate means falling behind, while competitors see compounding gains quarter by quarter as they’ve simply cracked the AI code.

Why the ROI gap is self-inflicted

The real cost, then, isn’t failed pilots, it’s half commitment. Many organizations treat AI as optional or experimental, then act surprised when it behaves that way.

The first step is to stop measuring AI like another IT project; measure it like R&D. Applying traditional ROI metrics to AI repeats the same mistake organizations have made during every major technology shift; expecting immediate returns from AI reflects industrial era thinking applied to a cognitive era transformation.

When email arrived, companies didn’t abandon it because quarterly earnings didn’t spike. AI is similar. The problem isn’t that it doesn’t work, it’s that it’s being evaluated with the wrong lens. CTOs and CFOs should treat AI budgets strategically, similar to how R&D investment may be considered.

While there are some short term savings, the real value lies behind the longer amortization windows, staged milestones and tolerance for early phase losses. This reframes boardroom conversations away from short-term justification and towards strategic capability building.

How can we fix this?

1. Learn from our failure

We encountered this issue firsthand. When we started integrating AI into our own delivery model, productivity pockets appeared, but there was no sustained internal buy-in and no systemic change to support those gains.

What worked was a deliberate, process-led shift. Rather than trying to scale everything at once, we focused on a small set of lighthouse projects backed by lean, cross-functional teams. AI was embedded across the full software delivery lifecycle, with attention moving away from tools and towards how work actually flowed.

The friction was never in the code, but in the systems around the code. How do approval chains work when compliance teams operate on weekly cycles, but AI agents move at machine-speed? How do organizations build reusable, secure components rather than rebuilding from scratch on every project?

Treating AI as part of the operating model, not a collection of tools, is where meaningful ROI starts to emerge. We're now applying the same model with clients, with faster shipping and fewer people.

2. Get ready to dig deep and change your foundations

One of the biggest reasons AI initiatives stall is because organizations simply layer new intelligence on top of operating models that were designed for a very different era. Many enterprises are still structured around legacy systems and linear approval chains built to standardize transactions, not to support real-time judgement.

AI needs context, access across systems and the ability to act. But in many organizations, data is locked inside heavily-customized ERPs, workflows are fragmented by function, and decision rights are buried in handoffs and committees. So before buying the shiny new tool, go back to basics. Fix the foundations.

The sharpest lever is decision rights. AI agents operate in seconds, but most enterprise decisions still route through weekly approval cycles. Until that gap closes, AI speed has nowhere to go. This also means confronting legacy estates honestly.

Not every core system needs to be replaced, but the processes and assumptions built around them may need to change if AI is to operate effectively.

Addressing structural foundations

This is uncomfortable work, but once the structural foundations are addressed - clearer decision rights, less fragmented data, outcomes-oriented teams - AI can dramatically compress the time it takes to rebuild.

This is where AI-native engineering pays off, as it allows businesses to go from treating AI as a feature, to building systems, workflows and organizations where AI is a first-class participant in how work gets done.

Instead of hard-coded logic and rigid process flows, AI native engineering centers on systems that are context-aware, continuously learning and capable of taking bounded action across the stack - from interpreting intent, to orchestrating workflows, to generating and improving code and decisions, in real time.

The competitive advantage comes not from access to better models, which are increasingly commoditized, but from how effectively an organization can embed those models into its operating fabric.

3. Build a portfolio dashboard, not project scorecards

Boards can’t evaluate AI as a portfolio if every initiative reports success differently. A standardized AI ROI framework, tracking utilization, business outcomes and strategic value, allows cross-initiative comparison and portfolio-level decisions.

It’s also important to manage expectations. Most organizations achieve satisfactory ROI on a typical AI use case within two to four years - far longer than the seven-to-twelve-month payback typically expected from traditional tech investments.

That said, where organizations align process, people and governance early, faster transitions are possible – especially with AI native engineering.

4. Put your whole heart into it

AI’s ROI problem is organizational, and leaders serious about this should be prepared to change how their company is set up.

AI doesn't fit existing structures and companies have to change to use it..

The cost of failure is real, but the cost of half-heartedness is higher, and far easier to overlook.

We feature the best robotic process automation software.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit