By Terence Tse
“It ain’t what you do, it’s the way that you do it,” says the song, and nowhere is this more true than when implementing AI.
For many companies, implementing AI has been a disappointment in terms of return on investment. Others don’t even do the measurement required to assess the benefits, if any, that their investment in the technology has brought them. This article explores the manifold possible reasons for AI ROI failure and sets out ways and means of avoiding it.
Companies Are Unable to Get ROI on AI Investments
A major study in AI has reported that companies worldwide invested some $582 billion in AI in 2025 alone.1 Strangely, such a vast sum is often spent even when the return on investment (ROI) on AI for enterprises remains stubbornly elusive. A survey of nearly 2,000 companies across 105 countries found that 88 per cent of organizations now deploy artificial intelligence in at least one business function. However, only 6 per cent report that AI has increased their EBIT by more than 5 per cent.2 This result aligns with another study of 300 AI deployments, based on 52 interviews and 153 survey respondents: 95 per cent of generative AI pilots show no measurable P&L impact whatsoever.3 A further report examining the experience of 1,250 executives arrives at an almost identical finding from a different angle: just 5 per cent of companies qualify as genuinely “future-built,” generating compounding returns from AI.4 Another recent study focusing on financial services found that only 14 per cent of respondents currently view AI as transformational for their organizational strategy and as conferring competitive advantage—signaling a deep execution gap between early-stage experimentation and institution-wide AI integration.5
With returns so hard to obtain, it is no wonder companies have begun to decouple investment from near-term returns; more than 90 per cent of companies surveyed in a recent study plan to continue investing in AI even if it doesn’t deliver immediate returns.6 So, why is it so difficult to achieve ROI in AI? The problem does not lie in the AI technology. Instead, the evidence available so far has pointed to five reasons that have much more to do with the businesses themselves.
Reason 1: The Wrong Clock
One reason arises in the business plan when an AI program is evaluated using the same financial lens as a conventional software deployment. The first problem is the frame of reference. A survey of 1,854 senior executives found that a typical AI use case takes two to four years to deliver a satisfactory ROI,7 about three to four times longer than expected for traditional technology investments, which typically only take seven to 12 months.8 Indeed, AI projects cannot be rushed; only 6 per cent of respondents saw returns within a year, and even among the most successful deployments, just 13 per cent achieved payback within 12 months.9
A typical AI use case takes two to four years to deliver a satisfactory ROI, about three to four times longer than expected for traditional technology investments.
Sometimes organizations cancel AI programs at the very moment they are poised to deliver. A recent survey of 782 infrastructure and operations leaders reveals a pattern: 57 per cent of those who had experienced at least one AI project failure said that it was due to their expecting “too much, too fast.”10 Yet, AI investments could be sunk costs that represent learning opportunities. In an in-depth report, 61 per cent of respondents had previously failed at least one AI project before achieving their current success.11
The growing interest in adopting agentic AI could further exacerbate this problem. Organizations have therefore begun to adjust how they measure ROI across AI technologies, using different timeframes or expectations for generative and agentic AI initiatives. For generative AI, ROI is most often assessed in terms of efficiency and productivity gains. For agentic AI, measurement is likely to focus on cost savings, process redesign, risk management, and longer-term transformation. Put differently, GenAI should be measured on a shorter-term impact basis, while agentic AI should be measured on a longer timeline, as it involves a more ambitious transformation. Hence, it’s no surprise that, among the respondents in a recent study already using agentic AI, just 10 per cent said they are currently realizing significant ROI from agentic AI.12
Reason 2: The Non-Existent Baseline
AI adoption must be measured. This isn’t just a best practice. It’s becoming an existential requirement for enterprises undergoing AI transformation. Yet, surprisingly, some organizations don’t measure what they get in return—at all. A report suggests that while 89 per cent of enterprises have adopted AI tools, only 23 per cent can accurately measure their ROI.13 Indeed, many companies deploy AI without first establishing a pre-deployment baseline. The result: there is no counterfactual or basis for comparison. This may seem a rare unintentional omission. Yet, in a survey of 1,006 C-suite executives across 11 countries and 32 industries, only 11 per cent of companies assess the value of AI for all use cases both before and after implementation. A further 32 per cent assess value only before implementation, but never validate outcomes.14
And this does not include the so-called “shadow AI”, which refers to employees’ unauthorized use of AI tools without formal IT approval or security oversight, and is not measured, if measuring is possible at all in the first place. A previous study found that 67 per cent of enterprises admit to lacking complete visibility into which AI tools their employees use.15 Shadow AI is not progress. If anything, it could prove damaging for companies; among employees using unapproved AI platforms, 57 per cent admit to entering sensitive company information.16
Without proper measurement, an initiative cannot demonstrate value, even when it is working. The most important investment a company makes in AI is not in the model or the solution; it is in the measurement that underpins them.
Reason 3: The Data Debt
When AI systems underperform, a knee-jerk response is to whip the model into shape. However, the problem often doesn’t lie in the model. It is the data. A survey of 325 senior leaders offers insights into the data debt problem.17 It found that 94 per cent believe that well-connected data, processes, and applications are crucial to successful AI adoption. Nevertheless, only 27 per cent say those elements are currently well connected within their organization. Worse yet, while 65 per cent of respondents say their structured data is at least partially AI-ready, only 39 per cent say the same about unstructured data (i.e., the emails, PDFs, call transcripts, and institutional records that underpin most day-to-day business processes). Furthermore, over half of the organizations cited data silos as their most common barrier to AI adoption.
The most important investment a company makes in AI is not in the model or the solution; it is in the measurement that underpins them.
Two other surveys corroborate these observations, with one reporting that data quality is among the most frequently cited barriers and opportunities. Thirty-six per cent of organizations identify improving data quality, integration, and system interoperability as their greatest opportunity to extract more value from AI in finance.18 Another study found that data availability and quality remain the primary pain points hindering AI adoption, while talent is an even bigger barrier.19
Reason 4: The Wrong Order: Technology First, People Last
In our recent book, Becoming AI Native: Charting the Next AI Frontier, we have found that workforce capability is another significant barrier, if not a bigger one than data debt. Unsurprisingly, then, the single most reliable predictor of AI value creation is the ratio of investment in people and process to investment in technology. It has been reckoned that the 5 per cent of companies that derive substantial value from AI allocate roughly 70 per cent of their AI resources to people and process redesign, with only 20 per cent to technology, and 10 per cent to algorithms.20 By contrast, most failing organizations invert this ratio almost exactly; analysis suggests that 93 per cent of enterprise AI budgets flow to technology, leaving just 7 per cent for the organizational layer—training, change management, governance, enablement, and workflow redesign.21
Recent research also shows that companies are much better off being people-focused than technology-focused. For instance, it was expected that by 2027, companies that emphasize AI literacy among executives would achieve 20 per cent higher financial performance than those that don’t.22 Companies that prioritize human-machine collaboration by redesigning roles, processes, and operating models are significantly more likely to realize measurable returns from AI than those taking a technology-first approach.23 Companies that double down on nurturing AI talent are therefore well rewarded. Companies that realize significant AI value plan to add three times as many full-time equivalents for AI upskilling as laggards and are twice as likely to involve employees directly in shaping adoption.24 In short, a key to getting value from AI investments is “humans × machines,” not “humans + machines.” The challenge is not adopting AI but evolving alongside it.
Reason 5: The Wrong Ownership
Value does not flow without ownership, and ownership requires explicit assignment. The question is not whether, but to whom ownership should be assigned. A recent survey found that most organizations assign accountability for AI value to the chief data / analytics / AI officer or to individual functional executives. Yet, this could be a mistake, as the survey also found that when the CFO is responsible for AI value delivery, 76 per cent of organizations report achieving a good amount of value. Under CIOs or CTOs, that figure falls to 53 per cent. Under functional executives, it drops to a mere 32 per cent.25 This should not be a surprise, given that the role of the CFO is evolving into that of an AI-driven transformation orchestrator, balancing finance, technology, and strategy while upskilling teams, managing risks, and driving measurable business value.26
As mentioned earlier, 61 per cent of the companies have failed in AI experiments before achieving their current success. This is partly due to misaligned ownership; teams initially treated AI as a technology project rather than as a process and change-management project. First attempts failed because they were applied to broken workflows and led by technical teams without business ownership.27
The Pattern Is Clear. So Is The Path Forward
The reasons for failure outlined here—mismatched timelines, absent baselines, data debt, the budget inversion away from people towards technology, and wrong ownership—are not random. Organizations that derive the most value from AI are not necessarily those with the most sophisticated technology. They don’t treat AI as a technology program but as an organizational redesign. By aligning incentives, redesigning processes, and reconfiguring organizational power, firms turn AI adoption into lasting business value. Those that see AI merely as a technical upgrade will inevitably fall short.
About the Author
Terence Tse is Professor of Finance at Hult International Business School and co-founder at the AI Native Foundation. He is also co-founder and Executive Director of Nexus FrontierTech.
