The AI project that never ships rarely makes headlines, but it is becoming one of the most expensive habits in enterprise technology.
A new study by GFT Technologies and Wakefield Research, which surveyed 945 CIOs and CTOs across 19 countries at organisations with annual revenues of at least $500m, finds that 84% have cancelled at least one AI pilot or project because of limitations in their legacy systems. For 22%, it has happened more than once. Almost all respondents (95%) say legacy technology delays their ability to deploy and scale AI, and 56% describe that delay as moderate or major.
The picture is of a technology industry that has bought the engine before checking the chassis.
Not an isolated finding
The GFT numbers fit a pattern that analyst firms have been describing for some time. Gartner forecast that over 40% of agentic AI projects will be cancelled by the end of 2027, owing to escalating costs, unclear business value or inadequate risk controls. Its analysts have also warned that integrating agents into legacy systems can be technically complex, often disrupting workflows and requiring costly modifications.
IDC has put a number on the cost of neglect. It predicts that CIOs who fail to address and remediate their growing data debt will face a 50% higher failure rate for their AI initiatives. That research, a MongoDB-sponsored study of 1,400 IT leaders in Asia-Pacific, also found that 89% of organisations acknowledge technical debt as a major obstacle to modernisation. Taken together, the message is that failure is not mainly about model quality. It is about what the models are plugged into.
Ambition is outrunning readiness
Confidence in the economics is wobbling too. Eighty-nine percent of respondents are concerned that global AI investment is growing faster than the business value it can realistically deliver, and 44% are very or extremely concerned. These are the people signing off the budgets.
The modernisation figures explain some of that unease. Only 15% of organisations describe their legacy modernisation for enterprise AI as complete or nearly complete. Another 49% have started and say they are on track, while 28% have begun but believe they are falling behind. Around 9% have not started but intend to, and 1% have neither started nor planned to.
A pilot can succeed on a clean dataset and a friendly integration. Enterprise-scale value needs reliable data, secure applications and systems that can talk to each other. Where those are missing, the pilot stalls, the business case weakens and the next AI proposal faces a more sceptical board. It is a cycle that feeds itself.
A security crisis waiting for a trigger
The security implications are stark. Ninety-three percent of executives believe that running AI on legacy systems without modernising them first could eventually trigger an enterprise-wide security crisis.
The concern grows as AI moves from generating answers to taking actions. An agent that can reach enterprise applications, databases and workflows needs far tighter controls than a chatbot, and it inherits every weakness of the systems it touches. Connecting a new, autonomous layer to old, poorly understood infrastructure multiplies the places where something can go wrong.
What makes this harder is that only 20% of respondents say their fellow C-suite executives and board members fully understand the risk. Technology leaders are being asked to speed up adoption while carrying a warning that most of their peers have not yet heard. For a CIO, part of the modernisation task is therefore persuasion: translating technical debt into a risk the board recognises, before an incident does it for them.
Geopolitics and regulation are redrawing the map
Legacy systems are not the only brake. Half of respondents (50%) say geopolitical developments have led them to limit where they deploy AI, 34% have reduced planned investment and 28% have cancelled projects. More than half (54%) say regulatory uncertainty has made them more cautious.
For multinationals, this changes what architecture means. Data sovereignty, compliance and model governance now sit alongside performance and cost. The question is no longer only which model to use, but where it can legally and securely run, and that is far easier to answer on infrastructure that has been modernised, documented and understood than on a tangle of inherited systems.
The workforce story is messier than the headlines
The study also challenges the idea that AI simply replaces people. More than 90% of respondents believe some public companies are citing AI to justify workforce changes aimed mainly at lifting their share price. In practice, 51% have hired staff specifically to review or correct AI-generated work, and 26% have rehired people they had previously let go.
The lesson is that automating tasks is not the same as removing the need for expertise. Organisations that cut deeply on the strength of a pilot are discovering that someone still has to supervise the output, handle the exceptions and answer for the decisions. The CIOs who plan for those roles at the outset will avoid the expensive reversal.
The personal stakes are high. Nearly 90% of respondents worry that a wrong workforce decision while scaling AI could put their own jobs at risk.
The race nobody advertised
The conclusion running through the data, and echoed by Gartner and IDC, is that the winners of the next phase will not necessarily be those with the biggest model budgets. They will be the ones that have done the unglamorous work: retiring or rebuilding fragile applications, cleaning up data, tightening security architecture and tying every AI investment to an outcome the business can measure. Gartner’s own advice leans the same way, pointing organisations towards rethinking workflows from the ground up rather than bolting agents onto old processes.
The AI race, in other words, is turning into a modernisation race. For the 85% of enterprises not yet nearly finished, the quickest route to the AI future may start with fixing what they already run.