Companies are spending heavily on Artificial Intelligence (AI), but many projects never become part of daily operations. In 2025, S&P Global Market Intelligence found that 42% of organizations investing in AI had abandoned most of their projects before production, up from 17% a year earlier. On average, they scrapped 46% of AI trials before broad adoption. An IBM survey of 2,000 CEOs found that only 16% of AI initiatives had scaled across an enterprise.
Some projects fail because the AI performs poorly. Many promising trials, however, stall because the company’s existing software, data, and working practices cannot support them.
Imagine a retailer testing an AI assistant that helps customers change orders. The team gives it a small set of accurate records, the AI identifies the order, checks whether a replacement is available, and calculates the price. Everything works well during the trial.
But daily operations are more complicated. Customer details can be in one program, orders in another, and inventory in a third. Thus, AI must collect information from all of them, update the order, and send a confirmation. If one program holds an old address or another reports yesterday’s stock level, the AI may quote the wrong amount or promise an unavailable item. It may also stop halfway, leaving an employee to investigate.
Deloitte reported that data problems caused 55% of surveyed organizations to avoid some generative AI uses. In an IBM survey, only 26% of chief data officers felt confident their organizations could extract value from documents, emails, and other unstructured information. For AI to work as intended, the company must know which records are correct, who owns them, and how quickly they change.
AI often needs several programs to exchange information within seconds. Many older applications perform one task inside one department. Some send files at fixed times; others share only limited information through connections built years ago. The scale of this problem can be enormous. In MuleSoft’s 2025 survey of 1,050 IT leaders, organizations used an average of 897 applications. Only 2% had connected more than half of them.
Older systems also carry years of business decisions, some of which exist only in instruction manuals or employees’ memories. For example, a retailer may give certain customers a special discount through an old pricing program that no one has documented. An accountant may know that one report always needs a manual correction, but such rules can remain invisible to AI.
Traditional security gives employees access according to their jobs. A salesperson can view an order but cannot change a customer’s credit limit. An AI assistant may cross several departments to complete one task. Too much access could let it change information it should not touch; too little could stop it halfway.
The order assistant may need to read inventory and change a delivery address, but not approve a large refund. The company must set such limits and send sensitive decisions to an employee.
The company must log each step so staff can see what information the AI used and which records it changed. They must also be able to stop or reverse an action.
Beyond software and computing power, an AI budget must cover data cleanup, application connections, documentation, security, and monitoring. IBM research found that companies that accounted for the cost of addressing old technology problems projected returns up to 29% higher. It also found that 81% of executives believed these problems limited AI success.
Kyndryl’s 2025 survey of 3,700 leaders found that 57% said technology problems delayed innovation. Participants ranked connecting AI with existing IT as the main obstacle to AI expansion. Modernization is therefore part of the AI investment, not a separate exercise.
Simply moving an old application to the cloud will not achieve this. Changing where software runs does not correct inconsistent information, undocumented rules, or weak connections. AI fails on legacy enterprise systems when companies treat it as a new feature that can sit on top of an unchanged business. Success begins with the less glamorous work of fixing the information, connections, and controls that AI needs. The model may power the service, but the surrounding systems determine whether customers and employees can depend on it. Start by assessing the systems your AI use case depends on, before you invest more in