When an AI project falls apart inside a company, the instinct is to blame the AI. Research from two separate corners now says that instinct is usually wrong.
Gartner predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. Gartner's own analyst on the report put it plainly: most of these projects are early experiments driven by hype and often misapplied, which hides the real cost and complexity from the people signing off on them.
Part of the confusion comes from the market itself. Gartner has flagged a pattern it calls agent washing, where vendors rebrand existing products like AI assistants, automation software and chatbots as agentic AI without any real new capability behind the label. A company that buys a rebadged product and expects it to handle judgment calls on its own is set up to fail before the first task even runs.
A separate study from MIT's Media Lab found something that lines up with this exactly. Researchers looked at 300 public AI deployments and interviewed dozens of executives, and found that 95 percent of corporate AI pilots delivered no measurable financial return at all. The report also found that companies buying AI tools from specialized vendors succeeded roughly twice as often as companies that tried to build their own from scratch.
Both studies point to the same root cause described by the specialists in this space. Most business processes were never written down properly. They run on habits, workarounds, and knowledge that lives in one person's head rather than in any document. A human employee can improvise around that mess because they understand context. An AI agent cannot. When it hits a situation nobody planned for, it either guesses or takes an action nobody authorized, and it does this at speed and at scale, which turns a small gap into a large problem fast.
This matters for any company thinking about handing a process to an AI agent, whether that is approving expense reports, routing customer complaints, or checking supplier invoices. The question to ask before buying anything is not "which AI model is best." It is "do we actually know, step by step, what happens in this process when something unusual comes up." If the honest answer is no, an AI agent will not fix that gap. It will expose it, immediately and repeatedly.
The path that seems to work, based on both pieces of research, is narrow and unglamorous. Pick one high volume task with a clear right answer, keep a person reviewing the edge cases, decide in advance what success looks like in numbers, and buy a tested tool rather than building one internally. None of that requires understanding how AI works under the hood. It requires the kind of process discipline that has nothing to do with AI at all, and everything to do with running a tight operation.