“Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied,” Anushree Verma, a Senior Director Analyst at Gartner, said in June 2025. She was announcing a prediction: more than 40% of agentic AI projects would be cancelled before the end of 2027, undone by escalating costs, unclear value, or inadequate risk controls.
The numbers that came with it made the point sharper. Gartner found 72% of supply chain organisations already deploying generative AI, and most reported middling results. A separate survey of 120 supply chain executives found only 23% have a formal AI strategy. The rest are running projects. There is a difference. One has a defined success condition; the other has a budget line.
The problem is not the model
RAND Corporation researchers James Ryseff and Anu Narayanan studied AI project failure and found the rate exceeds 80% — more than twice that of non-AI technology projects. Their diagnosis held across cases. The models were rarely at fault. The data underneath them was.
Gartner analyst Roxane Edjlali named the mechanism in February 2025: organisations will abandon 60% of AI projects unsupported by AI-ready data by 2026. A Gartner survey of 248 data-management leaders that quarter found 63% either lacked AI-ready data practices or weren’t sure they had them.
MIT’s Project NANDA studied the question from the other end. Its 2025 report “The GenAI Divide” drew on 52 executive interviews, 153 leadership surveys, and 300 public deployments. The finding: despite $30–40 billion in enterprise GenAI spending, 95% of organisations had seen zero return on their pilots. One Singapore retail bank ran 11 pilots over two years, spent $3.2 million, and shipped nothing. Its CFO killed the programme in the first quarter of 2025.
None of that is a story about bad models. It is a story about pointing sophisticated tools at data nobody cleaned.
What clean data actually looks like
GEP and the University of Virginia’s Darden School surveyed around 180 senior supply chain executives, in work led by Professor Tim Laseter and senior fellow Michael DuVall. Fewer than one in ten had scaled AI pilots across the enterprise. The ones that had shared three traits: automated data cleansing, real-time dashboards fed by trusted inputs, and digital audit trails running across supplier transactions.
Consider one automotive supplier documented in master-data case studies. Its ERP held 2,300 active supplier records. More than 250 were duplicates. Removing them stopped duplicate payments and ended months of manual reconciliation that had quietly poisoned every AI-assisted supplier analysis run against the data. The tools were already live. The records were the problem.
Cisco’s AI Readiness Index surveyed more than 8,000 chief executives. 81% admitted their data remains siloed. Feed a model a siloed dataset and it sees an incomplete picture — then answers with total confidence anyway. A better model does not fix that. It just makes the wrong answer more persuasive.
What separates production from pilot
The projects that ship tend to do three unglamorous things. They fix a CFO-legible baseline before the pilot starts, so “success” means a number rather than a feeling. They spend the majority of the budget on data engineering, not model selection — Gartner recommends at least 60%. And they run on a hard clock: reach production inside a fixed window or get cancelled. Not extended. Cancelled.
ASCM CEO Abe Eshkenazi has pressed the same discipline on the talent side. “Your technology spend has to be matched by your talent development spend, and we’re not there yet,” he said. His rule for the division of labour is blunt: “Allow the technology to do what it does; allow the people to do what they do.”
Verma’s 40% is not a verdict on AI. It is a forecast about what happens when organisations skip the data work and jump straight to agents. The infrastructure problem does not vanish under a smarter tool. It gets more expensive to reach.