Gartner projects that through the rest of this year, organizations will abandon 60% of AI projects. Not because the models underperformed, but because the data at their foundations wasn’t ready. This isn’t new. In fact, it is decades old.

A decade ago, Gartner put the failure rate for business intelligence projects at somewhere between 70% and 80%. The causes were the same as today and as during the dot-com era: inconsistent definitions, absent governance, and data nobody could trust. Back to the present: S&P Global found that 42% of companies scrapped most of their AI initiatives in 2025, up from 17% the year before.

The number of failures is accelerating. The reasons behind them, experts say, are not.

“If we look at why BI projects fail or why AI projects fail, it’s not the tech,” says Bob Bloem, senior director and head of data and analytics at AWS cloud consultancy and engineering firm Caylent. “They keep showing up because organizations continue to prioritize use case development or shiny new technology over the scaffolding that makes use cases viable,” he adds.

The Foundation AI Keeps Skipping

The three failures Bloem encounters in nearly every engagement recur across industries. They are all data problems, and they all existed long before AI showed up.

First, inconsistent data definitions across business units. That’s what Bloem calls the “customer definition” problem. For instance, marketing, finance, and operations each carry their own version of the same entity, with different downstream transformation logic, producing different answers from the same source data.

And without a defined way to enforce shared standards, no AI system can learn consistently. Second, missing or undocumented lineage. When a model produces a bad output, debugging requires tracing the cause. Without traceable lineage, that process becomes wasted guesswork. Finally, master data management (MDM) and data governance programs that exist on paper but have never been operationalized, allowing duplicate and conflicting records to propagate into feature stores and training pipelines before anyone catches them.

Charles Caldwell, SVP of product management at Redwood Software, an orchestration platform provider, notes that when it comes to data challenges, AI speed, and bad outputs, AI’s autonomy and speed work against AI initiatives. At human speed, when two teams pull from the same source and reach different numbers, the conflict surfaces in a meeting where someone notices, questions are asked, and judgment calls are made. At AI speed, those disagreements don’t surface as arguments. They get synthesized into outputs. “The dashboards never caught it,” Caldwell notes, describing a pattern his teams encounter repeatedly: the data problem was always there, but AI removed the human review that previously surfaced it before it became consequential.

Old Failures, New Casualties

Data governance and clear data ownership remain substantial challenges. For instance, in a scenario where one person manually corrects data discrepancies between systems and no one else realizes that the entire process depends on them, a single absence, resignation, or reorganization can silently break workflows that took years to build. When that hidden dependency surfaces, it does so in a failed billing run, a wrong number in a board presentation, or an AI model that spent three months learning from data nobody knew was wrong.

Andy Boettcher, chief innovation officer at DoubleTrack, a data quality research and Salesforce consulting firm, calls it “invisible tribal knowledge.” “It’s a massive risk,” he says. “Like the meme of the tiny block holding up the entire stack of blocks.” The same pattern appears across practitioner accounts: institutional knowledge required to reconcile data across systems lives in one person’s head, not in any system. When that person is unavailable, projects that looked stable immediately aren’t.

Jeremy Carmona, founder of Clear Concise Consulting, has run AI governance and data quality assessments across nonprofit, healthcare, and government clients. He finds ownership and definition failures occurring together in more than half of his engagements. “The definition problem is the quiet killer,” he says, “because everybody assumes it got settled years ago and nobody checks.”

His practical AI data readiness test? Firms are ready when they have definitions written down for the fields the AI will touch, a known, low duplicate rate on those records, and the required fields are populated. One human owns the data and is allowed to say no [to the AI initiative when the data isn’t ready]. “Almost no organization clears it cold,” Carmona says.

Chris Pardo, head of product for data and analytics at Allvue Systems, a software provider to private equity and private credit managers, explains how the private markets vertical illustrates what happens when the failure is fragmented data.

Pardo notes that, per Allvue’s own research, 56% of firms with significant assets under management still rely heavily on spreadsheets even after investing in purpose-built platforms. Only 8% of private market firms rate their data maturity as high. The failure here is that critical data cannot be uniformly accessed across systems. “You can’t train or run inference on data you can’t access,” Pardo says.

Bloem explains that organizations consistently prioritize building the use case over building the foundation on which the use case will depend. The initiative gets scoped, the vendor gets selected, the board gets a timeline, and the data environment gets audited six weeks in, by which point the model is already producing outputs that can’t be explained or corrected because the lineage isn’t there. Nobody agrees on what the inputs mean.

The chief data officer who wants to break that pattern has a narrow window. Carmona has a prescription: the assessment before scoping. The cheapest versions of the data mistakes are always the ones caught before the AI initiative goes live. The most expensive data issues are discovered after the AI initiative goes live. That version, practitioners report, is more common across verticals and engagements.