Most AI initiatives fail before the model is ever the problem. Gartner, MIT's Project NANDA, and RAND have each found, independently, that the leading cause of AI project failure is the data and reporting foundation underneath it, not the sophistication of the AI itself.
The scale of the problem
In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, rising costs, and unclear business value as the leading causes. Poor data quality was named first.
In 2025, MIT's Project NANDA published research on generative AI adoption inside large organisations and found that around 95% of generative AI pilots failed to deliver a measurable return. The researchers traced the gap not to the quality of the underlying models, most of which performed well in testing, but to how poorly those tools connected to the data, workflows, and decision processes already running inside the business.
RAND Corporation researchers, examining why AI projects fail more often than conventional information-technology projects, estimated that more than 80% of AI projects fail, roughly twice the failure rate RAND attributes to non-AI information-technology projects. Their analysis pointed to problems that had nothing to do with the model itself: unclear business objectives, and data that was not fit for the purpose the project assumed.
Read together, the three figures describe the same pattern from different angles: a large minority of projects are abandoned outright, a large majority of the pilots that do launch fail to produce a return, and total failure runs meaningfully higher than ordinary information-technology delivery. The numbers are not identical, because the three studies measured different populations at different stages. The direction is not in dispute.
The pattern behind three different studies
Gartner, MIT, and RAND surveyed different organisations, used different methods, and arrived at different headline numbers. What they agree on is the location of the failure. None of the three attributes the typical AI project failure primarily to model selection, model performance, or the sophistication of the AI itself. All three point to the data and reporting layer underneath it: whether the organisation's data is complete, consistently defined, and actually connected to the decision the AI initiative was meant to support.
This should not surprise anyone who has worked in management reporting since before AI initiatives existed. A model trained or prompted against fragmented, inconsistently defined data will produce fragmented, inconsistently defined output, just faster and with more apparent confidence than the spreadsheet it replaced. The model was not the bottleneck. It was never going to be.
Why this gets missed
AI initiatives are typically scoped, funded, and evaluated as technology projects: which model, which platform, which use case. Data quality is treated as an assumption rather than a deliverable, presumably fine because the organisation has been running on this data for years. It has been running on it. That is different from the data being complete, consistently defined, and reliable enough to support an automated system that will act on it without a person checking every output.
The organisations in the Gartner, MIT, and RAND findings that avoided this trap generally share one characteristic: they treated the data and reporting foundation as the project, not as a precondition to the project. Cleaning, structuring, and defining the data came first, on its own timeline, before a model was selected or a use case was built.
Fixing the data foundation after an AI system is already in use is more expensive than building it first, and considerably more disruptive, because by then people have already learned whether or not to trust the system's output. Once that trust is lost, rebuilding it takes longer than earning it would have the first time.
What a data foundation actually requires
A data foundation that can support AI initiatives, or any reporting that leadership needs to trust, requires the same things regardless of whether AI is involved. Data has to be extracted reliably from the systems it lives in, rather than assumed to be accurate because it has always been there. Definitions have to be agreed across every department that touches a metric, because a model trained on two departments' contradictory definitions of the same number will not resolve the contradiction. It will average it.
Data quality has to be validated before anything is built on top of it, not discovered afterward when an AI system produces a confidently wrong answer at scale. And the reporting or the AI system built on top of that foundation has to be designed around the decisions it needs to support, not around whatever data happened to be available.
This is not an argument against AI. It is an argument against sequencing it wrong. A system built on a solid data foundation inherits that reliability. One built on a fragmented, undefined foundation inherits that too, just delivered with more confidence and less visibility into where the numbers actually came from.
None of this is a technology decision. It is the same work management reporting has always required: agreeing what a number means, checking that the number is right, and building only after both of those are true. Gartner, MIT's Project NANDA, and RAND arrived at this conclusion from three different directions. Organisations still funding AI initiatives without addressing it are, on the evidence, more likely to be funding the next abandoned pilot than the next working system.