RAND spent months interviewing sixty-five data scientists and engineers, each with at least five years building AI systems, to find out why so many of their projects die. The finding wasn't about the technology.
More than eighty percent of AI projects fail. That's double the failure rate of ordinary IT projects that have nothing to do with AI. RAND's own words, from the report itself: twice the already-high rate of failure in corporate technology projects generally.
Here's the part that should stop every executive currently funding a pilot. The leading cause wasn't bad data. It wasn't weak infrastructure. It was leaders and technical teams never actually agreeing on what problem the AI was supposed to solve in the first place.
The Miscommunication Nobody Notices Until The Pilot Ends
Picture how this actually happens, because it rarely looks like an argument. A leader wants "AI for customer service." A technical team hears that and starts building. Nobody in the room defined what winning looks like: faster resolution times, lower cost per ticket, higher satisfaction scores, or something else entirely. Both sides believe they agree. Neither has actually said the specific thing out loud.
Eighty-four percent of the practitioners RAND interviewed named this exact dynamic, leadership-driven miscommunication, as the primary reason their projects failed. Not a data problem. Not a model problem. A conversation that never happened, or happened vaguely enough that two people walked away with two different projects in their heads.
This is why so many pilots technically succeed and still never scale. The model does what it was built to do. It just wasn't built to do the thing leadership actually needed, because nobody forced that translation to happen before the building started. By the time anyone notices, months of engineering time have gone into optimizing a metric nobody upstream actually cares about.
Why This Keeps Happening Even To Smart Teams
The RAND interviewees were not junior. Five years of experience minimum, across industry and academia. This isn't a competence problem. It's a structural one, and it explains why the failure repeats even at companies with genuinely talented people on both sides of the table.
Business leaders speak in outcomes. Faster. Cheaper. Better retention. Technical teams speak in specifications. Precision, recall, latency, throughput. Translating between those two languages is real work, and it's work that easily gets skipped because both sides assume the other one will handle it, or assume the translation is obvious when it isn't.
This asymmetry gets worse the more senior the leader sponsoring the project is, not better. A junior manager asking for something vague gets pushed back on immediately, because a technical lead has the standing to ask clarifying questions without risk. A senior executive asking for the same vague thing gets a nodding room and a team that quietly guesses at what was meant, because pushing back on the person funding the project carries a cost nobody wants to pay. The miscommunication RAND documented is not evenly distributed. It concentrates precisely where the stakes are highest and the willingness to ask a clarifying question is lowest.
MIT's own research on generative AI pilots found something that fits this pattern precisely. The five percent of pilots that actually delivered real returns weren't running better models than everyone else. Their teams picked one specific pain point, executed well against it, and moved. The other ninety-five percent were, in effect, the RAND finding playing out at scale: a team building competently toward a target that was never actually pinned down with enough precision to hit.
What Actually Fixes This, And Why Almost Nobody Does It
The fix costs almost nothing and gets skipped almost every time anyway, for a reason worth naming honestly. It requires a leader and a technical lead to sit in a room before any code gets written and translate a business outcome into a specific, falsifiable technical target, in writing, that both sides sign off on. Not "improve customer service." A number. A metric. A definition of done that a technical team can actually build toward and a business leader can actually recognize as success or failure when it arrives.
That conversation is uncomfortable in a specific way most kickoff meetings are built to avoid. It forces someone to admit they don't yet know exactly what they want, or forces a technical lead to push back on a request that sounds clear but isn't. Skipping that discomfort is what makes the miscommunication so durable. It's not that anyone chooses to leave the target undefined. It's that defining it precisely means someone has to say, out loud, "I'm not sure yet," in a meeting where everyone else seems confident.
Before your next AI initiative gets a green light, ask the question RAND's interviewees wish someone had asked on their projects: if you handed this specification to a technical team with no other context, would two different engineers build the same thing. If the honest answer is no, the miscommunication RAND documented is already baked into the project, and no amount of model quality or data cleaning fixes a target that was never actually agreed on.
There's a version of this test worth running retroactively, too, on whatever AI initiative your organization already has in flight. Pull the original kickoff document, if one exists, and read the success criteria out loud in a room with both the business sponsor and the technical lead present. Ask each of them, separately, what a good outcome actually looks like in specific, measurable terms. If the two answers diverge even slightly, that gap is not a rounding error. It's the exact failure mode eighty-four percent of RAND's interviewees pointed to, already live inside a project everyone still describes as on track.
The uncomfortable truth underneath all of this is that the fix was never technical, which is precisely why it keeps getting treated as one. Every postmortem on a stalled AI pilot reaches for a technology explanation, because that explanation doesn't require anyone in the room to admit they never actually agreed on what they were building. RAND's interviewees, sixty-five of the most experienced people in the field, said otherwise, and their answer costs nothing to act on except the discomfort of asking a clarifying question before the money moves instead of after.
References
Ryseff, James, Brandon F. De Bruhl, and Sydne J. Newberry. "The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed: Avoiding the Anti-Patterns of AI." RAND Corporation, published August 13, 2024.
Challapally, Aditya, et al. "The GenAI Divide: State of AI in Business 2025." MIT Project NANDA, 2025, as reported by Fortune, August 18, 2025.
