On stage in Amsterdam this month, I presented four years of research with one finding at its center that nobody in the room argued with: where the data was not ready, the AI project was not going to succeed. It was the clearest leading indicator we found.
The gap everyone feels and nobody names
The numbers around AI adoption are strange when you put them side by side. The technology has never been more capable. And yet the IBM Institute for Business Value reports that only 16 percent of AI initiatives successfully scale across the enterprise, while MIT's NANDA research found 95 percent of generative AI pilots never progress beyond experimentation.
That is not a technology gap. Models did not get worse. It is an adoption gap, and my paper for IntelliSys 2026, Bridging the AI Adoption Gap, spent 18 pages locating it. The shortest version: the bottleneck is organizational, not algorithmic. And the single most measurable piece of that bottleneck is the state of your data.
The 68/32 split
Buried in IBM's report, AI Data Quality: The Foundation for Trusted AI, is the statistic I quote to every client. Among organizations that describe themselves as AI-first, 68 percent report well-established data governance frameworks. Among everyone else: 32 percent.
Read that carefully. The organizations winning with AI are not the ones with the biggest models or budgets. They are the ones that did the unglamorous work of knowing what data they have, where it lives, who owns it, and whether it can be trusted. Data governance is not the compliance chore that comes after the exciting AI project. It is the difference between the 16 percent and everyone else.
What 214 implementations taught me
Between 2021 and 2025 I documented 214 AI implementation engagements with business leaders, nonprofit executives, and other non-technical decision makers. When we coded why projects failed, six failure modes kept recurring, and data-readiness gaps sat near the top of the list, alongside technology-first framing, its usual partner in crime. Organizations pick the exciting tool first and discover afterward that the data it needs is scattered, inconsistent, or simply absent.
The pattern was strong enough that when we built the CREATE Framework, we made one design decision that inverts how most vendors sell AI: an explicit data-readiness gate that sits before technology selection. In CREATE, you do not get to evaluate tools until your data has been assessed, and a failed assessment blocks the project from advancing. Not slows. Blocks.
One case study from the paper shows why. A workforce development nonprofit wanted AI-enhanced program matching. The data-readiness work revealed that their intake assessments captured too little about participant goals and barriers for any model to match well. Fixing the intake process came first; the AI came second. The result was a 23 percent reduction in early dropout and a 40 percent improvement in participant satisfaction. If they had bought the tool first, the tool would have failed, and AI would have taken the blame that belonged to the intake form.
Why data quality dominates
The research is blunt about the mechanics. When data meets quality criteria, accuracy, consistency, completeness, and timeliness, models learn efficiently, generalize, and hold up in production. When it does not, you get slow learning, unreliable predictions, amplified bias, and failures at exactly the edge cases that matter. Helpful-looking output on top of bad data is not a win. It is a liability with good bedside manner.
The uncomfortable math for leaders
Here is the conclusion from my paper that ruffles vendors: for organizations early in AI maturity, investment in data governance and quality infrastructure may yield greater returns than investment in more sophisticated AI algorithms. The upgrade your organization probably needs is not a better model. It is a mature data process: named ownership, documented sources, a provenance statement for anything that feeds a model, and an honest readiness assessment before any vendor demo.
That is the boring answer. It is also the one the evidence supports, from my 214 engagements and from IBM's research alike. Get the data right and the AI works. Skip it and join the 84 percent.
If you want the readiness assessment done properly, that is the first thing we do inside every engagement. We build, deploy, and run AI for organizations, and it always starts with the data. Book a free discovery call.
Sources
Gregory Richardson, Bridging the AI Adoption Gap: A Practitioner Framework for Non-Technical Decision Makers, in Intelligent Systems and Applications, Proceedings of the 2026 Intelligent Systems Conference (IntelliSys), Volume 1, Springer. IBM Institute for Business Value, AI Data Quality: The Foundation for Trusted AI (2025). MIT NANDA study as cited therein.