Nearly half of enterprise AI efforts are never making it to production, and often, it’s not because of the model. Companies that invest in a faster, smarter enterprise AI platform want it to revolutionise search, support, and decision making almost overnight. A few months in, adoption quietly drops off, leadership asks what went wrong, and the honest answer is almost never the model itself. It’s because the AI wasn’t provided with clean, fully connected, governed data to start with. This piece examines how this continues to occur, what the most recent research reveals, and what CIOs and IT leaders can do about it.
Imagine a mid-sized insurance company that deploys an AI assistant to enable claims adjusters to quickly locate policy information. On paper, it’s a simple win. The model is functioning correctly on its own — it can summarise documents, answer questions, and carry on a conversation. The issue lies in the things surrounding it: policy wording is stored in PDFs that no one has indexed; adjuster notes sit in a ticketing system that no one ever searches; and underwriting history is in a system that only a few senior staffers can access.
The assistant can’t see any of that reliably, so it fills the gaps with guesses — confidently and sometimes wrong. Within a few months, adjusters stop trusting it and quietly go back to searching manually. Nobody formally cancels the project. It just fades into the background: technically launched, never delivered.
This isn’t an isolated case. Three independent 2025-2026 studies, using three different methodologies, all point to the same underlying problem.
In July 2025, the MIT Media Lab’s Project NANDA found that 95% of organisations that have adopted generative AI have achieved no measurable return on their investment. Gartner reached a similar conclusion from a different angle, predicting in June 2025 that more than 40% of agentic AI projects will be cancelled by the end of 2027 — citing escalating costs, unclear business value, and weak AI risk management controls, not model performance.
For IT leaders, the latest data point is the easiest one to relate to. On June 8, 2026, the Institute for Business Value at IBM released its 2026 Tech Leader Study, which interviewed 2,000 CIOs and CTOs from 33 countries. It found that 77% of organisations say AI adoption is already outpacing their governance capabilities, and 70% say business teams are deploying AI faster than IT can even track it. Only 11% of technology leaders feel fully prepared for the scale of AI agent deployment expected over the next year — the same gap IBM’s own real-time data business, Confluent has been pointing to: enterprise AI agents stall because the data layer beneath them was never built for real-time, cross-system retrieval.
Unstructured and fragmented knowledge
Most institutional knowledge, from contracts to policy manuals to case notes to engineering decisions, exists in structured and unstructured formats and systems, none of which were built for machine retrieval. When an AI model is fed raw and unindexed documents, it produces exactly the confident-wrong-answer problem that IT professionals dread.
Governance that can’t keep pace
The IBM 2026 study has revealed that for organisations relying on manual governance, incident probability rises with AI adoption; those using AI governance tools instead see a 25% reduction in incidents. The average number of incidents involving AI agents that required human intervention for correction was 54 during 2025, with 37% of those incidents classed as severe and related to data exposure or security breaches.
Data access built for the wrong era
The security models used for decades divide enterprise data into separate swimlanes. That was a good thing before AI agents had to operate across departments in real time. Now it forces a “bad trade-off”: over-provision access and introduce risk or lock the data down and stall the project — just what happened to the claims adjusters above.
In most cases, a few simple indicators can be enough to determine whether a stalled project is a data problem:
If two or more of these sound familiar, upgrading to a bigger or newer model is the wrong move. What you need instead is a better-connected, better-governed data layer underneath it.
The ones that are ahead of the game aren’t the ones spending the most on their models — they’re the ones that are treating data readiness as the real AI programme. IBM’s research bears this out: companies that embed control and structure into their AI systems from the ground up have 16 times more AI agents than those relying on manual governance, and 18% higher operating margins. Organisations with robust data and financial discipline deploy 2.4 times more agents, with no higher AI budget, and are three times as likely to report that they’re fully equipped for AI use at scale.
In reality, this begins with a straightforward, honest AI readiness assessment, before the next pilot or the next agent:
Where is the data that’s already feeding your AI systems actually coming from? Is it accurate? Is it up to date? And who is responsible for managing access to it?
It sounds basic, and it is — which is exactly why it’s consistently the step enterprises skip on the way to shipping something faster.
With Vaultiscan, businesses can transform the data sitting underneath a stalled AI initiative into a resource their AI assistant can actually use, linking SharePoint drives, CRM and ERP systems, and case files into a unified, managed, retrievable layer.
It also enables IT and AI teams to make a larger leap: from purchasing newer models and hoping adoption sticks, to making data readiness the actual AI programme. That’s what makes the difference between an assistant employees quietly abandon within a few months and one that infrastructure teams come to rely on.
Most importantly, Vaultiscan enables enterprises to achieve three things at once: link disconnected knowledge, keep it governed, and make it trustworthy for AI. As a private, enterprise-ready knowledge layer, Vaultiscan transforms structured and unstructured data into a real-time, queryable knowledge base called Vaulti Lake; adds a secure assistant, Vaulti GPT, on top of that knowledge base; and enables engineering teams to build on that foundation without the data ever leaving their control via Vaulti SDK.
Rarely the model. MIT, Gartner, and IBM’s independent research all point to the same root cause: the data layer beneath the AI. Unstructured knowledge scattered across systems, governance that can’t keep pace with adoption, and access controls built for a pre-AI era combine to stall projects that otherwise launch fine.
A newer or bigger model doesn’t fix broken retrieval. Vaultiscan connects and governs the data an enterprise AI platform already needs — SharePoint, CRMs, ERPs, case files — so the model paired with it can retrieve accurate, current, permissioned information instead of guessing.
It’s a short audit of which data sources actually feed your AI systems today, how current and well-structured that data is, and who’s accountable for governing access to it. IBM’s research suggests this is the step most enterprises skip — and the one that predicts whether the next rollout sticks or stalls.
It depends on how fragmented the data is, but most organisations see a measurable improvement within weeks of connecting and structuring their core systems, not months of retraining or replacing the AI model itself.
The three independent studies from the past two years (2025-2026) all point to the same conclusion: enterprise AI isn’t failing because the models aren’t good enough. This is because the data it relies on was never prepared for AI use. Before you add another pilot, agent, or vendor, consider a simpler question— can your data really support the AI you’ve already invested? A proper AI knowledge base, plus a bit of AI data analytics, will get you further than another new-model demo.
The first step is tackling the data layer — structuring it, managing it, and integrating it with the systems already in place — so the AI you’ve already invested in works the way you want it to from day one. That’s the problem RSK Business Solutions started Vaultiscan to solve.
Want to see how it works with your data — not a hypothetical insurer’s? Schedule a demo with Vaultiscan.
Fill in your details and our team will contact you to discuss your use case and next steps.
Just a few details to begin.
By submitting, you agree to our Terms & Privacy Policy.
