Enterprises are facing persistent data management hurdles as they attempt to transition artificial intelligence projects into production environments, according to a joint study from Everpure and analyst firm Omdia. The findings show that 97% of organisations struggle to advance beyond pilot stages, with 62% encountering moderate to severe obstacles.
This implementation bottleneck is largely driven by storage and infrastructure complexities, as 68% of IT leaders rank data management as their principal challenge and 63% cite fragmented storage across disparate systems as a major barrier. Much of this friction stems directly from dark data, which consists of information that enterprises collect and retain but rarely reuse. Such repositories often include Redundant, Obsolete and Trivial material, known as ROT data, alongside commercially valuable records that remain difficult to identify or access. Indeed, 99% of organisations admit to harbouring dark data, with 37% reporting that these unmanaged assets constitute between 51% and 75% of their entire corporate information base.
data was
dark
25%
50%
75%
99%
Limited visibility complicates data use
Before organisations can successfully power AI systems, they must understand precisely what information they possess, where it resides, how it is utilised and whether it remains sufficiently current. The research shows that 58% of enterprises still lack basic visibility across their data estates, leaving useful business records intermingled with duplicate, outdated and low-value material. Everpure links this widespread blind spot to inflated storage expenses, increased exposure to security and compliance hazards, and persistent difficulties in identifying data capable of supporting analytics or modern AI workloads.
These operational liabilities explain why 76% of IT leaders now view dark data as a notable business hazard. Furthermore, 75% of respondents consider the ability to extract useful insight from these dormant assets essential to achieving AI success.
“Enterprises cannot build trustworthy AI on untrustworthy data. Redundant, Obsolete, and Trivial (ROT) data creates noise at the very layer that should provide context, increasing the risk of inaccurate AI inferences. The path to better AI isn’t simply adding more data, but ensuring the relevant, context-rich data is used which requires a comprehensive understanding of the data landscape to unlock business value,” said Ashish Gupta, General Manager, Data Management, Everpure.
Everpure recommends classifying data by value and risk
To address these operational bottlenecks, Everpure recommends assessing all enterprise data against its business value and associated risk profile before deciding how it should be governed. Under this framework, high-value information carrying low risk can be prioritised for immediate deployment in AI projects, whereas high-value data carrying greater exposure must remain subject to stricter governance, tighter access controls and dedicated security measures. Conversely, low-value information presenting elevated risk can be systematically secured, reduced or permanently deleted to limit regulatory and compliance vulnerabilities. In contrast, low-value and low-risk records should simply be archived or removed, helping organisations curb unnecessary storage costs while halting further data sprawl.
Alongside this tiering, the report advises businesses to identify and eliminate ROT data before feeding any inputs into AI workloads, whilst classifying remaining dark data to decide whether it should be utilised, governed, retained or permanently removed. Navigating this process requires confronting modern architectural complexity, as enterprise records are frequently scattered across software-as-a-service applications, hybrid cloud platforms, on-premises infrastructure, and traditional computing systems. Organisations ultimately need a clearer view of these multi-environment ecosystems before they can determine which information is genuinely suitable for AI deployment and implement the appropriate controls around it.
