Storage: the blind spot in AI projects
Artificial intelligence has become the top investment priority for businesses. They expect an improved customer experience, more efficient teams, and faster processes. Yet, according to the IDC study conducted in partnership with Everpure*, 54% of AI projects in the pilot phase never move into production. In other words, more than half of AI budgets fail to generate any return on investment.
The cause is not a lack of ambition or computing power; it lies with the data itself: its availability, its freshness, and its governance. Without these three pillars, no AI project can succeed. Traditional storage infrastructures, designed for other uses, are hitting their limits in the face of the latency and throughput demands of AI workloads.
Too much data, not enough quality
Also according to the study, 94% of organizations consider data quality important, or very important, for the success of an AI project. A nearly universal figure that clashes with the reality on the ground.
The three main culprits of poor data quality are well known: redundancy, the multiplication of siloed storage systems, and data obsolescence.
This diagnosis explains why 60% of organizations deem their storage infrastructure insufficient, whether for improvements or a complete overhaul. The finding holds true across all types of workloads, including generative, predictive, interpretive, and agentic.
A GPU that waits is costly
Investments in GPUs are among the heaviest in a data center dedicated to AI, yet these resources are often underutilized. Nearly half of organizations point to a need for better workload scheduling, and a similar proportion mention slow data delivery over the network. The bottleneck, therefore, is not the compute but upstream, in storage’s ability to feed processors at the right pace.
This reality applies to all categories of workloads, whether generative, predictive, interpretive, or agentic; each imposes its own constraints in terms of volume, latency, or governance. Agentic AI, in particular, crystallizes tensions: security and real-time data access are among the top two barriers cited to its deployment.
Data scientists waiting more than creating
The human cost of this inertia is equally tangible. In more than a third of organizations, data scientists spend more than half of their time waiting for data sets, GPU time, or IT requests. For profiles as scarce and costly to recruit as these, waiting represents a waste that is hard to justify.
Data readiness, even before skills or budget, emerges as the leading bottleneck for AI initiatives. A clear signal: improving data preparation directly reduces time to production, and thus the time to generate value.
Consolidate without chasing uniformity
Facing this reality, companies are turning to unified data platforms. Nearly half already run one or are seriously considering it. The goal isn’t to consolidate everything onto a single system, which remains technically out of reach for most organizations, but to reduce silos through a better-governed data circulation.
IDC speaks of data logistics: ensuring data arrives at the right place, at the right time, with the required quality and traceability. This logistics becomes the backbone on which any AI-ready infrastructure rests, referred to as AI-RDSI. The institute identifies four levels of maturity, from artisanal experimentation to fully automated mastery, passing through organizations that are just beginning to structure their practices.
What remains is a question of pace more than of resources. Organizations investing in a flexible and governed infrastructure do not merely avoid pilot failures; they build the advantage that will separate, tomorrow, those who truly leverage AI from those who keep dreaming.
Learn more about L’expertise Everpure.
*Global IDC study in partnership with Everpure, conducted with more than 1,300 organizations representative of different countries, sectors, sizes of companies and respondent roles, ensuring results reflecting a wide and balanced range of organizational experiences.
