The financial services industry is pouring money into GenAI, from $35 billion in 2023 to a projected $97 billion by 2027, with annual infrastructure spend on track to hit $400 billion by 2030. Every major bank, insurer, and asset manager has an AI strategy, and most have dozens of pilots running.
And yet most of those pilots are stuck. Not because the models are bad, but because the data infrastructure underneath cannot keep up.
The Real Bottleneck
The World Economic Forum clearly identified it in its 2026 AI Playbook: the number-two investment priority for financial institutions is "foundational data and technical capabilities." They note that firms are "prioritizing enterprise-grade data platforms, investing in modern data architectures that support real-time ingestion, large-scale unstructured data and governed access for AI models and agents."
That is a polite way of saying AI cannot deliver business value if it cannot reach your data.
Financial institutions hold some of the richest data estates on the planet, including decades of regulatory filings, transaction records, research archives, customer interaction histories, call recordings, compliance documents, and market data going back years. It is all there, but the problem is that it sits fragmented across dozens of storage silos, locked by data gravity, and inaccessible to AI workloads without weeks of staging and replication.
75% of IT projects fail to meet expectations, and AI project failure rates run even higher. The pattern is the same every time: the model works in the lab, then hits production and cannot access the data it needs, where it needs it, when it needs it.
Two Sides of the Same Coin
Every GenAI use case in financial services boils down to two things: making the data corpus work for the institution and making AI work for the customer.
These are not separate problems but two sides of the same coin. Get the data layer right, and the AI outcomes follow. Get it wrong, and the billions spent deliver only fragmented pilots rather than enterprise transformation.
The Data Side: Enterprise Intelligence
Think about what financial institutions actually need AI to do on the back end. Compliance teams spend over 60% of their time just locating and summarising information. Deutsche Bank built "Aggie" with Kodex AI to handle regulatory text summarisation and cut the process time from hours to minutes, while other institutions have significantly reduced the time to generate suspicious activity reports using cloud-based GenAI. These are not experiments but production systems delivering measurable results.
Risk modeling spans fraud detection, credit scoring, climate and ESG analysis, and pricing, and all of these disciplines require continuous retraining on massive unstructured datasets. Capital One and JPMorgan have reported significant reductions in false positives and improved detection rates with GenAI-powered fraud systems (IMF), while leading European insurers have scaled to dozens of AI applications that process millions of API calls every month.
Research tells the same story. Morgan Stanley uses OpenAI to give financial advisors instant access to insights across companies, sectors, and asset classes, drawn from its proprietary data. Other major banks have reported tenfold increases in prospecting efficiency, and Robinhood built Cortex to deliver professional-grade market intelligence on Amazon Bedrock.
All different use cases, all with the same dependency: they need the right unstructured data delivered to the right compute at the right time.
Here is the uncomfortable statistic: average enterprise GPU utilization is just 5%. Before AI jobs can begin, organizations often have to find scarce GPU capacity, reserve it, and migrate their data to wherever those GPUs are available. The challenge extends beyond GPU availability. Data gravity keeps data anchored to specific locations, while fragmented storage forces teams to replicate and stage datasets before AI workloads can begin. The result is greater operational complexity, slower AI initiatives, and higher infrastructure costs. It is no surprise, then, that 43% of organizations report rising AI infrastructure expenses, with agentic AI infrastructure leading the increase at 56%.
The fix is a unified, high-performance data layer that serves every AI workload from a single namespace, bringing together compliance archives, KYC (Know Your Customer) records, research corpora, market data, transcripts, and proprietary analytics into a single accessible location, without fragmentation, replication across teams, or staging. Sustained throughput keeps GPUs fed continuously so retraining happens on the team's schedule, not the infrastructure's.
When the data layer works, compliance analysts shift from retrieval to investigation, data scientists shift from waiting to iterating, research teams shift from data wrangling to market insight, and GPU procurement becomes a scheduling decision rather than a data migration project.
The Experience Side: Customer Engagement
Customer expectations have permanently shifted. 71% of banking consumers globally want an AI assistant in their bank app, with 82% wanting to approve agent actions before execution. The gap between what customers expect and what institutions deliver is widening every quarter.
Some banks are already there. ABN AMRO handles over 2 million text conversations and 1.5 million voice conversations annually through AI agents, with over 50% of them fully automated. Other leading banks have reported significant uplifts in customer engagement and savings applications through AI-driven personalisation, while call centre data collection time has dropped dramatically across the sector.
But the challenge is not building the AI agent. It provides the agent with access to everything they need to resolve a query in real time, including call recordings, conversation transcripts, account documents, emails, transaction histories, previous interactions, product holdings, and complaint records.
Fragmented storage means agents get partial context; partial context means escalation to humans, and escalation means that the 50% automation rate drops to single digits.
Multinational banks face an additional layer of complexity because customer data must remain sovereign in each jurisdiction, while AI services require unified access across regions, spanning GDPR in Europe, DORA for operational resilience, PCI-DSS for payments, and local data residency laws. Traditional storage architectures cannot satisfy these requirements without replicating data into separate silos per region, and that replication destroys the unified view the AI agent needs to actually help the customer.
The answer is the same as for the corpus use case: one data layer, sovereign per jurisdiction, accessible in real time from any AI service in any geography, with complete customer histories available to agents without pre-staging, without replication between channels, and without compromise on compliance.
When it works, agents resolve queries instead of humans triaging queues, and institutions deliver the personalised, instant experience customers demand while maintaining full regulatory compliance across every jurisdiction they operate in.
The Common Thread
Every GenAI use case in financial services depends on getting unstructured data to compute at the right time, in the right place, under the right governance. The model is never the bottleneck. The data infrastructure is.
That is where Qumulo changes the equation.
Qumulo delivers a unified data platform that spans on-premises, edge, and cloud environments, giving AI applications governed access to unstructured data through a single global namespace. Cloud Data Fabric eliminates fragmented storage silos, while Cloud Native Qumulo extends enterprise file services into the cloud with the scale and performance required for AI.
On the data side of the coin, Qumulo CloudBridging and GPU Liquidity enable institutions to rapidly expose enterprise data to available cloud GPU capacity without creating additional copies or undertaking large-scale migration projects. Cloud Data Fabric maintains a single view of distributed data, while NeuralCache keeps active datasets close to compute, accelerating model training and inference and helping organizations make better use of their GPU resources.
On the customer engagement side, the same platform gives AI agents secure, real-time access to complete customer context across channels and jurisdictions, without replicating data into regional silos. Financial institutions can deliver personalized experiences while maintaining data sovereignty, regulatory compliance, and a single operational view of enterprise data.
Underpinning it all is security. NeuralProtect helps safeguard data at the point of write, strengthening cyber resilience and ensuring the integrity of the data AI depends on. The result is a single platform that enables financial institutions to confidently scale AI across internal intelligence and customer-facing experiences.
The institutions that get this right will not just deploy AI but deploy it at scale, across every line of business, in every geography, on any cloud. The ones that do not will keep running pilots that never quite make it to production.
$97 billion will be flowing into financial AI by 2027. The question is not whether your institution will spend it, but whether you will have the data infrastructure to make it count.
Tom Tasker is a Principal Cloud Solutions Architect at Qumulo, focused on cloud data protection and ransomware resilience for financial services and regulated industries across EMEA and APJ. He is the co-author of the AWS Storage Blog post Resilience by Design: Building an Effective Ransomware Recovery Strategy and lectures at Loughborough University on Modern Data Architecture.