Tim Brophy reported on May 21, 2026 that a surge of AI pilots in financial services is failing to reach production because the underlying data infrastructure is fragmented and slow. Brophy, citing a conversation with Dr. Efi Pylarinou and Mike Sisk, says the bottleneck is not model quality but the inability of current data architectures to deliver the right inputs at machine speed — an issue that prevents organizations from converting model investment into business outcomes.
Operationally, firms run critical information across siloed systems, legacy architectures and manual spreadsheets that cannot satisfy AI requirements for latency or contextual relevance. Brophy and his interviewees list concrete platform requirements for production AI: millisecond access to signals, contextual retrieval that surfaces relevant background for each query, cross‑silo joins across legacy schemas, and built‑in governance and audit trails.
Market context underscores the scale of the problem. Dr. Pylarinou is quoted saying, "More than 40% of financial services are still managing their data in spreadsheets" and that "More than 50% have data that are locked in systems that generate that data." According to Brophy, organizations that rearchitected earlier are advancing with agentic and generative AI use cases, while firms relying on batch processing, traditional data lakes or fragmented stores are falling behind.
The business consequences are tangible: restricted or delayed data access undermines real‑time fraud detection, slows customer service responses and increases compliance risk. Mike Sisk puts the operational stakes plainly: "Data is the backbone of any AI success" and "Without a solid infrastructure, even the best models can’t deliver results," meaning that model development alone will not produce expected outcomes without infrastructure changes.
Security risks add urgency to the infrastructure problem. Brophy and Pylarinou point to a stress test in which an autonomous agent queried thousands of confidential files in about two hours at a major consulting firm, demonstrating how quickly agents can traverse systems when controls are absent. That episode highlights the need for per‑request authorization, strict access controls and forensic logging at machine speed to prevent unintended exposures.
For builders and architects the practical takeaway is straightforward: unify onboarding records, transactional systems and behavioral signals into a single platform that supports millisecond retrieval, relevancy scoring and policy‑based access. Doing so shifts firms from reactive reporting to proactive, machine‑speed decision‑making, enabling scale for agentic and generative AI use cases while reducing exposure to operational and security failures. The full discussion with Dr. Pylarinou and Mike Sisk is available as a recorded webinar referenced in the original report.
Sources
Replies (0)
No replies in this topic yet.
