Banking software provider Oradian says banks need to overhaul their core technology infrastructure to move generative artificial intelligence from experimentation to production, as most enterprise AI pilots fail to generate measurable returns.
Oradian, citing data from MIT’s Project NANDA 2025 report, said 95% of enterprise generative AI pilots deliver no measurable return, with the main obstacles to scaling AI lying in data engineering, regulatory governance, and workflow integration rather than model performance or prompt engineering.
The findings highlight a widening gap between AI experimentation and deployment in financial services.
McKinsey’s 2025 research cited by Oradian showed that 62% of organizations were experimenting with AI agents, but only 23% had scaled them into production.
S&P Global found that the proportion of enterprises abandoning most of their AI initiatives rose to 42% in 2025 from 17% a year earlier, while NVIDIA reported that only 21% of financial institutions evaluating agentic AI had deployed an active agent.
Oradian said traditional core banking systems were built over decades around human operators interacting with computer screens, making them poorly suited to software agents that need continuous access to data and the ability to execute transactions without human intervention.
Adding API gateways or chatbots to legacy systems does not address those constraints, the company said. AI agents require real-time data streams, machine-readable error handling and repeatable workflows that can be executed programmatically.
Oradian’s whitepaper proposes an “AI-native banking” architecture built around three layers: a Read Plane, an Act Plane and a Logic Plane.
The Read Plane would provide AI systems with a continuously synchronized, governed, read-only replica of production data, allowing agents to access current customer information without interfering with core transaction processing.
The Act Plane would combine comprehensive API coverage with real-time event streams, allowing agents to perform actions such as loan origination and credit-limit adjustments as events occur.
The Logic Plane would provide a controlled environment for banks to build, version, and govern proprietary credit models and workflows without waiting for software vendors to release updates.
The infrastructure question is becoming more pressing as regulators in Southeast Asia increase scrutiny of AI used in financial services.
The Bangko Sentral ng Pilipinas issued Memorandum M-2026-031 in June, establishing the STARS framework, covering sustainability, transparency, accountability, responsibility, and security in responsible AI deployment.
Indonesia’s financial regulator, OJK, introduced a full-lifecycle AI governance framework for banks in April 2025.
Oradian cited Philippine consumer fintech Salmon as an example of a financial institution using an AI-enabled core banking architecture.
Salmon deployed on Oradian’s platform in six months, subsequently expanded nationwide, and acquired Rural Bank of Sta. Rosa in 2024.
The company now serves more than 90% of its customers through digital channels and operates proprietary AI underwriting on the platform, according to the whitepaper.
The need to address the underlying technology stack comes as banks prepare for wider adoption of agentic AI. Gartner forecasts that 33% of enterprise software applications will incorporate agentic AI by 2028, up from less than 1% in 2024.
At the same time, Gartner expects more than 40% of agentic AI initiatives to be canceled by the end of 2027 because of rising costs, unclear business value, or inadequate risk controls.
For banks, Oradian said, the challenge is therefore shifting from proving what AI models can do to building systems capable of deploying those models safely, continuously and at scale.