As enterprise AI spending accelerates, banks are discovering that reliable data, orchestration and transaction controls, not models alone, will determine whether experiments deliver commercial returns.
Enterprise spending on artificial intelligence is no longer being driven by curiosity alone. Gartner has projected that outlays on AI platforms and models will reach $64bn, while IDC expects global enterprise AI spending to approach $400bn, underscoring a broad shift from experimentation towards systems that can deliver measurable commercial returns. In financial services, that change is especially visible as institutions push beyond customer-facing chatbots and proof-of-concept projects towards more ambitious, operational uses of AI.
That ambition is colliding with a much harder problem: execution. Gartner’s poll found that 55% of organisations were either piloting or had already deployed generative AI, but many banking leaders still face a substantial gap between testing and production. Deloitte has argued that legacy banking architectures were never designed for real-time AI decision-making, particularly where systems must reconcile data across multiple platforms and business lines. Backbase has made a similar point, warning that fragmented frontline infrastructure and inconsistent policy controls can undermine the promise of automation rather than improve it.
The practical barriers are most obvious in data and control. Finextra’s survey with FICO found that 40% of organisations are struggling with a surge in AI-generated synthetic identities and deepfakes, adding to the compliance burden around authentication and fraud prevention. That risk becomes more acute when an AI agent is allowed to trigger payments, liquidity moves or other live transactions using incomplete or poorly matched data. In that environment, banks often fall back on manual review layers and rigid rule-based overrides, which can preserve safety but also dilute the speed and efficiency gains they are seeking.
The emerging view across the sector is that successful adoption will depend less on the model itself than on the plumbing around it. Banks will need cleaner data, stronger entity resolution, and middleware that can validate and de-duplicate information before it reaches an AI system. They will also need tightly controlled transactional APIs, with explicit guardrails that check intent, risk and compliance before any action reaches live rails. The institutions best placed to benefit, analysts say, will be those that treat data architecture and orchestration as strategic assets, not afterthoughts.
Disclaimer: This article is intended to inform and educate, not to recommend or endorse any financial product, investment or strategy. Please consider your own financial circumstances and seek professional advice where appropriate before making financial decisions.
