Generative AI did not arrive in banking in 2022. Financial institutions have been applying machine learning, risk modelling, and advanced analytics for decades. What changed with the public release of large language models was not the technology. It was the expectation.Customers began to expect more intelligent, conversational interactions from institutions that were designed, by necessity, for stability and risk control. Boards faced pressure to articulate an AI strategy on a compressed timeline. Regulators accelerated scrutiny around explainability, resilience, and operational control. All of this happened at once.The result was a surge in experimentation. And experimentation, at scale, exposed a structural problem.The pilot is not the productProofs of concept operate in controlled conditions: curated data, limited integration, short-term cost tolerance. Production banking systems operate on an entirely different order of magnitude: tens of millions of customers, high-concurrency transaction volumes, continuous regulatory oversight, multi-year cost accountability. McKinsey's 2025 State of AI survey found that 88% of organisations now use AI in at least one function. Only 7% have achieved full enterprise-wide deployment.That gap is not about effort or intent. It reflects how banks are built.Successful pilots have often been mistaken for enterprise readiness. The hard part of AI in banking is not building models. It is absorbing extreme complexity, predictably, securely, and within defined economic boundaries, at scale.Why the data foundation is decisiveAI does not fail in banking because models are weak. It fails because the data underneath them is fragmented.Most large banks operate on data estates built over decades: core banking systems, risk engines, payments platforms, CRM environments, and regulatory warehouses, each rational when introduced, collectively complex. Traditional analytics can tolerate that fragmentation. AI at scale cannot.When models are deployed into live credit processes, fraud investigations, and compliance workflows, data inconsistencies surface quickly. Definitions that differ across systems begin to matter. Latency that was acceptable in batch reporting becomes a barrier to real-time decisioning. As agentic systems emerge, drawing on shared data across multiple simultaneous workflows, the quality of that foundation determines whether orchestration is manageable or unstable.Governance and economics must align with scaleAs AI expands across institutions, two further constraints come into view. The first is economic: consumption-based infrastructure models create cost volatility at enterprise scale that finance functions cannot easily forecast, and when costs cannot be modelled with confidence, investment committees hesitate. The second is governance: oversight cannot be retrofitted. Bias testing, model validation, audit trails, and human review need to be embedded into deployment processes from the start.European banks have a structural advantage here. The EU AI Act and DORA, often characterised as compliance burdens, are in practice design parameters. Institutions that have built governance infrastructure to meet regulatory requirements have built the same infrastructure that makes confident AI deployment possible. Clear rules reduce ambiguity. Ambiguity is what actually slows programmes down.The institutions that come out ahead over the next decade will not be those that experimented most aggressively. They will be those who build the foundations to operate AI reliably, repeatedly, and at scale.No#GenerativeAI #FinanceInnovationSimon Axon Global Financial Services Industry StrategistTeradata09 Jun, 2026