The opinions expressed here are those of the authors. They do not necessarily reflect the views or positions of UK Finance or its members.

 Somewhere in your organisation, there is almost certainly a model in production. Perhaps several. Trained on clean data, validated in controlled conditions, celebrated in a board update. What may not exist is a credible path from that pilot to enterprise-wide capability. That demands an honest conversation about why.

Across European financial services, AI investment has risen sharply. So has the volume of proofs-of-concept. What has not kept pace is the number of institutions willing to make the structural decision to stop treating AI as a series of experiments and start treating it as infrastructure.  That distinction matters more than it might appear. Experiments are funded differently, governed differently, and resourced differently to infrastructure. Until leadership makes an explicit choice about which category AI belongs in, the default is experimentation, with no defined end point.

Why the pilot mindset is now the problem

The dominant assumption in most institutions is that a successful pilot is evidence of readiness to scale. It is not. Pilots operate in conditions designed to produce success: curated datasets, reduced scope, and governance processes light enough to move quickly. They validate hypotheses. They do not survive production.

When banks attempt to scale, three gaps consistently emerge: fragmented data estates that collapse under shared workloads, governance frameworks designed for periodic review rather than continuous AI, and leadership structures that leave AI in data science teams rather than the operating model. None are technical problems.

Two decisions, one problem

UK retail bank

A UK retail bank secured board approval for a fraud detection proof-of-concept built on curated data. What the pilot had never been required to expose was the state of the underlying data estate: years of inconsistency across business units, independently built pipelines, and a monitoring framework designed for periodic review. Deployment stalled for months while the foundations were rebuilt. The pilot had worked. The organisation had not been ready for what came next.

Major international bank

A large international bank identified that its AI teams were independently rebuilding the same feature engineering pipelines across multiple business lines. Rather than continuing to fund individual model projects, the institution invested in shared feature infrastructure: a unified data layer that any team could draw on, with consistent definitions, governed access, and monitoring built in from the start. Subsequent AI use cases reached production in a fraction of the previous time, because the foundational work had been done once and made reusable. The return on each subsequent use case compounded from the first.

The difference was not technical capability. It was a leadership decision about what to fund.

Banks that have moved beyond pilots share one characteristic: they fund capabilities, not projects. Budget flows to the infrastructure that makes all future deployments faster, not to the individual initiative with the most compelling business case. That is a harder conversation than approving a pilot. It is also the one that separates institutions building durable AI capability from those perpetually relaunching experiments.

The window is narrowing

For UK banks still in the pilot cycle, the risk is not falling behind on model sophistication. The risk is falling behind on the systems, talent, and governance required to deploy that sophistication at scale. Those capabilities take years to build. They cannot be acquired in a vendor contract.  The institutions that made this choice early are now deploying AI at a speed and scale that is structurally difficult to replicate. The advantage is not the technology. It is the accumulated infrastructure, governance maturity, and regulatory confidence that allows them to move faster. Not despite oversight, but because of it.

At a recent Gartner event in London, analysts presented data from mapping over a thousand live agentic deployments globally. Ninety per cent were stuck in what they termed the Amnesia Zone: isolated tasks, no persistent context, no institutional knowledge accumulating, competing only on compute cost. That is not a pilot problem. It is an architecture problem. And it is the direct consequence of funding experiments rather than infrastructure.

The question boards should be asking is not "how many AI pilots do we have running?" It is "have we made the decision to fund the infrastructure that would allow any of them to matter?"

Most have not. The ones that have are pulling ahead.