Around 88 percent of leaders report using AI. Roughly 6 percent can demonstrate it is adding value. The distance between those two numbers is not a technology problem.
The gap is evidential, not technical
The figures are widely cited and worth sitting with, because the conclusion they point to is uncomfortable. The constraint on AI value in regulated firms is rarely model performance. It is the inability to demonstrate that performance to anyone who matters: an internal risk committee, an institutional buyer, a regulator.
Put plainly: you can say it happened, but if you cannot prove it to your own organisation, and you certainly cannot prove it to a regulator, it did not happen.
Consumer Duty has sharpened this considerably. The burden of proof has shifted. Firms must now demonstrate good outcomes rather than assert them. The chatbot that cannot handle the word complaint, the model that produces a plausible but incorrect answer, the automated process that cannot be reconstructed or explained: these are not technical failures. They are governance failures, and under the Senior Managers and Certification Regime they sit on identifiable desks.
The literacy gap is uneven, and the risk is asymmetric
Capability is unevenly distributed across most organisations, and the gaps do not carry equal risk. Compliance teams are being asked to make principles-based judgments in a domain they do not yet understand technically. Engineering teams understand the systems but not always the regulatory consequence. Leadership is accountable for AI-driven decisions without necessarily understanding what those decisions involve.
The standard for leadership is not technical mastery. A useful formulation is enough knowledge to be dangerous: understanding what the technology can and cannot do, what the risks are, and how to hold teams accountable for the decisions made with it.
The sharpest version of the problem sits at the customer-facing end, and it compounds. A customer contacts an organisation because an automated system has already failed them. They reach a human agent frustrated, having been given incorrect information or routed through a process that did not understand their need. If that agent then turns to the same or equivalent AI tools to find a resolution, the failure is not corrected. It is extended. Frontline staff are the point at which two flawed AI interactions converge on one already dissatisfied customer. That is an operational and Consumer Duty problem, not a training problem.
Incentives are producing deployments nobody needed
One structural pattern deserves attention on its own. In a significant number of financial institutions, budget is now allocated almost exclusively to work labelled as AI. The consequence is predictable. Teams that need funding for legitimate non-AI work retrofit an AI angle onto projects that do not require it, simply to access the money.
This is not a failure of individual judgment. It is a failure of incentive design, and it produces AI deployments that nobody needed, governed by nobody who understands them, solving problems that did not exist. The direction should run the other way. Leadership defines the strategic use cases, understands enough to ask meaningful questions, and then enables teams with the tools and governance to pursue them.
The question is not whether a firm uses AI. It is whether the firm can prove it deserves to.
What closes the gap
Three things, in order. A defined baseline before deployment, because without it return on investment cannot be demonstrated and drift cannot be detected. Compliance involvement upstream, in the design sprint and the architectural decision rather than at the point of go-live. And evidence generated by the architecture itself, so that the audit trail is a by-product of the system working rather than a reconstruction exercise after the fact.
Pilots do not usually stall because the technology failed. They stall because nobody established what success would look like before it started.
Author: Stephen Walsh, CEO, Keeper Solutions.
READ THE REPORT
This article draws on the AI Capability Building working session, convened by FinTech-Tables with Keeper Solutions and Higgs LLP. The full report sets out the keynotes, the three working group findings, the five key challenges and the ten takeaways in full.
