Financial services has stopped debating whether to use AI. According to Finastra’s 2026 State of the Nation research, adoption is now close to universal, with barely one institution in fifty reporting none at all, and the sector has moved from experimenting with the technology to running on it. The interesting question is no longer whether a bank uses AI, but what it has done with it.

At a panel on the future of banking at London Tech Week this month, a senior figure from one of the UK’s largest banks painted a picture of whether a fully autonomous bank exists yet and where AI in financial services is heading. The answer was no, and the more you sit with that answer, the more interesting it becomes.

Autonomy in banking is designed, not switched on

Autonomy in banking is not a single thing you switch on. It is a spectrum of decisions: where you trust a machine to act alone, where you require a human to confirm, and where the answer shifts entirely depending on the customer, the product, and the regulator. Getting that calibration right is genuinely hard, and the firms treating it as a configuration exercise rather than a design philosophy are going to find out the difficult way that they are not the same thing.

Layered on, or built in

There is a difference between a bank that has layered AI tools onto its existing processes and a bank that has embedded intelligence into the structure of how it operates. The first kind is doing digital transformation with a new vocabulary whilst the second kind is actually building something different. What makes the second kind possible, particularly in a regulated environment where every autonomous decision is a potential liability, is an AI architecture that can show its working. Not in the sense of producing a confidence score, but in the genuine sense of being able to say: here is the rule that applied, here is how it was applied and here is why the outcome was this one and not another.

Embedding requires AI that can show its working

Generative AI can sit on top to interpret what a client wants from an ordinary conversation, but that interpretation carries a randomness the execution never can. A bank that has merely layered AI on treats those as one system and hopes for the best. A bank that has built AI in knows precisely which layer is doing what, and which one it can be held to. Embedding, not layering, is what gets you to autonomy.