The right time to start again
Starting again makes sense when a technological change is so significant that rebuilding becomes cheaper and more effective than patching.
If a week is a long time in politics, twelve years is several lifetimes in technology. While Starling, Monzo and Revolut are still referred to as neobanks, in reality, the technological assumptions on which they were conceived are from another era. When I founded Starling in 2014, we pioneered the shift to smartphones becoming the customer's main bank branch. I could see that people no longer wanted branch-based banking. Instead of visiting a building, they wanted to simply open an app, inspect their accounts and click to make payments. That is what we built and the rapid growth of neobanks proved my vision. Since then, the incumbents have caught up with decent mobile apps of their own and what began as a challenge to the industry has become the way it operates.
Today, another transition is on the way and this one is even more significant. Software is no longer only being used by people. AI agents can write and test code, operate applications, and complete tasks. Increasingly, they also act for customers. The question has moved a very long way from how to put a better bank on a phone. We now need banks where some of their builders, operators, and users are machines.
A new generation of neobanks
I can already predict what the naysayers will say. These are the same sort who told me more than a decade ago that banking systems were too complicated, too expensive and too important for a new bank to attempt to build. The new rebuke will almost certainly contain the additional piece of advice that no one should keep starting again and building a new bank every time a new piece of technology appears. There is, of course, some truth to the latter sentiment. Most advances can and should be incorporated into existing systems. Incremental advances in AI, payments or cloud infrastructure do not justify starting again. However, starting again does make sense when a technological change is so significant that rebuilding becomes cheaper and more effective than patching an existing offer.
We need a new generation of neobanks to serve customers who want to take advantage of the freedoms offered by agentic commerce. These agents take away the repetitive, time consuming administration behind tasks such as general shopping, paying bills, cancelling unused subscriptions and finding a bargain holiday. Today's banks require customers to manually verify and authorise every individual payment initiated by an agent, defeating the purpose of autonomous financial assistants.
Existing banks, including the first generation of neobanks, are not prepared for agentic commerce because they were built on the assumption that a human is the principal user. The customer opens their banking app, reads the information presented on the screen, and decides which button to press. This is why we need to completely rethink our relationship with banks and this time it is a much, much bigger task than building a fully functioning app. The good news is, those same agents which are making life better for consumers, also hold the key to this next phase of banking.
When Starling was built, skilled engineers had to translate every requirement into code, examine one another's work, create tests and maintain extensive documentation. Even though Starling was still much smaller and operated faster than a traditional bank, that still required significant teams. Agents can now help with all these activities. They can turn a precise specification into software, generate tests, inspect code for inconsistency, document decisions, and examine systems for vulnerabilities. They can simulate unusual events and investigate why something did not behave as expected.
A smaller team that does more
Building with agents creates a different economic model for a new bank. A bank built in this way will require a smaller, highly experienced team, who oversee the development of its own technology. The process of making the changes needed for agentic commerce to work as it should will move more quickly from each agreed requirement to a tested release.
It is not the case that this is simply a cheaper initial build, either. Day-to-day maintenance accounts for much of the lifetime cost of banking technology. The same tools that built the bank can then help maintain and operate it. Since agents fully understand the specifications, code, tests and operational records, they can help trace failures, keep documentation current and test proposed changes continuously. This is how AI agents should be used.
Why, though, a new bank? Yes, existing neobanks can add agent features to their technology, just as traditional banks eventually produced good mobile apps. Supporting agents properly means more than adding another interface. It requires a bank to be able to identify each agent, understand who it represents, enforce a bounded mandate and retain evidence of every decision. Retrofitting those capabilities across existing customer journeys, while updating existing controls is possible, but it is a substantial undertaking.
A new bank can make these requirements part of its existing architecture from the very beginning. Every service can be designed for agents to interact with directly, where every permission can be limited and monitored with the potential to be revoked independently. It can also be built and operated by a much smaller team using agents to write, test and maintain its technology.
The customers themselves will neither know nor care which parts of their bank were built with agents and nor should they. They should, however, notice the result. They will be able to use whatever trusted agents they choose without surrendering control of their money. Meanwhile, a smaller cost base should create room for better rates and lower charges. A bank that can change its technology more quickly should be able to solve problems and introduce new services faster.
The next-generation agent-native bank does not have to cut a large operation down to size, not least because it can avoid creating unnecessary layers in the first place. An AI-native start-up is not a conventional business that later adds on AI tools. Its technology, workflows and operating model are designed from the outset around people directing agents to perform much of the routine building, testing and admin. This does not mean employing no people; it means concentrating people where their judgment, experience, and accountability matter most. Starting again is not an overreaction. It is the only way to build a bank equal to the agentic age.
