Artificial intelligence is now everywhere in banking. Every strategy presentation mentions it, every institution says it is investing in it, and every board is asking management how quickly it can be deployed. Accenture surveyed 500 American advisors last year – and 96% believe generative AI will transform client service.
In many cases, however, what is being deployed is closer to a demonstration than a transformation with real impact on the business – systems that perform well in controlled environments but would struggle when connected to the complexity of real banking infrastructure.
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
But the biggest risk for banks today is not moving too slowly on AI, it is deploying it in the wrong order.
Why data infrastructure must be the starting point
Across the industry, many banks are prioritising visible AI – chatbots, assistants and customer interfaces – before fixing the foundations that make AI actually work. Data infrastructure, governance and explainability are often treated as secondary problems. In reality, they are the starting point.
This pattern is easy to understand if you look at how AI discussions unfold inside banks. Boards are under pressure to demonstrate progress quickly. The easiest way to show that progress is to launch something visible, and a chatbot is the fastest way to do that.
Technically, deploying a chatbot is relatively simple: feed a large language model internal policies and documentation and it can answer questions from employees or clients almost immediately. But strategically it only scratches the surface of what AI can actually do. The real opportunity lies much deeper, in the way banks organise and use their data.
Financial institutions hold one of the richest datasets in the economy
They know where we spend our money, how our financial behaviour evolves over time and how major life events shape our finances. For many customers, banks hold a record of decades of financial activity. Through KYC, those same institutions also collect a significant amount of information. And it does not stop there: banks sit on multiple layers of structured and unstructured data – much of it highly valuable, if it is properly cleaned and prepared for AI models.
But having the equivalent of several lives of client data doesn’t mean anything if you can’t take advantage of it.
Much of this information sits across legacy systems built over thirty or forty years. Data is stored in different formats across departments and custodians. In some cases, critical client information is still locked inside scanned documents or handwritten forms.
Deploying customer-facing AI on top of fragmented, inconsistent legacy data, effectively automates errors at scale
When banks attempt to deploy AI on top of this environment, the outcome is predictable. Garbage in still means garbage out – even with the most advanced models.
Put simply, the technology is ready, but the plumbing is not.
This is why the correct sequence for AI deployment in banking is simple: data first, models second. Without clean, standardised data, even the most powerful AI engine is useless.
The models themselves are no longer the main challenge. The technology already exists and continues to improve rapidly. What determines whether AI works in banking is the quality and structure of the information it receives.
This issue becomes even more important when regulation is involved.
Modern AI systems rely on neural networks with billions of parameters. In practice they function as black boxes. Even their creators cannot fully explain how every output is generated.
For regulators and compliance teams, that creates understandable hesitation. If an AI system recommends an action, flags a transaction or supports a financial decision, the institution must be able to justify the reasoning behind it.
This is why compliance teams are often seen as blocking AI adoption. Yet compliance may also become one of the biggest beneficiaries of AI.
$60bn: The annual global cost of financial crime compliance
According to LexisNexis, the global cost of financial crime compliance now exceeds $60bn annually. Much of this work involves manual processes such as client onboarding, document verification and regulatory reporting. AI has the potential to transform these processes, but only once the underlying information is structured and accessible. Without that preparation, automating compliance simply adds complexity.
Another common misunderstanding in the industry is the idea that AI in banking essentially means chatbots. Just as important is measuring whether AI systems actually work. Banks need to track metrics such as accuracy, task completion and hallucination rates to understand whether models are improving operations or simply generating convincing answers.
The real transformation will not come from a single model, but from a multi-agentic system with several agents that combine specialised models and are working together as a team. One model can analyse financial data, another can interpret regulatory requirements, another can assess tax implications and another can evaluate market conditions.
This is much closer to how decisions are made inside a bank. Clients are not advised by a single individual, but by specialists working together – across investments, risk, tax and compliance. AI is starting to mirror that structure.
Together they can replicate the collaborative decision-making that currently requires teams of specialists. There is also a clear reason why this transformation matters now.
Baby Boomers and Gen X currently hold around 70% of global wealth
Over the next two decades, an estimated $124 trillion will transfer to their heirs. So, institutions that don’t adapt to digitally-native clients risk losing 70% of the next generation before they’ve even made an introduction.
Those future clients will expect digital services that are fully part of their lives – with faster insights, more proactivity, more empathy, and hyper-personalised advice. At the same time, the financial history that makes personalised advice possible already sits inside the legacy systems of traditional banks.
If institutions can organise and mobilise that information, they will have a powerful advantage. If they cannot, newer digital platforms will quickly catch up.
Banks do not have a technology problem, what they have is a data problem.
Institutions that take the time to organise and structure the information they already hold will unlock enormous value from AI. Those that focus only on surface-level tools will find that the technology delivers far less than expected.
In banking, as in many industries, the difficult work usually matters more than the visible one.
Alvaro Morales, Co-founder & Chairman of Flanks
