Debunking GenAI hype in banking

In the expectation of a slew of 2026 predictions making wild speculations about how generative AI (GenAI) is going to change banking beyond recognition, I propose that a more measured approach is needed to uncover the real value that the technology can deliver.  

Last year, I predicted that there would be some consolidation, recalibration and stabilisation in the market and, as a result, we would see a much higher quality of GenAI applications across the banking sector from improving the customer experience to optimising back office processes.  

My projections held true. Throughout 2025, institutions spent time exploring what is possible, relevant and achievable within the banking context, and then drilled down into what was suitable for their specific legacy architectures and technological environments.  

This trend will evolve into more practical actions and initiatives over the next 12 months to provide greater clarity around where GenAI shines versus where it’s not applicable. 

Determinism versus stochastics 

But to attain clarity, it’s important to understand the difference between traditional AI and GenAI. While the former uses deterministic algorithms, the latter is built on stochastic principles, using probability to model systems that appear to vary in a random manner. This means that the same input could generate different outputs.  

However, this isn’t acceptable for fully automated financial operations, which require high reliability, predictability and transparency.  

As such, I believe that GenAI will be most suitable in settings where there’s human intervention. For example, the technology is well-suited for conversational scenarios with tasks that require human oversight but can benefit from GenAI suggestions. Banks can use the technology to launch more interactive user interfaces, where customers can interact with the bank as they would a human, moving beyond simple frequently-asked questions. 

This year will also see a reincarnation of voice assistants in banking, which was subpar and abandoned with early chatbots based on a simple natural language processing, such as Alexa and Google Assistant. Some banks are already looking into using GenAI to recognise voice and generate responses to serve the customer segment who prefer talking to their bank, rather than pressing buttons or touching screens.  

In the back office, banks can leverage GenAI to provide guidance to their employees and accelerate certain tasks. While there has been much concern that staff would be made redundant by GenAI, instead banks should look to use GenAI to improve efficiency and help their staff do more, which will have a positive impact on customer experience as processes will take much less time to complete. 

For example, efficiency can be gained in compliance processes, which are comprised of much manual, redundant technical work, such as analysing documents and summarising text. Instead of a compliance team spending a week analysing hundreds of documents, they could do it in 30 minutes with GenAI.  

The increased efficiency could either mean less people will be needed or more work could be done. I believe the latter is what will happen as there’s much more demand than existing processing power – the bottleneck is because banks can’t process enough applications in a working day. Once they can accelerate the process, the funnel will open up and institutions will see more demand. Instead of reducing the number of employees, banks should look to serve customers faster and better. 

Agentic AI hype 

There is an enormous amount of buzz around agentic AI, or fully autonomous decision-making, but that isn’t going to happen anytime soon because of the difficulty in predicting outcomes with GenAI. It can produce different outputs from the same input due to elements of randomness and probability in its design, which also means it’s not possible to explain why a specific output was generated. Without traceability, regulators won’t be able to ensure that the institution is doing the right thing. 

In addition, agentic AI doesn’t understand the specific context for each individual – the models are generalised, not contextualised. This means an AI agent will determine the statistically most probable scenario in general, not in my particular context. And incorporating individual context before making the decision is not an easy task to do in an automated way.  

Of course, the better the data the higher the probability that the outcome will be good. But there is no visibility into which data was used to train the AI agent, so we can’t determine how much bad data is in the model that will drive decisions about my finance.  

Therefore, I wouldn’t outsource my financial decisions to GenAI because I can’t be sure as to the outcomes. Perhaps they would be good five times out of 10, but the other five outcomes could be suboptimal. 

Many are rightfully questioning whether this is the correct technology to delegate any fully autonomous financial task execution to. Providing advice is one thing, but letting GenAI decide on my behalf? The technology is not built for that.