From Hype to ROI: Making AI Pay Off

AI in banking is moving from hype to hard reality, where real value will come not from widespread experimentation but from disciplined use cases, strong data foundations, and careful alignment between AI capabilities and business needs.

The banking industry seems to have reached a tipping point with AI. Everyone is talking about it, but the results remain mixed. What is really going on? What is going on is a classic technology adoption pattern, accelerated to an unusual speed. When ChatGPT launched, it became the fastest-growing consumer product in history. Suddenly, every board wanted an AI strategy. The problem is that the capability curve and the expectation curve have diverged sharply. AI today is exceptional at certain things — summarisation, search and retrieval from large document sets, paraphrasing, translation, and code generation. But it is mediocre to poor at others — sustained multi-step reasoning, working with very large unstructured datasets, and tasks where precision is non-negotiable. Banks that matched the right capabilities to the right problems early are seeing real returns. Those who tried to force AI into every process are disillusioned.

So how does a bank figure out where AI actually fits?

I use a simple framework built around four conditions. The first is an information gap — is the right person lacking the right information at the right time? Think of a relationship manager who cannot answer a client’s question because the product knowledge lives in a document they have never read. The second is repetition — is someone spending hours on a task that is largely mechanical? Regulatory clause searching is a prime example. The third is viability — is the task simply impossible without AI? Analysing sentiment across every customer support interaction, for example. And the fourth, which is the one most banks overlook, is error tolerance — what happens when AI gets it wrong? If the cost of an error is manageable and detectable, you have a strong candidate. If getting it wrong means a mispriced loan or a missed fraud alert, you need a very different deployment model.

Let us talk about the data challenge. Banks are data-rich but often insight-poor. How does that affect AI deployments?

It is the elephant in the room. Every bank I engage with has vast amounts of data — core banking records, CRM data, product documentation, emails, regulatory filings, and market data. The issue is that this data exists in silos with inconsistent quality and fragmented pipelines. Most banks are somewhere in the early stages of what I call the data value chain — they capture data reasonably well, do some harmonisation for specific purposes, but have not reached the level of organisational knowledge where AI can truly shine. No amount of model sophistication compensates for poor data. Before asking, “What AI model should we use?” every institution should be asking, “Is our data ready for any AI model?”

"The difference between AI hype and AI value is simple: context, clean data, and disciplined deployment.”

There is growing discussion about contextengineering as the real differentiator in AI. Can you explain what that means in practical terms?

This is perhaps the most important concept for banking technologists to understand. The large language models themselves — GPT, Claude, Gemini — are largely commoditised. Every institution has access to the same models. What differentiates a mediocre AI application from a genuinely useful one is the context you provide. Context engineering is the art of giving the AI everything it needs to solve the problem: the right instructions, the right background knowledge, and the right tools. A simple example — if you want AI to extract financial metrics from SEC filings, you need to provide it with the document, the definition of each metric you care about, verification logic, and access to computational tools. The model is the engine, but context is the fuel. Banks that invest in context engineering will extract dramatically more value from the same models everyone else is using.

Measuring AI’s impact has proven surprisingly difficult. What metrics should banks focus on?

This is an honest challenge. The research shows that productivity gains from AI are real but uneven. Studies have found meaningful improvements in customer support efficiency — around 14% more issues resolved per hour — with even larger gains among less experienced staff. But other studies, including one on open-source software development, found that AI actually increased completion times despite developers believing it had helped. The takeaway is that impact is highly dependent on the task, the people, and the systems. One factor I think is enormously important but essentially unmeasurable is fatigue reduction. AI can absorb the most draining, repetitive cognitive work, leaving people to focus on judgment and creativity. That does not show up neatly in a productivity dashboard, but it is transformational.

What are the biggest risks banks face as they scale AI adoption?

Three stand out. The first is bias — AI models trained on historical data will reproduce and sometimes amplify historical biases. In banking, where decisions affect people’s access to credit, housing, and financial services, this is not just a technical problem; it is an ethical and regulatory one. The second is explainability. Most AI models cannot tell you why they made a particular decision. For credit scoring, fraud detection, and loan approvals, regulators will increasingly demand clear explanations. The third, which I believe is underappreciated, is the cost question. Running large language models is expensive — significantly more so than traditional computing. Today those costs are partly subsidised by the model providers, but we do not know how long that will last. Every AI use case needs a clear cost-benefit analysis.

What would you say to a banking CEO who is under pressure to show AI results?

I would say: resist the temptation to do something visible but shallow. Start with operational efficiency use cases where you can measure impact and where the tolerance for AI error is reasonable. Build your data foundations seriously — it is not exciting, but without it nothing works. Invest in people, because the skills gap is real; your workforce needs to understand how to work alongside AI, not just how to use a chatbot. And keep perspective. AI is a genuinely powerful tool, but it is a tool. The institutions that will win are not the ones with the most AI initiatives; they are the ones that deploy AI thoughtfully against problems that matter.