Every bank I speak to has an AI initiative. Most have several. And yet McKinsey found in late 2024 that most banks have still not delivered revenue growth or efficiency gains at scale from AI. In banking, one of the most data-rich industries on the planet, that should be alarming.

Where is the issue then? The models are ready. The computer is available. But most institutions are treating AI as a technology project when it is fundamentally an operational transformation. Pilots fail to scale not because the technology underperforms, but because they were never connected to a real workflow, a real decision, or a real outcome from the start — and because explainability, governance and audit trails were an afterthought rather than a foundation.

Deloitte's 2026 Banking & Capital Markets Outlook identifies the barriers plainly: fragile data foundations, legacy infrastructure, compliance pressure and internal resistance to change.

That is the divide I see widening every quarter. Not between banks that have AI and banks that do not. Between banks that have embedded AI into how they actually work and those still running pilots.

For a while, AI in banking meant basic chatbots and simple customer service copilots. Useful at the margins, invisible in the core. That is changing rapidly.
In onboarding, lending, underwriting and risk management, AI is now inside the decisions that matter — not summarising them after the fact, but shaping them in real time.

But this shift exposes a critical distinction that too many institutions are missing. Large language models are exceptional at interpreting and summarising vast amounts of text. They are not, on their own, appropriate for regulated credit decisions. Those decisions require predictive models that are explainable, repeatable, auditable and completely consistent — and the quality of those models depends entirely on the data behind them. Models informed by patterns across many lenders, not just a single institution's own book, will make fundamentally better decisions, which is why data depth, not model choice, is becoming the real differentiator, and the gap is widening.

Speed without trust is worthless in banking. And trust, in this context, is not just about customer confidence — it is about regulatory compliance.

If a bank cannot explain how a system made a credit decision, what data it used and why a specific outcome happened, that model will never reach production. And the regulatory bar is about to become explicit. Under the EU AI Act, credit scoring and creditworthiness assessment are classified as high-risk, which means they are subject to strict requirements on transparency, data governance and human oversight. At the time of writing, the high-risk obligations for credit decisioning are expected to move from August 2026 to December 2027, pending final publication in the Official Journal, and many institutions are quietly treating that as breathing room.

That is precisely the wrong response. Explainability cannot be retrofitted into a model already running in production. The institutions that will clear the bar comfortably are the ones designing for it now — treating governance as a foundation, not a feature to bolt on before an audit. They have stopped thinking of compliance as a constraint on AI and started treating it as a design principle. They do not choose between speed and control — they architect for both simultaneously. Responsible governance is not the brake on AI adoption. It is the gating factor that separates pilots from enterprise.

Sergio Segura of nCino discusses responsible AI governance in our Forecasting Fintech webinar with Mambu and Google Cloud. To watch the full webinar, click here.

The industry debate about AI replacing jobs is a distraction. It mistakes the symptom for the problem.

The right question is not whether AI will replace bankers. It is whether your bankers are spending their time on work that actually requires a banker. Right now, in most institutions, the answer is no. Highly trained credit analysts spend the majority of their day gathering data, testing covenants and building reports. Relationship managers spend more time on administration than on relationships. Loan processors spend their hours chasing documents rather than closing deals.

Agentic AI changes this reality fundamentally. These are systems that do not just answer questions but act, orchestrating complex multi-step processes autonomously. When an AI operates continuously, monitoring risk exposure at 3 a.m., testing covenants before the banker arrives, alerting on early warning signals in real time, the banker's role does not disappear. It is elevated. They start the day with analysis done, recommendations ready and their attention freed for judgment, strategy and the client.

The real decisions AI surfaces are not only technical. They are human too: how your people spend their time, what they are trusted to do, and why the best of them will choose to stay.

The institutions getting this right treat AI as a colleague — one that works continuously, embedded within existing workflows, handling the parts of the job that don't require a banker's specific expertise. Their workflows become smarter and more efficient, unconstrained by the hours people work.

AI alone will not replace your bankers. But bankers who can leverage AI will replace bankers who cannot. Institutions that build AI-augmented workflows today are not just improving efficiency. They are building the environment that the next generation of banking talent will choose to work in. In a sector where the war for skilled underwriters, analysts and relationship managers is intensifying, that is a strategic advantage that goes well beyond productivity metrics.

All of this — better data, explainable models, AI woven into the work — delivers far more when the pieces are connected than when they stand alone. Which leads to the decision that matters most over the next 18 months. It is not which AI model to deploy, or which workflow to automate first. It is the orchestration layer: what governs how AI-driven decisions flow across the institution.

Think of it as the difference between a bank that has ten capable AI tools and a bank where intelligence moves between them. In the first, a credit insight sits in the underwriting system; a risk signal sits in monitoring; a cross-sell prompt sits in the CRM — each useful, none aware of the others. In the second, the orchestration layer connects them, so a covenant breach flagged at 3 a.m. reaches the right banker with the right context already assembled. That is what makes intelligence compound rather than accumulate.

It is also why the gap widens every quarter. The banks that get this right will not just deploy AI faster — they will compound its value across every function, while their competitors are still wiring up one use case at a time.

Knowing what to build for is the first step. Knowing how to build it is the second. The execution comes down to six choices — not technology choices, but operational ones. If you are planning your next move, keep the approach simple and focused on measurable outcomes:

Prioritise workflows where AI has an immediate, measurable impact: onboarding, lending, underwriting, portfolio monitoring. Do not start with the most visible use case. Start with the most impactful one.

Ensure your data is clean, connected and accessible to your analytical models. AI amplifies what it operates on. If you amplify well-structured, high-quality data embedded in well-designed processes, you get extraordinary results. If you amplify fragmented data and broken workflows, you get chaos at machine speed. Data readiness is not a technical prerequisite. It is a strategic decision.

Explainability, as I argued earlier, is not a feature to add later. Some decisions must be fully explainable and auditable; others benefit from AI's ability to interpret, summarise and synthesise. Knowing which is which — and deploying the right type of model for each — is one of the most consequential design choices you will make. Confusing the two is one of the most common and costly mistakes in banking AI implementations.

The winning bet, then, is not the model. It is the orchestration around it, the data depth beneath it, the workflows that carry it and the governance model that holds it accountable.

The technology is ready. The question is not whether to move. It is how fast, and whether you are building on the foundation that will actually scale.

The time to move is now.