Banking has always been an information-intensive business. Every credit decision, every compliance check, every customer interaction generates data. For most of the industry's history, the capacity to act on that data was bounded by human bandwidth. AI is changing that constraint by significantly extending what a well-structured institution can do with the data it already holds.

The opportunity is large. McKinsey estimates generative AI alone could add up to $340 billion in annual value to the banking sector — primarily through enhanced risk management, operational automation, and improved customer decision-making. But the evidence also shows that realising that value is harder than most pilots suggest. Deloitte's 2024 Financial AI Adoption Report found that only 38% of AI projects in finance meet or exceed their ROI expectations.

That gap between potential and realisation is the most important thing for any bank's leadership to understand before making significant AI commitments.

Traditional credit scoring draws from a narrow set of variables, including payment history, income and outstanding liabilities.

AI credit models can analyse hundreds of variables simultaneously, including non-traditional signals such as utility payments, rent history, and transaction behaviour patterns, hence expanding credit access while improving risk accuracy. For markets where thin credit files have historically been a barrier, this is a meaningful structural shift.

The caveat worth noting: a McKinsey survey of 44 financial institutions found that more than two in five have slowed AI use-case development because of disappointing outcomes, often linked to rushing deployment without sufficient model validation. The speed gains in credit processing are real, but they require robust testing environments and ongoing model governance to hold.

Regulatory compliance is one of the largest cost centres in banking. It is also one of the clearest productivity opportunities for AI because much of compliance work involves highly repetitive pattern-matching against large volumes of documented transactions and disclosures.

Institutions using AI for compliance and settlement processes report cost reductions of up to 40%, according to BCG research. Separately, over one‑third of organizations now automate a majority of their compliance tasks using AI, with the most mature deployments delivering significant reductions in reporting and audit preparation time, according to recent compliance automation surveys.

Generative AI is already being used to draft regulatory reports, synthesise policy documentation, and flag anomalies across transaction data — work that previously required teams of analysts. The productivity improvement is most pronounced in AML screening, KYC documentation review, and audit trail generation.

The governance consideration here is significant. Regulators across most jurisdictions now expect explainability from AI-driven compliance decisions. That means black-box models are not appropriate for this domain. Institutions deploying AI in compliance workflows need audit-ready decision trails as a design requirement, not an afterthought.

The most visible AI applications in customer-facing banking have been conversational, including chatbots, virtual assistants and personalised alerts. The evidence here is more mixed than in credit or compliance, largely because customer expectations are high and tolerance for AI errors in service interactions is low.

What does work well is using AI to inform and accelerate human customer interactions, rather than to replace them entirely. AI-enhanced onboarding systems have reduced average customer onboarding times by around 50%, compressing processes from 20 to 30 minutes to under 10 mins, while increasing data accuracy through automated validation and reduced manual entry. Banks deploying AI-assisted onboarding also report material improvements in early-stage customer retention, with some seeing up to a 30% uplift within the first six months, reflecting both better customer experience and more reliable onboarding workflows

Bank of America's Erica assistant handles millions of customer interactions monthly, providing personalised financial insights and proactive account notifications — a use case that works because the scope is narrow and well-defined. The model does not speculate; it retrieves and summarises structured account data. That specificity is what makes it reliable.

Wealth and asset management is seeing particularly strong returns from AI adoption. At firms like JPMorgan, AI-enabled advisor tools helped drive a 20% increase in gross sales in asset management between 2023 and 2024, while allowing advisors to expand client coverage. Critically, AI functions as an analytical layer ahead of advisor conversations—synthesising portfolio data, market signals and client history—rather than operating during live client interactions. This model preserves human judgment while materially improving advisor productivity and commercial outcomes.

Physical banking is evolving. And one of the least-discussed but most operationally significant applications of AI in banking sits inside branches:

Computer vision:

Computer vision systems in branches can analyse customer flow in real time, identify congestion points, measure service wait times, and generate staffing recommendations dynamically, converting camera feeds that previously existed only for security into operational intelligence assets.

OCBC Bank has deployed facial recognition across its ATM network, processing two million monthly withdrawals with no PIN required — a frictionless authentication model that simultaneously improves security and customer experience. In parallel, BBVA demonstrated that computer vision-enabled KYC verification, conducted via video call, reduces the account opening process from hours to minutes.

These are examples of production systems generating measurable operational outcomes that point toward a version of branch banking that is more efficient, more personalised, and more secure than what most institutions currently operate.

The institutions achieving consistent, scalable returns from AI share a structural approach that distinguishes them from those running isolated pilots. McKinsey's analysis identifies five markers of high-performing AI banks: a bank-wide AI vision with measured ROI, a full-stack approach combining generative AI with analytical AI, domain-level reimagination of workflows, multi-agent systems for complex processes, and reusable AI components rather than point-in-time builds.

The data infrastructure question comes first. AI models are only as reliable as the data they are trained and operated on. Fragmented data, such as, inconsistent across core banking systems, CRM platforms, and risk engines, produces unreliable model outputs, regardless of model sophistication. This is not a technology problem; it is a data governance problem that requires executive-level commitment to resolve.

Human-in-the-loop design is the second non-negotiable. In credit decisions above certain thresholds, in compliance determinations with regulatory consequence, and in customer-facing communications, AI should function as a decision-support layer, not a final decision-maker. The institutions that pushed past early disappointment and continued iterating on AI in credit are the ones now seeing consistent returns, according to McKinsey's chief risk officer research. Those that prioritised early ROI above model quality were more likely to abandon programmes prematurely.

BCG research shows that organizations building specialist AI teams with deep domain-specific financial expertise—rather than relying on generic machine-learning groups—can unlock up to 60% efficiency gains and 30–40% cost reductions in targeted functions. These gains consistently appear where AI is shaped by people who understand financial products, workflows, and control environments. In this context, domain knowledge is not optional; it is what separates models that perform reliably in production from those that appear effective in pilots but fail under real-world complexity.

Three developments from the past twelve months are materially changing what AI can do inside a bank, and how quickly boards need to form a position on them.

McKinsey's analysis of approximately 4,000 fintechs shows that incumbent banks trail fintechs significantly in deploying AI with measurable business impact — not because of model access (both have it) but because fintechs move faster from experimentation to production deployment and face fewer legacy integration constraints.

92% of global banks report active AI deployment in at least one core banking function. But deployment breadth is not the same as deployment depth. The majority of those deployments remain in pilot or single-function mode. The competitive pressure from fintech and from AI-native neobanks is a present threat, and one that is accelerating.

McKinsey's Global Banking Annual Review 2025 projects that AI-driven cost savings will compress margins industry-wide over time, as competitive pressure passes efficiency gains to customers rather than retaining them within institutions. The strategic implication: the banks that move earlier capture cost advantages that later movers will not be able to replicate at the same margin.

AI in banking is neither a silver bullet nor a speculative bet. It is a set of specific, well-understood tools, each with a defined scope of reliability and a known set of failure conditions.

The institutions that will benefit most are those that approach AI as infrastructure rather than initiative. That means treating data quality as a precondition, not a parallel workstream. It means designing human oversight into production workflows from the start. And it means measuring AI performance with the same rigour applied to any other core banking system.

The current question for any bank's leadership is whether the institution is building the underlying conditions — data, governance, domain expertise, and architectural discipline — that allow AI to perform at the level the business case requires.