Eight ways banks can move AI from pilot to performance

Corporate, commercial and small business banks can turn this around and become first movers. We spoke to banking leaders and EY corporate banking and technology teams in June and July 2025 and found eight ways for banks to seize the AI opportunity.

1. Pivot to external use cases

Today, AI is primarily used for lower-value, internal applications. But every bank executive we spoke to believes the technology is most advantageous when used to benefit the customer. Examples include halving the time it takes to complete a loan application form by pre-populating answers; providing immediate answers to account-related queries through chatbots; or conveying more personalized advice by empowering relationship managers with AI-generated insights.

Despite the potential, banks remain cautious about external use cases due to concerns over hallucinations — AI generating incorrect or noncompliant advice — and unclear ownership, as many AI initiatives are led by teams not directly connected to client-facing functions.

Banks are also overlooking an opportunity that could be even more rewarding than customer experience: a total rethink of their operating models and the services they provide.

“Today, banks are in a race to build their AI capabilities and deploy impactful use cases,” says Matt Cox, EY Global Corporate, Commercial and SME Banking Consulting Leader. “They need to enter the race to use AI to transform their business. That’s the first-mover advantage.”

2. Empower the business

At most banks, technology teams guide AI investment. Paradoxically, most banks acknowledge this isn't ideal. Detached from the client-facing side of the business, technology teams tend to deploy AI in the parts of the business they know well, back-office operations.

Matthew Parker-Jones, Global Head of Product for Global Transaction Banking at Scotiabank, underscores this perspective: “We’ve empowered the business to drive an outcome with AI, whether that’s better client experience or lower costs. We will likely end up going slower, but the impact will be more sustained. You need clarity from the top of the house – the CEO – that this is expected from business leaders. Otherwise, you’ll default to a central team running things.”

Instead, the business itself should lead on AI investment. Business teams are in a position to understand exactly how AI could improve customer experience and create growth and will therefore deploy the technology in areas where it's likely to generate higher returns.

3. Don’t ignore ROI

Assessing the ROI of AI is complex, and in the short term it's often lower than anticipated. Complicating matters further, AI implementation typically involves changes to workflows and processes, making it difficult to isolate its direct impact. This has led some banks to disregard ROI calculations altogether.

History suggests caution in expecting immediate returns. In the early days of internet banking, initial projections of cost savings were premature. Real financial benefits emerged only years later. AI will follow a similar trajectory, with long-term gains far outweighing short-term results.

At the very least, a crude assessment of ROI, perhaps using A/B testing, can help prioritize use cases. It’s also helpful to distinguish between individual use cases and foundational investments in data, technology and human capabilities. These foundational investments must proceed – regardless of ROI justification.

“ROI should not be ignored, even if it’s crudely estimated,” says James Sankey, EY EMEIA Corporate, Commercial and SME Banking Leader. “It helps you to go into this journey with your eyes wide open. Many assume the ROI will be positive, but when you take into account all of the technology costs, it may be negative in the first two years.”

4. Create an AI platform for sustainable success

Many banks struggle to scale AI because they build use cases from scratch. While this approach may be quick, it often leads to long-term challenges — AI use cases become reliant on a patchwork of enabling technologies, tools, and capabilities.

A more sustainable tactic is to build a platform of foundational capabilities that can be reused across any use case. These capabilities include optical character recognition (OCR), machine learning, retrieval augmented generation (RAG) structures, vector databases and prompt libraries. In the long term, this platform approach improves scalability and reduces costs by avoiding duplication of capabilities and the need to maintain multiple underlying systems.

Without a unified platform, banks risk siloed AI agents that limit process integration and client service improvements. “Today, banks rely heavily on AI embedded into platforms they have used for years,” Cox explains. “But these are siloed, which makes it impossible to link together processes in a service like credit. And without this linkage, it’s impossible to change how credit is provided and to improve client service offerings. A platform approach enables this.”

5. Explore modern ways of fixing data issues

Incomplete or poor-quality data is the top barrier to scaling AI. The standard approach to this challenge is to devote human resources to it. But this has its limits and is expensive, so many banks are exploring AI-powered tools that help address data quality. AI needs good data – but it can also help make data better. Banks are already seeing improvements in data validation and compliance using emerging tools.

“We helped a large bank use AI to understand and interpret data used in credit underwriting and validate whether it was correct in the underlying record,” says Adam Smith, EY Americas Corporate, Commercial and SME AI Banking Lead; Managing Director, Financial Services Consulting, Ernst & Young LLP. “It generated a significant uplift to approximately 90%, allowing employees to focus on specific issues that are most likely to be wrong.”

Banks should explore how new AI-powered tools can address data challenges and reduce the need for manual remediation. These tools might still be maturing, but they can help today.

6. Re-evaluate the balance of cloud and on-premises

As banks use AI more widely and technology matures, the required computing power grows drastically. This raises the question of whether to use capacity in the cloud or on-premises. Some prefer the scalability and flexibility offered by cloud. Others have adopted a hybrid approach, blending cloud with the security and control benefits provided by on-premises.

In reality, the significant upfront cost of building large-scale computing infrastructure means it's only viable for larger banks.

But for those that can, the benefits are substantial. “We’ve built our own GPU infrastructure that allows us to build, deploy and maintain AI-powered banking applications,” says Niranjan Vivekanandan, EVP and Chief Operating Officer, RBC Commercial Banking. “This provides us with enhanced security, privacy and sovereignty. That’s vital because trust is central to our relationship with customers.”

Banks also need to reassess assumptions about cloud cost and vendor dependency. As AI workloads grow, cloud pricing models are evolving, and reliance on a few providers raises strategic concerns.

7. Assess future skills requirements

Insufficient technology skills could wreck AI ambitions. When asked about the challenges of creating value from agentic AI, 58% of banks highlighted a lack of technology skills and capabilities. So, banks are aware that AI creates skills challenges.

There are two areas to address. First, the entire workforce needs to be upskilled with the competence and confidence to use AI-powered tools. A combination of targeted training courses and change management programs is essential. Second, banks must add specific capabilities to their technology teams as AI is scaled, including software engineers, user interface and user experience specialists, and business analysts.

“Whether it's AI and data engineers, application developers or those with cybersecurity expertise, banks need anywhere between three and five times the number of people they had five years ago,” says Sameer Gupta, EY Americas Financial Services AI Leader.

Attracting and retaining this talent is just as important. Banks may need to signal their AI ambitions externally to appeal to candidates who might not see banking as a natural fit. Internally, offering varied projects and clear career paths can help retain in-demand talent.

8. Renew attention to risk

Scaling AI, especially in customer-facing applications, comes with risks and rewards. The quality of the output of AI-powered tools is their main concern, especially if they provide advice direct to clients. The EY Responsible AI Pulse survey found that two-thirds of banking leaders consider unreliable AI output a major to moderate concern, and 48% worry about false AI-generated information being taken seriously.

Banks also worry about data governance, especially when proprietary information is used to train large language models (LLMs). As Osamu Abe, Chief of Staff for Asia Pacific, MUFG Bank explains, “Data leakage is one of the greatest risks for us. We might not want our proprietary data to be pooled with that of other companies in LLMs – we don’t know what we’re mixing it with. We’d value being able to curate the set of data specifically for our business and for our needs.”

How can banks mitigate these threats? As a starting point, they'll need a more sophisticated approach to model risk management and model governance for the complex, language-based generative AI (GenAI) models. And that means new skills.

Second, they'll need to address concerns about transparency at the outset of designing an AI application and the surrounding workflows. For example, designers working on knowledge management tools could specify that any advice created by LLMs must reference source material. A thorough risk assessment should also determine when human involvement is needed.

Banks should also work closely with risk teams to define the sequence of checks required before scaling GenAI use cases. ING has taken this approach. “GenAI brings a lot more risk dimensions relating to privacy, intellectual property, vendors, data and decisioning, so you need to find a way of using the technology safely,” says Bahadir Yilmaz, Chief Analytics Officer at ING. “We have 140 checks that need to be assessed for GenAI, though some of these are being automated.”