• Banks’ fear of falling behind competitors is helping drive billions of dollars of AI investment
• Accenture found just 20% of banking leaders are seeing widespread and sustained value from AI initiatives
• Individual productivity improvements are failing to translate automatically into system-wide gains
• Some banks are optimizing costs by matching different AI models to specific tasks
• Return on investment is expected to become a much greater priority as AI spending faces increased scrutiny
• AI agents and parallel workflows could deliver the larger productivity gains banks are seeking

Banks’ fear of missing out on artificial intelligence is helping fuel a spending boom across the industry, even as many institutions struggle to demonstrate that their investments are producing meaningful returns.

Global banking expenditure on AI exceeded an estimated $40 billion last year, with financial institutions continuing to invest heavily as executives increasingly regard the technology as essential to remaining competitive.

However, new research from Accenture suggests the scale of spending is running ahead of banks’ ability to extract sustained value.

Just 20% of banking leaders surveyed by the consultancy said they were seeing widespread and sustained value from their AI initiatives.

The research covered 212 retail banking and 110 capital markets C-suite executives across 20 countries between April and June.

Mike Abbott, Accenture’s global banking lead, told Banking Dive that confidence in AI remains high despite difficulties measuring its financial impact.

“There’s a high level of confidence they’re going to get something out of it, but it hasn’t exactly been measurable just yet,” Abbott said.

Banks are currently finding some of the clearest benefits by incorporating generative AI into established and repeatable processes, including call center operations, underwriting, marketing and content production, and regulatory reporting automation.

But institutions remain more focused on reducing costs than generating additional revenue from the technology.

According to Abbott, one problem is that many banks have approached AI primarily as a way of making individual employees more productive rather than redesigning entire processes around the technology.

“Task-wise productivity – getting 10%, 15% for each person – does not add up to system-wide productivity,” he said. “They’re giving AI tools to individuals and saying, ‘Hey, do your job a little bit better.’”

Most institutions have yet to fundamentally reorganize workflows around AI, he added.

Competitive pressure is nevertheless encouraging continued investment. Abbott described a definite “FOMO feeling” within banking, driven by concerns that institutions risk falling behind rivals if they fail to adopt AI quickly enough.

In some cases, he acknowledged, that pressure has resulted in wasted spending.

More experienced institutions are now becoming increasingly selective about which models they deploy.

Large language models may be reserved for complex tasks where their capabilities are necessary, while smaller or open-source models can potentially handle document ingestion and simpler functions more efficiently and with fewer hallucinations.

Banks are also developing models using their proprietary transactional data, potentially giving institutions greater ability to tailor AI to specific financial applications.

The next stage is likely to bring much tougher scrutiny of costs. Abbott expects return on investment to “take center stage” next year as banks question whether expensive advanced models are necessary for particular tasks and prioritize applications capable of producing measurable financial benefits.

One potentially significant opportunity involves replacing traditional sequential processes with AI-enabled parallel workflows.

Mortgage processing, for example, typically involves a series of consecutive stages. AI agents could potentially perform multiple elements simultaneously, dramatically reducing processing times.

Similar approaches could allow banks to build a single AI agent for tasks such as handling lost or stolen cards and deploy it across call centers, websites, mobile applications and branches.

Software development could undergo comparable changes. Rather than following the traditional design, build, test and deploy sequence, AI could allow elements of the development process to operate concurrently.

Abbott said some leading banks are already experimenting with such approaches and expects them to become considerably more mainstream.

But realizing AI’s potential will require more than buying increasingly sophisticated models. Banks will also need to invest in existing employees and develop the expertise required to redesign processes around the technology.

The industry may therefore be approaching a decisive transition. The first phase of banking’s generative AI revolution has largely been characterized by experimentation, competitive anxiety and rapidly increasing expenditure. The next is likely to demand evidence.

As Abbott put it, “Right now we’re in euphoria, and that’s reflected in the spending.”