TLDR
- Most lenders now treat AI as a strategic priority, with 84% calling it critical to their business, according to Vijay Mehta, Experian’s executive vice president of global solutions and analytics.
- But many institutions are still struggling to demonstrate clear financial returns from AI investments.
- Banks are deploying AI first in operational tasks such as customer service, fraud detection and workflow automation while keeping humans involved in credit decisions.
Lenders are embracing artificial intelligence as a strategic imperative, but many are still struggling to prove it is paying off.
That is the central finding of a new Experian study of over 200 senior decision-makers in credit, fraud and analytics roles in the U.S. and U.K. The survey shows that 84% of lenders say AI is a “high or critical priority” for their strategy, and 89% say it will play a critical role across the entire lending lifecycle – from origination to collections.
“We’re hearing that across the country,” said Vijay Mehta, Experian’s executive vice president of global solutions and analytics, in an interview with The AI Innovator. “But then, the other big part of this is that they’re struggling to find ROI.”
According to the survey, 38% of respondents reported struggling to see return on investment from current AI implementations.
The study, titled “AI in Lending: What lenders think, and why it matters,” is based on a focused survey of risk professionals – primarily chief risk officers, along with data scientists, analysts and some IT leaders – across the U.S. and U.K. Mehta said that focus on lending and risk differentiates the research from broader AI surveys.
The goal, he said, was to understand how AI is affecting lending workflows and to inform Experian’s own product roadmap.
From experimentation to strategy
According to the report, AI is no longer viewed as a technology experiment but as a business strategy. Nearly nine in 10 lenders believe AI will play a critical role across the lending lifecycle, underscoring a shift from pilot projects to enterprise-wide ambition.
Mehta said financial institutions generally fall into three camps: a top tier of around 10% pushing the envelope and embedding AI deeply into workflows, a middle group of about 70% integrating AI more cautiously, and a smaller cohort of 20% “standing on the sidelines” because of fear or uncertainty .
Unlike many other industries, banking operates under strict regulatory oversight, especially in the U.S. and the U.K., which shapes how AI can be deployed. “The use of AI with regulated data is going to be different than other industries where they don’t have the same constraints,” Mehta said.
That regulatory context helps explain why lenders are prioritizing use cases that are easier to measure and govern.
Efficiency first
When asked what they want from AI, lenders most frequently cite operational efficiency (78%), improved credit decision accuracy (77%), and better risk mitigation (61%) .
Mehta said efficiency remains the “lowest hanging fruit,” particularly in repeatable tasks such as those in customer service and engineering. “Velocity is a great metric,” he said, pointing to throughput, reduced friction and fewer handoffs in workflows as measurable key performance indicators.
In lending, that might mean streamlining onboarding or automating elements of customer acquisition, while keeping humans in the loop for final underwriting decisions. “There is still some reluctance within the regulated space where you are making a decision about whether someone gets a mortgage or a new line of credit,” Mehta said. “You can automate all of the things up to that point, but you’ve got to have human involvement.”
Fraud and identity protection are also prominent AI applications. Mehta said lenders recognize that “the bad guys are going to be using AI to attack us, so we have to use AI to help prevent that.”
The ROI problem
Despite strong interest, proving financial returns remains difficult. The report identifies uncertain ROI as the second-largest barrier to adoption, after integration complexity.
Mehta said some lenders are measuring success by counting pilots rather than by assessing value in production. “They’re saying, ‘let’s measure success with the number of use cases or pilots we have versus saying, ‘let’s measure success with real value metrics.’”
That disconnect may partly explain why 38% of lenders report difficulty demonstrating ROI. Mehta said the industry is still defining what AI success looks like, whether in terms of cost savings, improved loss ratios or shareholder value.
Cost dynamics are evolving as well; high cost was the third factor cited by survey respondents as a barrier to adoption. Mehta said token costs for large language models have declined over the past two years, while cloud, compute and storage costs have also trended downward. However, he cautioned that long-term pricing models for foundation AI providers remain uncertain.
Trust and data foundations
Data quality emerged as the top factor influencing trust in AI vendors and the survey noted regional differences as well: U.K. lenders prioritize ethical AI and regulatory compliance, while U.S. lenders emphasize innovation and transparency.
“AI is only as strong as the data that you have,” Mehta said.
He described AI-ready data as requiring a trusted semantic layer, proper lineage and monitoring, role-based access controls, and clean data hygiene. In large language model deployments, such controls can reduce hallucinations through techniques such as retrieval-augmented generation and vector stores.
While some data preparation is required, Mehta pushed back on the notion that institutions must overhaul their entire data architecture before starting. “I don’t believe you have to go through this massive data transformation in order to use AI,” he said. Instead, he advises starting with a specific use case and identifying the data assets needed.
Integration complexity remains the top barrier to adoption, reflecting the reality of legacy core systems. Mehta said that in many cases AI can be added through APIs and model context protocols without replacing entire systems. For more disruptive workflow changes, however, larger system replacements may be required.
Experian positions its Ascend platform as a way to help lenders move from pilots to production at scale. Mehta said the platform supports more than 2,000 customers and enables a transition from deterministic workflows to agent-assisted ones without tearing out foundational systems.
Agents and the road ahead
Lenders are becoming more comfortable with AI agents, though full autonomy remains distant. Currently, companies are progressing from AI assistants or chatbots to deploying “digital teammates,” but not agents that act completely independently. “We’re still a little ways away from true autonomous agents,” Mehta said.
Looking ahead, Mehta expects deeper reasoning capabilities, multi-agent frameworks and multimodal systems to shape lending workflows. Over time, he predicted, lending will become more conversational, with customers interacting with a primary agent that coordinates with specialized sub-agents behind the scenes.
For now, however, the industry’s challenge is less about technological possibility and more about disciplined execution “AI success is defined by speed, accuracy, and measurable outcomes,” the report concluded. Lenders may see AI as essential, but the next phase will hinge on turning enthusiasm into demonstrable results.
