GRAPEVINE, TX — Despite all the hype around agentic AI, the near-term payoff in financial services still comes from streamlining unsexy, everyday tasks.
Processing Content
Executives speaking at SAS’s annual conference last week said the most immediate return on investment will come from high-volume back-office work — KYC onboarding, compliance documentation and claims processing — areas where human reviewers can catch mistakes before they become costly.
Agentic AI is in its “terrible twos” stage, Adolfo Lopez, senior vice president of corporate technology at JPMorganChase, said. “Most of the work that is being done with agentic AI is assisted or delegative.”
With the technology still finding its footing, AI agents that support human decision-making rather than make decisions outright comprise most of the use cases, executives said.
Sri Raghavan, principal of data and AI enterprise strategy at AWS, suggested agentic AI will slowly take on more responsibility, with the “human in the lead” orientation defining how tools are being deployed right now.
“I don’t foresee a near or medium-term future where humans are not there as a part of the process,” he said. “It’s not necessarily going to replace them for some of these really important value-added tasks,” including personalized investment recommendations for wealth management clients.
He warned that handing too much autonomy to agents in some high-risk situations can go “spectacularly wrong.”
Automating tedious, menial processes
Raghavan singled out know-your-customer compliance during new customer onboarding as low-hanging fruit for agentic AI efforts.
Meanwhile, insurer Allianz said it deployed agentic AI to process high-volume, low-risk claims.
Andrea Pohlman, head of consulting sales at Allianz Technology of America, said in higher-stakes cases such as catastrophe claims — where a human is the ultimate decision maker — the technology has still been valuable in speeding up processing times. Across the board, she said a bigger goal is cutting commercial underwriting timelines down significantly (from six to eight weeks down to six to eight hours), a need made more pressing by talent shortages.
To decide which use cases make sense, Lopez recommended organizations start small and iterate from there.
“Go back to your organization and find that one process that eats up a significant amount …. of high-impact human hours, and automate that process,” he said.
For Raghavan, getting the data right is a prerequisite. He brought up an example of a client managing 146 data sources across spreadsheets, flat files and data platforms. Rationalizing those assets ensures AI rollouts will be effective, he argued.
“If I had my druthers, that’s one of the first things I’d focus on,” he said.
Build versus buy
Lopez said JPMorganChase uses a mix of buying and developing tools internally; the bank buys the model but builds its control layer in-house.
“The guardrails that actually feed the data are all being built in-house, but the model is coming from outside,” he said. “We build a nervous system and then we leverage it to gain the muscle memory to be able to perform routine tasks on a regular basis.”
Panelists emphasized that the industry is still navigating the tension between agentic AI’s promise, and the practical reality of deploying it in a financial services environment where any shortfall could have major consequences for users.
“We’re in a very highly regulated industry,” said SAS banking advisor Stephen Greer. “We deal with very complex problems that have extremely high consequences for failure and very, very low fault tolerance, and that weighs heavily on the decisions we make when we think about how we might implement this.”
