Why Some Banks Are Blocking AI Access

It’s out of my hands … until it really is out of hand.

Citigroup and Bank of America are the latest major banks to restrict AI agents’ access to client accounts, a recent report in the Financial Times found. The tools in question, however — namely Meta’s Muse, a personal AI assistant that can manage spending and track budgets — are booming, as people increasingly rely on them for financial advice.

“There’s really no undo button on a wire transfer,” said Danny Lohrfink, cofounder of Wealth.com. “AI should inform the financial decisions, but it should not be able to execute them fully unsupervised.”

Rogue One

If an AI agent were given access to a client’s accounts, not only could it “go rogue,” said Mike Wilson, CEO and co-founder of Hamachi.ai — it could potentially also leak client data. “That data could go anywhere,” he added. “That agent could be compromised [and] do things on your behalf that you don’t want to be done, and that’s super scary.” Further, a recent report from LexisNexis found that AI fraud is costing financial services firms $5.47 per dollar lost, with AI involved in about 20% of fraud cases. 

Others in the industry have voiced concerns that if AI is able to shift cash into high-yielding accounts, it could cause banks to lose deposits, the effects of which would trickle down through the whole industry. Part of the reason banks in particular might be hesitant to let AI in is that their business model relies, in part, on making money off cash deposits, Lohrfink said. “If it was as easy as instructing an agent to always be looking for the best place to hold your $1,000 to maximize your yield … most people would do that,” he added. “But that’s not good business for banks because that friction keeps money sitting in cash that can be then loaned out at higher rates.”

Meeting My Maker (and Checker). If an AI model makes financial decisions in someone’s best interest, the output is only ever going to be as good as the data it’s trained on, he said. “If you let an LLM do its thinking in its frame, it’s trained on data that can be outdated,” he said. “It might be referencing old tax laws or old legislation when running its calculations.” There should always be a human in the mix when it comes to executing on trades or wire transfers, he said. 

“I would ask every AI vendor the question: What can your agent do without a human being in the loop to say ‘yes?’” he said. “If the answer includes being able to move money or execute trades, I’d be really cautious about it.”