30% trust the bank chatbot: RFI Global finds customers hold banks to a stricter AI bar

Dan Latimore, who runs North America research at RFI Global and banks with Bank of America, asked Erica whether he should buy a CD. Erica sent him to the CD page of the mobile app ^[15]. He put the same situation to Anthropic's Claude, framed as $15,000 sitting in checking, and got back federal deposit insurance, indicative rates, the possibility of Fed cuts and the liquidity trade-off ^[16]. The institution holding the transaction history produced a link. The system holding one sentence of context produced the reasoning. That is a deposit-pricing conversation, and it happened somewhere the bank cannot see.

The survey numbers around that anecdote are more interesting than the headline. High trust runs 33% for product questions, 32% for fraud alerts, 30% for learning about money, 28% for budget insights and 27% for reviewing finances ^{[2][3][4][5][6]}. Five tasks of wildly different difficulty, six points of spread, a 30% average ^[20][21]. Answering a product FAQ and reviewing someone's finances are not the same engineering problem, but respondents grade them as though they were. They are not scoring capability. They are scoring the counterparty, which is roughly what Latimore says when he calls it tough to disentangle trusting the bank's AI from trusting the bank ^[12].

His prescription follows from that: stronger fraud mitigation and useful round-the-clock support, then climb the trust ladder ^[11]. The awkward part is that neither of those improves an answer, and the answer is where Claude won. Most bank assistants still serve pre-written libraries, some with natural language understanding mapping the question to a canned reply ^[13]. Latimore's own verdict is that they beat scrolling through menus and not much else ^[14]. A scripted assistant is cheap to keep running and quietly gets worse every time the free comparison improves.

Rebuilding is not a front-end project. Wenni Wu, chief growth officer at the seven-year-old Piermont Bank, says the hard part is holding data security and risk management steady while the technology goes in ^[17], and that AI is only as good as the data it can reach, with bank data spread across various systems that first have to be cleaned up ^[18]. On the customer side she expects small-business owners, who lack the device and access discipline of large firms and run their finances across a pile of separate software, to force the fragmentation question faster than banks would choose ^[19].

Then there is the arithmetic on the remedy. 78% say they hold banks to a higher AI standard than tech firms ^[8]; 63% say they would trust their bank's AI more if a human were clearly responsible for the outcome ^[9]. Different questions, so treat the 15-point gap as an outer bound rather than a measured deficit ^[23]. Even read generously, a named human owner is endorsed by fewer people than impose the stricter bar, so accountability language does not by itself buy parity. And on the easiest task in the battery, product questions, at least 67% still withheld high trust ^[22]. RFI frames the stakes as competitive, with consumers already taking financial questions to Claude and ChatGPT ^[24]. The competitor in that framing has no deposits to defend and no compliance review on its answers.