Since May, ChatGPT can link thousands of banks and brokerages through Plaid. In turn, Plaid acts as a bridge between bank accounts and financial apps. Other frontier AI models offer similar financial integrations through alternative tech stacks. For example, Anthropic’s Claude can connect bank data via the official Plaid Dashboard MCP Servers using the Model Context Protocol (MCP), while models like Gemini connect Plaid via Google Cloud and Google Wallet, and open-source tools and third-party providers also offer financial connections.
“ChatGPT, Claude, and other LLMs have made it incredibly easy to share data–whether through uploading static statements or linking live data feeds of accounts,” says Anders Jones, CEO and co-founder of Facet, a flat-fee financial planning and investment management firm.
“The pitch that the model companies give is that that level of information gives them enough context to give financial ‘advice’ (in quotes, because it is not advice by the legal definition),” Jones says.
In September, Vanguard reported that one in three Vanguard investors use AI. Thirty-seven percent of Vanguard investors who’ve used AI for money questions say it gave them wrong or misleading answers, and 44 percent say they are not sure about the quality of the answers they are getting from AI.
This Morning
The daily digest for entrepreneurs and business leaders
Jones explains that there are a few problems with relying on models: first, they aren’t good at math. “They use inference to provide answers, which essentially is a guess at what the right output should be,” says Jones. For certain things, it works well, says Jones. Chatbots can hold their own in fairly advanced conversations. “For others, it doesn’t, and number crunching is one of the key places that it falls down,” Jones adds.
“Second, it doesn’t have nearly enough context to provide a clear financial picture,” Jones explained. When you sit down with a human advisor, there is a level of connection and emotional intelligence that naturally seeps into the conversation, Jones adds. “With an LLM, it is only as good as the data that is in front of it, and there is no way a person can upload all of the nuances and complexities of their life,” says Jones.
AI models have been extensively documented to be prone to hallucinations, while a Stanford study found they also heavily engage in sycophantic behavior (overly agreeing and people-pleasing) to drive user engagement. Jones adds that AI also has a way of coming off as extremely authoritative–right, up until they are proven to be wrong.