Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Use Generative AI for Personalized Financial Advice
Transcript
- Lucas: So Wealthfront recently rolled out a feature they're calling 'Ask AI' — it's a chatbot powered by a large language model that sits inside their app and answers client questions about their actual portfolios. Luna: Not just generic financial FAQs, right? It's tied to your account. Lucas: Exactly. It ingests your holdings, your risk profile, your tax situation — and then you can ask something like 'Should I sell my Apple shares to harvest losses this quarter?' and it will generate a response specific to you. Not a canned paragraph. Luna: I've seen demos. The responses include actual ticker symbols and tax implications. It feels like having a junior advisor on call. Lucas: And that's exactly the use case. Traditional robo-advisors like Betterment or Schwab Intelligent Portfolios are basically decision trees — they rebalance, they tax-loss harvest, but they don't converse. This is the first wave of conversational wealth management. Luna: We talk a lot on this show about how technology is changing the nuts and bolts of business. If you find these conversations useful for what you're building or running, listener support is what keeps this show ad-free and independent. Lucas: Yeah, it's a small thing that makes a big difference. You can contribute at buy me a coffee dot com slash fexingo. Luna: Thanks for that. So Lucas, how is Wealthfront actually building this? Are they fine-tuning an open-source model? Lucas: They are. They're using a fine-tuned version of Meta's Llama 3, and they've trained it on a corpus of annotated financial advisor transcripts, regulatory documents, and anonymized client questions. The key is they've also built a retrieval-augmented generation layer — RAG — that pulls real-time portfolio data before the model answers. Luna: So the model isn't just guessing. It's looking at your actual unrealized gains before it suggests a tax move. Lucas: Correct. And that's the difference between a clever chatbot and a tool that could actually get you in trouble if it were wrong. Wealthfront reports that user engagement — measured by logins per week — jumped 40 percent after they launched the feature. Luna: Forty percent is huge. What about the quality of the advice? Is the SEC comfortable with this? Lucas: That's the billion-dollar question. The SEC has said that ai generated advice falls under the Investment Advisers Act if it's personalized — which it is. So Wealthfront had to register the model as part of their advisory process. They also run every response through a compliance layer that checks for prohibited language — things like guaranteed returns or missing risk disclosures. Luna: I've read that some firms are even using a second AI to audit the first one's outputs. Lucas: Exactly. A 'judge model' that evaluates the primary model's reasoning. Wealthfront doesn't disclose exactly how they do it, but the industry standard is to have a separate LLM review every response for regulatory alignment before it reaches the user. Luna: And bias? Financial advice has a long history of steering people in ways that benefit the firm, not the client. Lucas: Right. Wealthfront says they test for 'advisor bias' — meaning the model should not recommend Wealthfront's own products unless they're genuinely optimal. They published a transparency report showing that in 94 percent of test cases, the AI recommended a fund outside Wealthfront's proprietary lineup when appropriate. Luna: That's actually pretty reassuring. So what does this mean for the traditional financial advisor? Lucas: I think it's more of a complement than a replacement. A lot of advisors spend half their time answering basic questions — 'what's my asset allocation?', 'should I refinance?'. If AI handles those, the advisor can focus on complex situations: estate planning, concentrated stock positions, behavioral coaching during market drops. Luna: But smaller neobanks — Chime, Current, the likes — they don't have Wealthfront's resources. Can they do this? Lucas: They can, through API access. Companies like Writer or Jasper now offer fintech-specific LLM endpoints that include pre-trained compliance guardrails. A neobank could plug one in and have a basic financial Q&A bot in a few weeks. The challenge is data: you need enough transaction history to personalize, and most neobanks have less than five years of data per customer. Luna: So the incumbents have an edge in data, but the startups have agility. Lucas: Exactly. And the incumbents are moving. Fidelity is testing a similar tool for their robo-advisor. Vanguard has published research on using LLMs to generate personalized withdrawal strategies for retirees. This is becoming table stakes. Luna: Let's zoom in on that retirement use case. How does the AI handle the uncertainty of future returns? Lucas: It doesn't predict returns — at least, responsible implementations don't. Instead, they use Monte Carlo simulations in the background and the LLM communicates the range of outcomes. So you ask 'How much can I withdraw per month?' and the AI says 'Based on your portfolio, there's an 85 percent chance you can safely withdraw $3,200 per month for 30 years.' That's the output of a simulation, not the model guessing. Luna: That's interesting because the user experience is a conversation, but the math is still traditional stochastic modeling. Lucas: Precisely. The LLM is the interface, not the calculator. That's a crucial distinction for regulators and for user trust. Luna: Are there any cautionary tales so far? Any firm that launched an AI advisor and had to pull it? Lucas: Not a full pull, but there was a case last year where a fintech's AI gave advice that inadvertently violated state usury laws on loan recommendations. The model didn't know that combined APR from multiple products could exceed a state cap. They had to add a geo-aware compliance module. So the industry is learning the hard way that financial regulation is hyper-local. Luna: Right — what's fine in New York might be illegal in Texas. So the AI needs to know where the user lives. Lucas: And their tax bracket, and their marital status, and a dozen other variables. That's why Wealthfront's approach of fine-tuning on advisor transcripts is smart — those transcripts already contain all those contextual nuances. Luna: One thing I haven't heard you mention: cost. Running LLM inference at scale is expensive. How do these fintechs make the unit economics work? Lucas: Great question. Wealthfront has disclosed that each AI conversation costs them roughly 12 to 18 cents in compute. Compare that to a human advisor call which costs maybe $8 to $12 per interaction. So the economics are dramatically better, as long as the AI doesn't hallucinate and cause a compliance issue that leads to fines. Luna: Eighteen cents versus ten dollars — that's a no-brainer on cost. But the risk is reputational. Lucas: Exactly. One bad answer that gets shared on social media could undo years of brand building. So the firms that are doing this well are the ones that invest heavily in guardrails and testing, not just the model. Luna: So where does this go next? Will we see AI advisors that proactively reach out — 'Hey, I noticed your cash position is high, want to invest?' Lucas: That's the holy grail, but it's legally tricky. Proactive advice triggers fiduciary duty in a way that reactive advice does not. Some firms are experimenting with 'nudges' — non-specific prompts like 'Market conditions have changed. Would you like to review your portfolio?' — and then the conversation becomes advice only after the user opts in. Luna: It's like the opt-in is the line between marketing and advice. Lucas: Exactly. And regulators are watching closely. The SEC's 2025 risk alert on AI in wealth management specifically flagged proactive advice as a focus area. So firms are treading carefully. Luna: It feels like we're at the very beginning of this. The Wealthfront tool is version one. I wonder what version three looks like. Lucas: I suspect version three is a multi-agent system where one agent handles tax, another handles estate planning, another handles insurance — and they coordinate. The user asks one question and the system orchestrates a response across domains. That's where the real value creation will be. Luna: And probably where the real regulatory complexity arrives too. Lucas: No doubt. But for now, the fact that 15 percent of Wealthfront users ask at least one AI question per week tells you the demand is real. People want accessible, personalized financial guidance — and they're willing to trust a machine to give it. Luna: As long as the machine gets it right. Lucas: Right. And so far, the industry is learning that getting it right means building the technology, the compliance, and the trust all at the same time. That's the challenge — and the opportunity.