The personal financial advisor costs $200 a month

On Thursday OpenAI launched ChatGPT Finance. US subscribers on the $200-a-month Pro tier can now connect their bank and brokerage accounts to ChatGPT through Plaid, and ask the chatbot questions about their spending, their portfolio, and what to do next. An OpenAI spokesperson framed it cleanly: "We're bringing the experience of having a personal financial advisor to ChatGPT."¹

I want to take that framing seriously, because it deserves to be taken seriously, and then I want to ask the question the framing is designed to skip: which people are we talking about, and what is actually being handed over in exchange.

The product is real, and the demand is real. OpenAI says 200 million people a month already come to ChatGPT with finance questions.² That's not a fabricated number to justify a launch; that's a description of what people are already doing with a tool they have. Personalised financial guidance has, for most of its history, been a service available to people wealthy enough to pay a human adviser by the hour or by assets-under-management percentage. The bottom three quintiles of US earners have never had a financial advisor and were never going to. If a chatbot can give a single mother in Akron a coherent answer about whether to pay down a 24% APR credit card or contribute to her employer's 401(k) match, that is, on its own terms, a useful thing.

I want to hold that thought, because the rest of this piece is going to push hard in the other direction, and I don't want the pushback to flatten what's genuinely on offer.

Who can actually use this. ChatGPT Pro costs $200 a month. That's $2,400 a year for a software subscription, on top of whatever the user is already paying for the financial accounts being connected. In the US, the median household income is roughly $80,000; a $2,400 annual subscription is three percent of pre-tax income for the median household and a rounding error for the top decile. Pew's consumer surveys consistently find that paid software subscriptions over $20 a month are concentrated in the top two income quintiles.

So the launch population is, by construction, people who already have investment accounts to summarise, who already have enough discretionary spending that a dashboard categorising it is interesting rather than painful, and who already have $2,400 a year to spend on software. The person staring down a benefits cliff because a raise will cost them their childcare subsidy, the person deciding whether to take the 401(k) loan to cover the boiler, the person trying to understand what an IDR plan does to their student loans — these are the people for whom personalised financial guidance would be most valuable, and they are not the launch cohort. They are, at best, the eventual mass-market cohort, after the product has been shaped around the preferences and complaints of people who pay $200 a month.

This is not a small detail of rollout sequencing. Products inherit the priorities of their early users. A financial guidance tool built to satisfy people optimising their backdoor Roth conversions and their tax-loss harvesting will, when it eventually trickles down, be a tool with those affordances and those instincts. The questions a Pro subscriber asks ChatGPT about their finances are not the questions a Walmart cashier needs answered, and the shape of the product follows the shape of the early feedback loop.

What users are handing over. Plaid is the pipe. Users authenticate to their bank through Plaid's interface, Plaid retrieves transaction history and account balances, and that data flows to OpenAI. Plaid's existing infrastructure already connects to apps like Venmo, Robinhood, and Coinbase, so the integration itself is unremarkable. What's worth attending to is who Plaid is, and what they have agreed to in the past.

In December 2022, the FTC settled with Plaid for $58 million over allegations that Plaid had collected more financial data than consumers were told about, and used it for purposes consumers had not consented to.³ Plaid's post-settlement consent architecture is better than its pre-settlement architecture, and the company will tell you, accurately, that it operates a "permissioned data" model where users authorise each connection. But the user experience of granting Plaid access, click through a few screens, type your bank password, click confirm, is not meaningfully different from the experience that produced the FTC complaint in the first place. Anyone who has used Plaid knows that the disclosure language sits in a place where almost no one reads it, and the granular controls over what data is shared and for how long are, in practice, things almost no one configures.

Now add OpenAI as the consumer of that data. OpenAI is not a Registered Investment Adviser. It is not regulated by the SEC for the guidance it offers. The disclosures that would be required if a human advisor sat across a desk and said the things ChatGPT is going to say are not required here, because the guidance is positioned as informational rather than advisory. This is the same legal positioning every personal finance app uses, which is a fair point and one I want to name: ChatGPT Finance is not categorically more lawless than Mint or YNAB. It is, however, categorically more capable of sounding authoritative, more capable of producing personalised recommendations on demand, and operating at a scale none of those apps have approached. The regulatory frame was designed for tools that show you a pie chart of your spending. It was not designed for a tool that will, on request, tell you whether to sell your Tesla stock.

Consent that was confusing when it covered a spending dashboard becomes something else when it covers a system that will give you advice you will act on.

The data pipeline is built before the business model is announced. This is the part I find hardest to look away from. OpenAI is currently not running ads, not taking referral fees from financial product providers, not monetising the Finance feature in any way other than through the Pro subscription itself. Sam Altman has said publicly, on multiple occasions, that he doesn't want ChatGPT to be an ad product. I have no reason to doubt he means it today.

But the consent users are granting today is not bounded by today's business model. They are connecting bank accounts, granting read access to transaction histories, and authorising a multi-year relationship with a company whose business model has changed substantially every eighteen months since it was founded. The pipeline is being built before the commercial terms of its eventual use are disclosed, because they cannot be disclosed, because they have not been decided. A user who consents in May 2026 to OpenAI seeing their transaction data for the purpose of spending analysis is, functionally, also consenting to whatever OpenAI's 2028 business model decides to do with that data, subject only to whatever opt-out mechanism exists at the time.

This is the enshittification trajectory in slow motion. Step one: build the product as pure utility, no ads, no commercial muddiness, no conflicts. Step two: get users to grant the access required for the product to work. Step three: notice that the data is enormously valuable and that competitive pressure is rising and that the public markets are asking when, exactly, this company plans to be profitable. Step four: introduce a "personalised offers" feature, or a "we found a better savings account" referral, or a tier where ads are how you avoid paying. None of this is hypothetical: it is what happened to Mint, to Credit Karma, to essentially every consumer finance app that started as utility and ended as a lead-generation business for credit card and loan providers.

I am not predicting that OpenAI will follow this trajectory. I am pointing out that the architecture of the launch makes it available to them, that the consent users grant today does not protect them against it, and that the people best positioned to demand stronger commitments, well-resourced early adopters who could organise, complain, and switch, are also the people whose data is least vulnerable to the worst forms of monetisation. The harm of being targeted by predatory financial product ads based on your transaction history is concentrated, like most harms in this story, on the people who will arrive at the product last.

The pattern, not just the feature. ChatGPT Health launched in January. ChatGPT Finance in May. OpenAI acquired the Hiro finance team in April. Intuit integration is on the roadmap. The pattern is legible: OpenAI is moving into the life-critical verticals first — the domains where the data is most intimate, the dependence is most sticky, and the commercial value of the relationship is highest over time. Both verticals launched Pro-only.

I don't think this pattern is a conspiracy. I think it is what any commercial entity does when it has identified high-value verticals and a sequencing strategy for entering them. But "what any commercial entity does" is exactly the frame that the original promise of ChatGPT as a general-purpose assistant was designed to soften. The chatbot you trust because it helped you draft an email is the chatbot you are now being asked to trust with your retirement accounts, and the trust is being asked for at the moment when commercial pressure on OpenAI to monetise more aggressively is rising, not falling.

What's actually being offered. A genuinely useful product, to a slice of the population that least needs it, in exchange for a long-tailed data relationship with a company whose commercial future is unsettled, mediated by an infrastructure layer with a documented regulatory history, outside the disclosure framework that would govern an equivalent human service. All of these things are true at once.

I am not telling you not to use it. I am telling you that "we're bringing the experience of having a personal financial advisor to ChatGPT" is one sentence describing one part of what's happening, and the parts it leaves out are the ones that matter most to the people who, three years from now, will be the ones using it.

Footnotes

¶