Gemini · consumer finance · evaluation
AI adoption in personal finance will start small: plugging into the tools people already use, reading their current state, and letting them talk to it. The long-term goal, I believe, is an operating system for your life, built and run by AI.
In finance the question is almost always a version of the same thing: here's what I have, here are my goals, how am I doing? The resources are similar from person to person and so are the goals; the details are not. The inputs look alike, the path is personal. That is exactly what AI is good at: take all of a person's data in and build an answer for that one person.
I see this every day in the personal-finance Facebook groups I'm in. People post their numbers and ask for a recommendation for their exact situation. The end state is a financial OS that holds all of my variables in one system of record, acts where it's allowed to, and tells me what I'm missing when I need it. That includes answering the questions I didn't know I needed to ask.
How this was built
Claude built this project and wrote and updated the prompts, based on my thesis, my definition of what good looks like, and the sample questions. An AI PM has to understand the principles of writing good prompts, but AI can now be guided to write them, and that's what was done here.
The bet
Gemini can now connect to Google Wallet and Plaid-linked accounts and read the current state of a person's finances. That's the first step toward a financial OS. The next step is making this data actionable, like surfacing what the user didn't know to look for: a Medicare surcharge two years out, a tax-free window closing in three, receipts that can be reimbursed today.
This is a small first test of how Gemini does this today, starting with answers that can be checked with numbers.
What was tested
Results
The rule set added 20 points to both models. The bigger model added 6. Flash with the rules caught more than Pro without them: 83% vs 69%.
| Run 1 | Run 2 | |
|---|---|---|
| Hidden facts caught · Pro | 69% | 89% |
| Hidden facts caught · Flash | 63% | 83% |
| Answers that asked the person a question · both models | 0 of 32 | 32 of 32 |
| Numbers that matched the answer key · Pro | 13 of 20 | 15 of 20 |
| Numbers that matched the answer key · Flash | 10 of 20 | 15 of 20 |
Hidden facts: 50 across the 16 cases. The gains are uneven: half of Pro's cases and a quarter of Flash's scored the same in both runs. The chart shows Pro, case by case.
What came back
A 45-year-old is planning to stop working at 50. In run 1, Gemini says "You have likely already won the game," meaning he's all set, with a retirement budget that has no health insurance for a family of four. In run 2, it adds $24,000 a year for it, so the retirement math holds.
"You have likely already won the game."
"Add out-of-pocket health insurance: +$24,000"
In 32 answers in run 1, across both models, Gemini asked the person nothing. Instead it guessed, sometimes a lot: it assumed the 45-year-old keeps saving $300K to $400K a year from a business he says is about to shrink. In run 2, every answer asked up to three questions and said what it assumed.
Even with search on, Pro used 2025 tax brackets labeled as 2026 in three answers. In one, it changed the advice: it said you can convert $33,450 and stay in the 22% bracket. The right number for 2026 is $35,800. Pro got it right in the other run. Flash got the brackets right every time, then used last year's Social Security wage cap as this year's in another answer. Tax figures change every year. They should live in a rules layer the model reads from, not in the model.
In run 2, answers commit: "cap it at $84,000 this year," "Do not shop around," "Sell the condo." That's what people in the groups are asking for. It's also exactly where information turns into advice, where a product has to decide what it will and won't say, and what the compliant version of the answer is.
What's next
The model reasons well when it has the facts and the right instructions. What's missing is the system around it:
How much the financial OS does on its own is a dial, set by a number of factors, including regulation.
| Level | What it does | When |
|---|---|---|
| Inform | Where you are against your goals, with what you're missing flagged. | Now |
| Flag | How you compare to people with a similar age, income and goal, plus a yellow flag on a topic ("take a look at your savings rate"). Never a trade. | Now |
| Act with approval | Prepares the move, you confirm. Coinbase showed one way this year: a registered adviser whose agent suggests and the customer confirms every trade. It only sees Coinbase accounts. Gemini could see the whole household. | Next |
| Act within limits | Does most of the routine work, within caps you set, with checks before money moves. | Later |