There is a reasonable suspicion that AI in personal finance is mostly a new coat of paint on an old business: take a product that was hard to sell, add a chat box, raise the price. Applied to most of the category, the suspicion is correct.
But there is a version of this that survives the scepticism, and the way to find it is a single question. Does the tool calculate, or does it predict?
That one distinction sorts the entire market, and it sorts it against most of the marketing.
Calculation is checkable. Prediction is not.
If a tool tells you that three of your funds share 41 percent of their holdings by weight, you can verify that. Download the holdings files, match by ticker, multiply, add. It will be right or it will be wrong, and you can find out today.
If a tool tells you a stock will outperform next quarter, you cannot verify it at the moment you act on it. You find out later, once, with no way of knowing whether the outcome reflected skill or the coin landing heads.
Both claims are sold in the same interface, with the same confidence, often by the same product. The first is genuinely valuable. The second is a horoscope with a Bloomberg font.
What the calculation layer actually finds
The reason this matters is that the checkable arithmetic keeps finding things people did not know about their own money.
Weight drift. A position bought at 8 percent of a portfolio and never touched can sit at 22 percent after two good years. Nothing was bought. The exposure tripled anyway. Almost nobody notices, because brokerage apps sort holdings alphabetically or by today’s move, never by how much of your outcome each one controls.
Look-through overlap. An investor holds a total-market fund, a large-cap growth fund and a technology fund and reports three positions. Look through to constituents and the same seven or eight mega-cap companies occupy the top of all three, because every market-cap-weighted index puts the largest companies first by construction. True exposure to any one of them is routinely two or three times what the investor assumes.
As those shared holdings outperform, they grow as a share of every fund that holds them and as a share of the portfolio. Concentration increases automatically, with no decision, until something reverses it.
Benchmark-relative return. Your broker shows gain against cost basis, which blends how the asset performed with when you bought it. The useful number is the same holding measured against a benchmark over an identical window, dividends included on both sides. A position up 9 percent in a year the S&P 500 rose 14 percent went up and still cost you.
Drawdown correlation. Correlation computed over five years is dominated by the calm stretches when diversification was never tested. Conditioned on drawdown windows, assets sitting comfortably at 0.3 routinely converge above 0.8. Diversification that disappears exactly when it is needed is not diversification.
Where thematic investing fits, honestly
This site has covered the rise of thematic ETFs, and the look-through problem is particularly acute there.
A thematic fund is a bet expressed as a basket. That is a legitimate and sometimes efficient way to hold a view. The trouble is that a great many thematic funds, once you look through them, are substantially the same handful of large caps you already own through your index fund, wrapped at four to twenty times the fee.
Sometimes the theme is real and the holdings are genuinely distinct. Sometimes it is a label. The only way to know is to compare constituents against what you already hold, which is precisely the calculation almost nobody runs before buying. If you are looking at, say, an AI infrastructure theme, the useful question is not whether the story is good. The story is always good. It is what share of that basket you already own.
So which tools are worth using
The category is uneven and the field narrows fast on four tests.
Does it read your real accounts? Manual-entry tools go stale in a month. A read-only brokerage connection stays current without effort, and that decides whether a tool gets used once or quarterly.
Does it take custody? Robo-advisors like Betterment and Wealthfront solved connectivity by owning the account. That works and it costs an annual percentage of assets, and it means moving your money to use the product. Reading the account you already have is a different trade and, for most people already at Fidelity or Schwab, a better one.
Can you argue with it? Real questions are narrower: does this fund do anything my index fund does not, how did I actually do over three years, what am I doubling up on.
Does it stay in its lane? The tell of a serious product is that it declines to tell you what will go up.
We build Walnut on that last principle specifically, and on the strength of the other three we think it is currently the best AI investing app for someone who already has a broker and wants a straight answer about what is in the account. It connects read-only, answers in plain English, and does not forecast. Our comparison of the best AI investing apps names the rivals and where they win.
The bottom line
The scepticism about AI in personal finance is well aimed at the forecasting half of the category and badly aimed at the measurement half.
Concentration, look-through overlap, benchmark-relative return and drawdown correlation are forty-year-old arithmetic. They are also almost never run by the people whose money is at stake, because running them takes an evening and a spreadsheet. Automating that is not a gimmick. Automating the prediction is.
This article is informational and not investment advice. Walnut is not a registered investment adviser. Past performance does not indicate future results. Investing involves risk, including the possible loss of principal.
