For a long time, my job was making personal finance easy for regular investors to understand through human-centered design. With the rise of AI, and an increasing number of people hoping to use it to make more informed financial decisions, that work is no longer hemmed in by a single dashboard or experience.

So when Perplexity launched their Finance product, I was genuinely excited. I connected my Fidelity account first. It linked, but the performance number it returned was off. Not close enough to be a rounding error. When pressed, it advised me to consult my Fidelity dashboard for the real number. Then I tried Robinhood; it started to connect, got most of the way through, and then flashed a message I didn't expect: "too many holdings." And canceled out entirely.

Perplexity is one of the most thoughtfully built products I've seen, and I think what they're building with Finance is genuinely exciting. This isn't a failure of their AI. It's a data problem that their AI had no way to detect or correct.

That's the thing I want to talk about. Because most people using these tools have no idea it's happening, and I think they should.

The problem isn't AI.

When you connect a financial account to any AI platform, the data that arrives isn't really your data. It's a translation, passed through an aggregator, recoded into a schema built for recordkeeping rather than intelligence, then handed to a model that has no way to know what got lost along the way.

The AI does exactly what you asked. It reasons over what it received, sounds confident, and gives you a clean answer. If the underlying data was wrong, everything downstream is wrong too, with no warning.

This isn't an edge case. It's the default state of most financial data pipelines right now.

Here are the 4 most common issues:

1. Security misclassification. Most aggregators handle U.S. equities, mutual funds, and ETFs well enough. Collective investment trusts, private securities, municipal bonds, options, futures — these get misclassified or dropped. If you hold anything outside the standard universe, there's a real chance the AI doesn't know what it is.

2. Transaction semantics. A dividend reinvestment arrives coded as a purchase. A cash sweep reads as a withdrawal. A return of capital shows up as income. These aren't rare — they happen in nearly every portfolio, every month. Every calculation that touches those transactions quietly inherits the mistake.

3. History gaps. Most platforms start measuring from the day you sign up. When a brokerage connection drops and reconnects, the gap disappears rather than getting filled. Ask an AI about your 10-year track record and it may be reasoning over 8 months of data.

4. Stale snapshots. Some aggregators push yesterday's data. Some push last week's. When an AI tells you how your portfolio is doing today, it may be describing how things looked a while ago.

None of these are problems that a smarter model solves. They're data infrastructure problems. Better reasoning on top of wrong inputs still produces wrong answers.

Here's why it's harder to fix than it looks.

The obvious question is: doesn't Plaid handle this? Plaid connects to 12,000+ institutions, and does that job extremely well. But connecting to data and understanding what the data means are two different problems.

The interpretation work—knowing that a specific transaction code means dividend reinvestment rather than a purchase, that this security identifier is a CIT and not an ETF, that a gap in history needs reconstruction rather than acceptance—that's domain knowledge. It's built up over years of working inside financial institutions, understanding not just the formats but why those formats exist and where they break down.

It doesn't come from an API. It can't be shortcut. And it only becomes visible when an AI gives you an answer that's confidently, cleanly wrong.

The good news is this is a solvable problem. The industry is moving fast, and I expect the best AI finance products to close this gap over time (and it's something my company has been working on for a while). But right now the gap is real, and it matters.

What to do with this

If you're using AI for your finances, a few things worth trying: ask it a question you already know the answer to. Find a transaction you remember and see how it's categorized. Ask about an account you've held for years and see how far back its memory goes.

You'll get a quick read on whether the data layer is telling your AI the truth.

These tools are worth using. The vision behind products like Perplexity Finance is the right one: your financial picture should be queryable, understandable, and in one place. Getting the data layer right is what makes that vision actually work.

Kate is Head of Design & Marketing at Truthifi. Truthifi Connect is a secure MCP server that normalizes and repairs financial data before your AI ever touches it. truthifi-connect.ai