AI Just Beat My Financial Advisor. Ask It This One Thing First.
In this article
**Marcus Webb** — Infrastructure engineer turned tech writer. Writes about AI, DevOps, and security.
> **Bottom line:** I ran my actual portfolio — $340,000 across a 401(k), a taxable brokerage account, and a Roth IRA — through Claude 4.6 and ChatGPT 5 and compared the output to what my fee-based advisor had recommended for three years.
Both models caught a tax-inefficient asset placement my advisor never flagged, worth roughly $1,900 a year in avoidable drag.
But the AI only became useful after I asked it one specific question: "What are you assuming about me that I haven't told you?" Skip that question and you get generic, dangerously confident advice.
Ask it, and the model starts doing the one thing a good advisor does and a mediocre one skips — interrogating its own assumptions before handing you a number.
I fired my financial advisor on a Tuesday morning in July 2026.
Not because he was incompetent — because I'd just watched an AI model catch a mistake he'd been paid $3,400 a year to not catch, for three straight years.
That sentence should make you uncomfortable, whether you're pro-AI or skeptical of it. It made me uncomfortable too, and I write about this stuff for a living.
The Setup: Three Years of "Trust Me"
Back in 2023, I hired a fee-based advisor through a boutique firm — 1% of assets under management, the standard model.
He was pleasant, credentialed, and did the normal things: rebalanced annually, put bonds in the 401(k), stocks in the taxable account, sent a quarterly PDF with a pie chart on it.
I never questioned the allocation logic because I didn't know enough to question it.
That's the entire business model of human financial advice for people who aren't finance nerds — you're paying for trust, not necessarily for verification.
In June 2026, out of curiosity more than distrust, I dumped my full account statements — balances, holdings, cost basis, my tax bracket, my age, my timeline to retirement — into Claude 4.6 and ChatGPT 5.
Two separate sessions, same data, no cross-contamination.
I wanted to see if a model without a fee incentive would reach the same conclusions as a human being who was, in fact, getting paid a percentage of my balance every single year.
I did not expect what happened next.
The Core Insight: The Question That Changes Everything
My first prompt was lazy, honestly: "Here's my portfolio, is this allocation good?" Both models gave me confident, plausible-sounding answers.
Claude said my allocation was "reasonably diversified for your age." ChatGPT said something almost identical. Neither one asked me a single follow-up question.
That's the failure mode nobody talks about with AI financial advice — it will happily answer with total confidence and zero clarifying questions, exactly like a bad advisor rushing through a review meeting to get to the next client.
**Confidence without inquiry is the tell that you're getting a template, not analysis.**
So I tried something different.
My second prompt was: "Before you answer, tell me what assumptions you're making about my risk tolerance, my timeline, and my tax situation that I haven't actually told you."
What Changed When I Asked
Both models stopped and did something my human advisor had never done in three years of quarterly meetings — they listed their assumptions out loud, in writing, where I could see and correct them.
Claude 4.6 wrote: "I'm assuming you have a 20+ year horizon based on typical retirement planning, but you haven't confirmed your target retirement age.
I'm also assuming your bond allocation in the 401(k) is intentional rather than a default — can you confirm which account currently holds your bond funds versus equity funds?"
That question was the whole ballgame.
I checked my 401(k) allocation, and it was inverted from what it should have been — my bond funds, which throw off taxable interest, were sitting in my taxable brokerage account, while growth-oriented equity funds sat in my tax-advantaged 401(k).
It's a textbook asset-location mistake. Fixing it doesn't change your returns; it changes how much of those returns the IRS takes.
The Number That Made Me Cancel the Call
I ran the corrected allocation through both models and then had my (now former) advisor confirm the math on a call.
**The asset-location error was costing me approximately $1,900 a year in unnecessary taxes** — money that had been quietly leaking out of my accounts since 2023, invisible on the quarterly PDF because the PDF only showed performance, never tax efficiency.
Over three years, that's close to $6,000. My advisor's fee over the same period was north of $10,000.
I want to be precise about what happened here, because it's easy to overstate: the AI didn't outperform the market, predict a stock, or do anything a good fee-only fiduciary advisor with a CFP designation couldn't have caught in five minutes.
What it did was ask a structural question — "what am I assuming?" — before generating an answer, and my advisor, in three years of meetings, never once did that out loud.
The Reality Check: Where This Breaks Down Fast
Here's where I have to stop you before you close your brokerage app and start taking retirement advice from a chatbot.
Neither ChatGPT 5 nor Claude 4.6 is a fiduciary. Neither one is legally accountable if the advice is wrong.
When I asked Claude about a specific tax-loss harvesting strategy involving a wash-sale window, it gave me an answer that was subtly wrong about the 30-day rule when substantially identical securities were involved across my Roth and taxable accounts — an error a real CFP caught in about four seconds.
**Models are also bad at things that require knowing your full life, not just your account balances.** They don't know your job security, your marriage, whether your parents will need care in five years, or whether you're the type of person who panic-sells at the first 20% drawdown.
My advisor, for all his blind spots on tax efficiency, did know that about me, because he'd asked.
There's also a selection bias I have to own: I only ran the numbers that were easy to verify against public tax rules.
Asset location is a solved, mechanical problem — exactly the kind of thing an LLM trained on tax code excels at.
Retirement planning involving actual uncertainty about your future self is a different animal entirely, and I wouldn't hand that over on faith.
The Takeaway: How to Actually Use This
If you're going to try this yourself, don't just paste your portfolio into ChatGPT and ask "is this good." That question invites a confident, generic answer — the AI equivalent of small talk.
Here's the workflow that actually produced value for me:
- **Ask it to state its assumptions first**, before it gives you any recommendation. If it doesn't ask clarifying questions unprompted, force it to.
- **Feed it real numbers** — account types, cost basis, tax bracket, actual holdings — not vague descriptions. The value is in the specificity.
- **Cross-check anything involving tax law or legal rules** against a second source.
Wash-sale rules, contribution limits, and estate law change yearly and models get details wrong with total confidence.
- **Keep a human for anything involving your actual life circumstances** — job changes, family, risk tolerance under real stress.
That's not a tax question; it's a "who are you" question, and it deserves a real conversation.
- **Treat this as a second opinion, not a replacement.** The value I got wasn't "fire your advisor." It was "verify your advisor," which is a much cheaper and much less controversial use of the technology.
I didn't cancel my advisor relationship because AI is smarter than a CFP.
I canceled it because a free five-minute conversation with a model caught something a paid annual relationship didn't, and once I saw that gap, I couldn't unsee it.
Have you actually checked your own portfolio's tax efficiency, or are you trusting the quarterly PDF the same way I did for three years? What did you find when you asked?
