Have you used AI with your finances?

Conversation
1

I’ve been tracking the growing trend of using Generative AI within personal finance, and I recently came across a pair of fascinating CNBC articles discussing a new study published in the Journal of Financial Planning. The data shows a massive shift in how people approach their money, noting that roughly 66% of Americans (and up to 82% of Gen Z and Millennials) who use GenAI have leveraged it for financial advice.

However, the study—which tested seven major platforms including ChatGPT, Claude, and Gemini—raised some critical structural warnings that are worth discussing, especially for a data-conscious community like ours:

The Illusion of Authority: Andrew Lo, director of MIT’s Laboratory for Financial Engineering, noted a significant concern with LLMs: “no matter what you ask it, it’ll always come back with an answer that sounds authoritative, even if it’s not.”

The Precision Trap: The research explicitly pointed out that, counterintuitively, AI isn’t inherently great at crunching numbers or performing precise financial calculations. Asking it to run a numerical analysis on your own specific situation remains highly risky.

The Efficiency Gap: The process often devolves into an unstable sequence of trial and error—sometimes requiring more than 20 prompts—leaving the user with an output they still have to manually double- and triple-check against their actual data.

As a community that thrives on clean data pipelines, it makes me wonder: Why leave the deterministic, 100% accurate outcome of a structured database or spreadsheet for a probabilistic system that requires manual auditing and a 20-prompt guessing game?

Have you tried incorporating ad-hoc AI prompts into your financial workflows? Where do you draw the line between a helpful linguistic tool and a data integrity risk?

Would love to hear your thoughts!

CNBC article published 7/7/2026

CNBC article published 4/18/2026

2

Very interesting findings, and I can concure that I fall into the same belief, that AI is more authorative and that it should be good with numbers. It’s good that so far Tiller isn’t using it to do anything with the numbers, it’s using it to categorize, something it can be much better at. This makes me happy I’ve been pumping the brakes on using AI for my finances, and will likely drag my feet a bit longer until there’s evidence that it can be trusted.

3

That study seems spot-on. I am trying different budgeting templates. I wanted to test AI to see if it could run some calculations. I copied a column of my current 2026 ‘actuals’ from one of my budgets and pasted it into ChatGPT. I asked it to look at the totals of each cell, increase each total by 2.8% (2027 projected inflation rate), and list the new totals in a separate column. The calculations were correct. The formatting was a train wreck, and it seemed to have an issue separating income from expenses. I fixed the formatting easily enough. However, you are correct. In the time I spent doing all of that, I probably could have easily figured out what formula to use and how to apply it to the data I wanted to change. I did this mostly out of curiosity but also because I am somewhat new to spreadsheets and budgeting. I’m a gold-card carrying member of the ‘I’ll just ask AI’ club. I generally surmise that AI is smarter and quicker and tasks I don’t know how to do. But in all of the time and eventual success I had, I’m not sure I learned anything except that I am dependent on AI for a lot I could probably do myself. I hope this isn’t TLDR, but your post clicked with me. Thanks!

4

@jpfieber Please realize that I am not anti-AI. I like you use it for many things! I think it’s wondrous technology! But when it comes to finances my conclusion , like yours, based on evidence is that it’s not ready for prime time. Thanks for sharing!

5

⁠@lisaluisi⁠ Thank you for sharing such a transparent and perfect example! You hit on the exact core of the issue: the hidden cost of time and the absolute necessity of auditing the output.

Even when the underlying math happens to be correct, the time spent copying, pasting, correcting a formatting issue, and manually verifying that it didn’t mistake an income line for an expense completely erases any perceived shortcut. As you beautifully pointed out, a simple deterministic spreadsheet formula (like ⁠=A1*1.028⁠) gives you an instant, 100% reliable result on the first try—and more importantly, it helps you build confidence in mastering your own data.

There’s absolutely nothing wrong with curiosity or being part of the “I’ll just ask AI” club for text or ideas, but stories like yours show why structured sheets and databases remain the gold standard for money. Really appreciate you adding this insight to the thread!

6

Formula copied and pasted. Thanks! I’d like to get better at manipulating data on Google Sheets. Any good recommendations?

7

To be honest I’m not a spreadsheet guru. So I wouldn’t be one to ask. But you are in the right community there are many out there who are better at formulas than me. I’m more of a database guru. Middlin spreadsheet user. Hope this doesn’t disappoint. :smiley: (AI is good at formulas and programming in general as you may know)

8

I would say that as you are an ask AI club member, you can ask AI; I ask AI lots of questions to help me figure out the formula I need or double check I’ve done things correctly. I don’t have my data in AI, but it has definitely helped figure out things for getting data results. I have found that sometimes I beat my head against a wall with the answers, but that then makes me look at the data and ask again if needed. I have a few times gotten the issue solved by looking at the formulas in other sheets then applying it where I need it or asking AI to help update the formula to what I need to do.

9

@casilverthorn.96 Spreadsheet formulas and programming as well as knowledge questions are things that AI is good at. To help you get to the answer. Problems arise when people think it’s a magic elixir or wand to wave at financial advice based on pure mathematical outcomes. The technology is just not there. It was not designed for that. But because it’s in fashion now for companies to boast of their AI capabilities for marketing purposes it can lead to bad outcomes. It is covered in the articles what AI is designed for!

10

I see you got the summary above from the two articles you linked for precision, but couldn’t find it in the actual paper linked: Do Different Generative Artificial Intelligence (GenAI) Tools Provide Different Financial Recommendations? | Financial Planning Association
also watched the tickok that was attached, so it was probably in a longer form interview.

I rather disagree and if anything he was probably saying a different thing , computers are great at crunching numbers or performing precise financial calculations. it’s the recommendation and application of laws and situations where errors occur, ironically that’s where he mentions that’s also where humans also fail.

11

@ctlee Thanks for joining the thread! It’s true that computers are great at crunching numbers. AI is a software (Large Language Model or LLM) designed to predict the next token (or word) not for crunching numbers. Here is text from the 4/18/2026 CNBC article.

IMG_1984
IMG_19841206×2622 393 KB

It’s true that sometimes AI gets it right (calculations). But not all the time and requires the data to be doublechecked and triple checked. People don’t bother to check it and take it at its word that’s one of the problems!

12

Is this from the paper or one of the articles?

I agree that applying complex financial laws and personal situations is incredibly difficult—and yes, human advisors can fail at this too.

But there is a massive difference: a human advisor is legally bound by a fiduciary duty and carries malpractice insurance. If an AI recommendations engine gives out-of-date or hallucinated advice based on incomplete data (or bad prompting), the user has zero recourse when the IRS or audit comes knocking.

Thank you for pointing this out as well, it highlights why relying on probabilistic models for high-stakes financial decisions is so risky!

Hope I understood your point correctly. Thanks!

13

from the short interview video in the article you linked

Also i think we’re all thinking of different examples and nuances when a bunch of these statements are made,
sure, the llm itself is a prediction engine which makes it’s math as error prone as anything else, but when it codes to do the math [2310.03731] MathCoder: Seamless Code Integration in LLMs for Enhanced Mathematical Reasoning ie python seems like it acts like a calculator.

But there is a massive difference: a human advisor is legally bound by a fiduciary duty and carries malpractice insurance.

yes and no; I’m assuming you mean fiduciaries are bound by fiduciary duty, most advisors are not, there has been changes since the 2016 DOL rule making. If you meant they had duty of care, then yeah

14

@ctlee Thanks. Unfortunately I am not a TikTok subscriber so I didn’t see the video interview and didn’t read that in the article. Did I get the gist of his (your) point? Or was I off base? I apologize if I was countering the wrong thing without seeing it -David

15

That’s interesting, i wonder if you have an adblock that’s blocking it? because those videos like other embed videos play without an account. (side note i had edited my response above since you posted). I believe you got the gist of it; i think we’re just arguing on different nuances?

16

⁠@ctlee⁠ Thanks for sharing that paper! It actually illustrates the exact point I’m making perfectly.

If we look at the results in that abstract, the absolute state-of-the-art model they developed still only scores 83.9% on grade-school math (GSM8K) and a mere 45.2% on advanced math.

Even if we treat Python execution as a “calculator,” a system that has a 16% failure rate on basic grade-school arithmetic is fundamentally unsafe for personal finance. In accounting and database administration, we don’t tolerate a 16% margin of error—it has to be 100% deterministic, 100% of the time.

As for the legal semantics—whether an advisor is a bound fiduciary or held to a duty of care—the core point remains: there is a real human carrying professional liability. If an AI “calculator” hallucinates a tax bracket or misses a decimal point on an investment calculation, the user holds 100% of the liability and the financial loss.

I think we can both agree that AI is a fantastic tool for code syntax, but when it comes to executing the final numbers on a financial ledger, nothing beats a deterministic system. Thanks for the great debate!

IMG_2022
IMG_20221206×2622 488 KB

I’m not even going to say that this does not have a future. But right here and now it’s flawed. -David

17

@ctlee Also the video finally played for me once your editing was done! So you were correct I don’t need TikTok to see it. I caught the tail end of it though. So I didn’t see the part where Lo made that comment about human advisors. But I believe you when you say that he did! Not sure if the article follows the video word for word but it’s worth reading. Hope you do read both!

18

I need to explain that even though models are using Python scripts to make accurate math calculations - when it comes to financial planning, tax preparation, and accounting aggregations on flat files it’s lacking because they incorporate nuances, rules, and gray areas the model is not designed to handle with its scripts. But with a proper semantic model with SQL and DAX cues some of this could be overcome.

By defining the business rules, data relationships, and accounting measures explicitly in a relational backend before the AI ever touches it, you remove the guesswork. The SQL database enforces the structural schema, the DAX measures handle the precise financial logic, and the AI is left doing what it actually does best: translating natural language into clean, deterministic queries.

A Python sandbox gives an LLM a calculator, but a semantic layer gives it a brain. Until people combine the two, running AI blindly over raw flat files or without rules for it to apply, it remains a high-risk gamble for personal finances.

19

I found this take by @thomp679 to be interesting

Boldin and Claude by thomp679

Boldin is a Smart Money/Retirement Asset calculator that has a free version and a paid version (I think that runs.144.00) and 2800.00/yr if you want a CFP involved. It’s deterministic and pretty popular.

@thomp679 tells of his journey to get Claude working with Boldin verifying the results. It took many attempts from what I gathered from his post.

Another deterministic retirement software is marketed by Lawrence Kotlikoff called MaxiFi planner. I’ve actually used it in the past (pre-AI boom) and found it to be useful in my own situation. Maxifi planner is 149.00/yr and is his most popular tier.

If there are any others you used or are using AI prompts (or products) to get financial planning done feel free to post in this thread!

20

Early on, when ChatGPT was the big new thing (fall 2022), I sometimes took a screen shot from Excel or Google Sheets and just pasted it into Chat GPT and asked it things like, “please suggest the formula to increase every data item in Column G by 2.8%. What would I put in cell H2 to get this started?” If appropriate, I scrambled the data a little bit for some added privacy.

With Google Sheets recently, I think Claude and Google Gemeni can access your google sheet directly if you let it (again, privacy concern…), so you can say, “please look at https://googlesheeets/MyBudget/blahblah. Create a new worksheet in that workbook that will help me visualize my spend vs my income. I prefer bar charts over line graphs”.

To summarize, all of the AI tools are pretty good with spreadsheets if you just ask them for what you want. Unlike a few of the comments here, I haven’t seen anything really botched up when it comes to formatting or math, though I’d still check things carefully for sure. I have seen it just not do what I wanted, so I asked again and tweaked things until I had what I needed. (Sometimes I asked AI to make the changes, other times it was quicker to make them myself).

Also, you can use the AI tool to help you learn the spreadsheet skills. AI is a pretty good teacher. So if you ask, "I bought an item for $1000 in 2010 (cell G10). What’s the formula I could put into later cells in row 10 to show the value of that $1000 in the subsequent years? Then, ask it to explain the math behind that formula, and it will. :slight_smile:

Or, “Please look at the data on this worksheet** and suggest which chart format is best for my audience which is comprised of (whoever is in your audience).”

** For your worksheet, either upload, paste in the sheet or link to it. Create a copy with fake data if you need to for privacy.