Personal finance is mostly arithmetic and decisions. AI is genuinely useful for the first part and surprisingly bad at the second. Treat it as a financial advisor and you'll get bland advice that sounds confident and means nothing. Treat it as a fast calculator with opinions and you can save a real amount of time.
I've been running my budget through Claude and ChatGPT for about a year. Here's what works, what doesn't, and where I draw the line on what I share.
Don't connect AI to your bank
Skip this. Several apps now offer "AI-powered" budgeting that links to your accounts via Plaid and feeds your transactions into a model. The pitch is convenience. The risk is that every coffee, paycheck, doctor's bill, and PayPal nudge to your kid's college roommate becomes training data, gets stored in a vendor's database, and shows up in support engineer queries when something breaks.
You don't need that. Export a CSV from your bank. Strip account numbers. Paste it into ChatGPT or Claude. You get 95% of the value without the integration surface.
If you're using a tool like Cleo or Copilot Money that markets AI features, read the privacy policy. Look specifically for whether transaction data is used to train models and whether it's shared with third-party LLM providers. Both have made changes here that are worth rechecking annually.
Three things AI does well
Categorizing transactions. Paste a month of transactions and ask for them grouped by category: groceries, restaurants, subscriptions, transport, and so on. Claude is noticeably better than ChatGPT at this because it doesn't try to over-explain. Ask for a CSV back. Now you have something you can sort and filter in any spreadsheet.
Spotting subscription creep. Once your transactions are categorized, ask: "Which of these are recurring monthly charges?" You'll get a list of subscriptions you forgot you had. The first time I did this I found four I wasn't using. $43 a month, $516 a year. The model didn't tell me to cancel them. It just made them visible. I made the calls.
Stress-testing decisions. "I'm considering a $2,400 expense. My emergency fund is $X, monthly expenses are $Y, here's my income. Walk me through the trade-offs." This works because the math is rote and the prompts are clear. The model can't tell you whether the expense is worth it, but it can show you what it does to your runway and whether you can absorb a job loss after.
Three things AI does badly
Picking specific investments. Anything that sounds like "should I buy NVDA right now" gets a wishy-washy answer at best or, worse, a confident wrong one. Don't use it for this. Use a real advisor or read a primary source.
Tax advice with edge cases. Generic tax questions (what's a Roth IRA, how does the standard deduction work) are fine. Anything specific to your situation, especially across state lines, with self-employment income, or with crypto, verify with a CPA. The model will give you an answer that's 80% right, and the 20% will cost you.
Budget targets pulled from thin air. If you ask "what should my budget be?" you'll get the 50/30/20 rule restated. That isn't advice. Your budget depends on where you live, what you owe, and what you want. Start with your actual spend, not a Reddit heuristic.
A workflow that holds up
Once a month I do this in about fifteen minutes:
- Export last month's transactions from my checking and credit card.
- Open Numbers or Excel, strip the account numbers, save as CSV.
- Paste into Claude with: "Categorize these transactions. Return as CSV with columns date, vendor, amount, category."
- Paste the result into a spreadsheet I keep year over year.
- Ask Claude: "What changed from last month? Anything that looks unusual?"
That last prompt is the one that earns its keep. The model catches the gym charge that doubled because of a "facility upgrade fee," the streaming service that bumped from $14.99 to $17.99 quietly, the $400 Amazon order I forgot was for a dishwasher.
Tools worth a look
Most "AI finance" apps are wrappers. The exceptions worth trying:
- YNAB isn't AI-native, but pair it with a Claude-assisted monthly review and you get the budget discipline of YNAB plus the pattern-spotting of an LLM.
- Monarch Money added LLM-powered categorization recently. It's decent at recurring charges and its privacy stance is better than most.
- Beancount plus your own scripts, if you're technical and want full control. Pipe your ledger to a local model via Ollama and never let your data leave the machine.
The pattern across all of them: keep the AI on the analysis side, keep the account access narrow, and never let a chatbot push a transaction.