Which tool to choose
There's no magic app that cuts expenses for you: you need the right AI for your goal.
- You want to understand where the money goes and what to cut, right away and free → a general-purpose chatbot (ChatGPT, Google Gemini, Claude). You upload the statement, it analyzes and tells you what to eliminate. It's the fastest route and it works in Italy too, because it works on the file you give it, without connecting to the bank.
- You want to keep expenses under control every day, not just once → a tracking app localized in Italian. Spendee is among the most appreciated for managing personal and family finances, has a well-translated interface and lets you monitor income and outflows manually or with bank synchronization in the Premium plans. Money Manager is the alternative with bank sync in the Premium version and a complete free version for manual management.
- You're buried in subscriptions and have lost count → AI on your statement finds them all in one go. Dedicated apps like Rocket Money do it for a living, but the automatic account connection and the cancellation service are designed for US banks. In Italy it's best to have the AI find the subscriptions and then cancel by hand.
For most people the starting move is the first: the chatbot on your statement. The rest of the guide follows this path.
How to do it
From computer or phone the path doesn't change, except where noted.
- Download the statement. Log into your bank's online banking and download the last 2-3 months of transactions. Look for a button like "Export", "Download transactions" or "Transaction list". If you can choose the format, take the CSV (a tabular file, readable by the AI without errors); if the bank only gives PDF, that's fine too.
- Remove the data that identifies you. Before uploading, delete or hide name, full IBAN, address and card number. The AI only needs the date, description and amount of each transaction.
- Upload the file. Open ChatGPT (chatgpt.com or app), Gemini (gemini.google.com) or Claude (claude.ai). Use the paperclip icon to attach the file. If the version you use doesn't let you upload attachments, open the file, copy the transaction text and paste it directly into the chat.
- Paste the prompt below and send.
The prompt to paste:
You are a household budget advisor. I'm attaching the transactions from my account over the last few months (date, description, amount). Do this:
1. Group all outflows into clear categories (housing, groceries, transport, restaurants, subscriptions, shopping, other) and tell me how much I spend per month in each, in euros.
2. List all recurring charges and subscriptions you recognize, with the monthly and annual amount.
3. Indicate the 5 items where I can save the most, with the possible cut in euros per month and how much that adds up to per year.
4. For each cut, propose a concrete action (cancel, reduce the plan, change provider).
Use only the data I gave you: if an amount isn't clear, flag it instead of estimating it.
- Verify the numbers. When the AI gives you back the categories and totals, spot-check 4-5 items: compare the sum of one category against the real transactions. It's the step that separates a real decision from a house of cards.
- Act on the first three cuts. Cancel the subscription you don't use, call the provider for the lower plan, set a cap on the category that exploded. The saving becomes real only here.
Expected check: after point 4 you should have in front of you a table with your spending categories in euros and a numbered list of quantified cuts ("Cancel X: −€12/month, −€144/year"). If the AI gave you generic advice without figures, reply: "Give me everything in euros per month and per year, using only my transactions".
A concrete example
Marco, an office worker, is convinced he spends "around €200 a month" between dinners and coffees. He downloads three months of transactions from his account in CSV, hides IBAN and name, uploads to ChatGPT with the prompt above.
Result: the AI calculates €410 average a month between restaurants, bars and delivery, more than double his estimate. In the list of recurring charges, two music streaming subscriptions pop up that overlap (one taken on a trial and never canceled, €10.99/month) and a gym he hasn't been to since January (€39/month). Cuts identified: double streaming (−€132/year), gym (−€468/year), a €250 cap on eating out (−€160/month, −€1,920/year). Marco cancels the two subscriptions that same evening and imposes the limit on eating out. Saving put into practice the first month: over €200.
When it does NOT work (and how to fix it)
If the AI makes up the numbers
It's the most serious risk: with PDFs the AI sometimes gets things wrong or creates transactions out of nowhere. It crunches the document and produces a table that is sometimes accurate, sometimes confidently invents nonexistent transactions or stops halfway through the file. Fix: give the file in CSV instead of PDF (the numbers stay exact), work on a few months at a time and spot-check the category totals against the real transactions before deciding. If the numbers don't add up, ask the AI to redo the extraction item by item.
If you're worried about privacy
Uploading your own transactions to an external service is a legitimate concern. Practical fix: the AI doesn't need to know who you are. Leave only the three columns date-description-amount, remove IBAN, name and card numbers. The transaction description ("SUPERMARKET PURCHASE", "SUBSCRIPTION X") is enough for the analysis. For an extra level, in ChatGPT you can turn off the use of your chats for training in the Data Controls settings; Gemini and Claude offer an equivalent option. If you need absolute confidentiality, run a model locally (for example with LM Studio or Ollama): the data never leaves your device.
If the bank only gives you a messy PDF
Some banks export PDFs with overlapping columns, unreadable. Fix: first ask the AI to do only the cleanup work, then analyze the resulting table. Attach the PDF and use this prompt:
Extract from this bank statement all transactions into a table with columns Date, Description, Amount, Type. Format the result as CSV, so I can paste it into Excel. Don't estimate or invent items: if a piece of data isn't readable, write "to verify".
Check the table, fix the errors, and only then launch the expense analysis.
If it connects your bank but you're in Italy
The most automated apps (Rocket Money first) shine in the US, but the account connection and automatic cancellation don't cover Italian banks. Fix: use the AI to identify the subscriptions from the statement, then cancel by hand from the service's app or the store where you subscribed. For ongoing monitoring, rely on a localized app like Spendee or Money Manager.
A tip from someone who actually uses it
The one-off analysis uncovers the waste, but doesn't keep it away: after two months the subscriptions reappear and the dinners explode again. The trick is to turn it into a monthly ritual. Save the prompt in a note on your phone and on the first of every month repeat the operation on the statement of the month just closed. Ten minutes and the AI tells you right away whether a category has grown compared with the month before. Manual tracking increases awareness because you review every purchase, while automatic tracking is more convenient but leads to the "set and forget" attitude. The monthly analysis sits in the middle: convenient like the automatic, aware like the manual.
Frequently asked questions
Which free AI should I use to analyze expenses?
For this task all three major ones work fine: ChatGPT, Google Gemini and Claude have free versions that read uploaded files or pasted text. If you already have a Google account, Gemini is the most immediate; ChatGPT is the most proven for extracting tables from PDFs. The practical difference on analyzing a bank statement is minimal: choose the one you already have open and get started.
Do I really have to connect my bank account?
No, and that's the advantage of the AI method: you work on a file you download and clean up yourself, without giving anyone your bank credentials. Automatic connection is convenient in dedicated apps, but before relying on an app check whether the free version includes data export: if it's missing, the history of your transactions stays trapped the day you decide to switch.
How much can I really save?
It depends on how much waste you've accumulated, but the most common mistake is to underestimate it: almost everyone discovers forgotten subscriptions and one out-of-control category. After entering income and outflows, the analysis shows how much you have left after bills, debts and savings, and flags whether there's an imbalance. Don't expect the AI to cut for you: it hands you the quantified list, you make the cut.
Doesn't the AI risk giving me wrong financial advice?
Here's the misunderstanding to bust: in this use the AI doesn't tell you how to invest nor decide for you. It sums and categorizes your real transactions and shows you where the money goes. The "advice" is arithmetic on your data, not a market forecast. The only real risk is the calculation error or the invention of items, and for that there's the checkpoint: verify the totals of one or two categories before acting. That way you work on your own numbers, not on the opinions of a machine.