I Asked 3 AIs to Find My ‘Hidden’ Bank Fees. One Found £180.

The Introduction

I hate bank fees. Not the obvious ones, like the £5 monthly charge for a “premium” account I definitely don’t need. I’m talking about the stealth fees—the foreign transaction markups, the “notification” charges, the poor exchange rates.

Last month, I traveled to Europe and used my debit card. I thought I was being smart. But when I got my statement, the numbers felt… off.

I wanted to audit my last 3 months of statements to find every single penny of wasted money. But staring at a PDF for 3 hours is my idea of hell.

So, I ran an experiment. Could an AI do the audit for me?

I pitted ChatGPT-4o, Claude 3.5 Sonnet, and a specialized finance AI called “Cleo” against my raw bank data. The results were genuinely shocking.

The “Level 1” Prompt (The Failure)

First, I tried the lazy way. This is what 99% of people do, and it’s why they think AI is useless for finance.

I copied my statement text and pasted it into ChatGPT with this prompt:

The Lazy Prompt:

"Read this bank statement and tell me if there are any hidden fees."

The Result: It gave me a polite, generic lecture. “As an AI, I can see you have some transaction fees, but I cannot determine if they are ‘hidden’. You should check your bank’s terms and conditions.”

The Critique: Useless. It was too scared to give an opinion. It just defined what a fee is without analyzing my data. Score: 1/10.

The “Prompt-First” Strategy (My Secret Sauce)

To get a real answer, I had to treat the AI like a forensic accountant, not a chatbot. I used my “Forensic Audit” prompt chain.

Step 1: The Data Cleanser (Safety First)

Before I fed anything to the AI, I needed to make it safe. I didn’t want to upload my account numbers.

My Prompt:

"Act as a data privacy expert. I am going to paste a bank statement. First, write a Python script to strip out all account numbers, addresses, and names, but KEEP the transaction descriptions, dates, and amounts. Do not analyze yet. Just clean the data."

Why this works: It forces the AI to acknowledge privacy first. I ran the script locally, and then I had a safe dataset to paste back in.

Step 2: The “Fee Hunter” Classification

Now for the real work. I needed to force the AI to categorize “stealth” fees, not just obvious ones.

My Prompt:

`”Now, analyze the cleaned data below. I want you to categorize every single transaction into 3 buckets:

1. ‘Essential’ (Rent, Bills, Groceries)
2. ‘Discretionary’ (Eating out, Shopping)
3. ‘Friction Costs’ (This is the important one. Include ANY transaction fee, non-sterling transaction fee, ATM fee, or subscription that looks like a duplicate).

Output the ‘Friction Costs’ as a table with columns: Date, Merchant, Amount, and ‘Suspected Reason’.”`

Why this works: By inventing the category “Friction Costs,” I forced the AI to look for waste, not just “fees.”

Step 3: The “Forex” Audit

This was the killer step.

My Prompt:

"Focus on the transactions marked 'Non-Sterling'. Compare the amount charged to the historical GBP/EUR exchange rate on that specific date. Calculate the 'spread' percentage I paid vs. the market rate. Is it higher than 1%?"

The Results (The “Wow” Moment)

ChatGPT (using Code Interpreter) chewed through the data in seconds. Here is what it found using my “Forex Audit” prompt:

• The “Non-Sterling Transaction Fee”: I knew about the £2.50 per withdrawal fee.
• The “Hidden” Spread: What I didn’t know was that my bank was also charging me an exchange rate that was 2.8% worse than the market rate.

The Math:

• Total spent abroad: £2,000
• “Visible” Fees: £15
• “Hidden” Exchange Rate Markup: £56

It also found:

• A duplicate Spotify subscription (I was paying for my own AND a Family plan). (Cost: £10/month)
• A “inactive account” fee on an old savings pot I forgot about. (Cost: £5/month)

Total “Friction Cost” found: £180 over 3 months.

Dixon’s Verdict

This experiment paid for itself 10 times over in five minutes.

• ChatGPT-4o (with my prompts): It was brilliant at the math and the currency conversion. It found the “invisible” exchange rate fee that I would never have calculated manually. (9/10)
• Claude 3.5: Great at categorizing the subscriptions, but it couldn’t do the live currency conversion math as well. (7/10)
• Cleo: Good for a general overview, but it didn’t catch the specific forex spread. It’s better for budgeting, not auditing. (6/10)

The Takeaway: Don’t just look for “fees.” Look for “friction.” And use the Forex Audit prompt if you’ve traveled recently—you are probably losing more than you think.

Call to Action

I’ve added the exact “Forex Audit” Prompt I used to my AI Toolkit Page so you can copy-paste it.

If you want to stop paying these forex fees entirely, check out my review of Monzo vs. Revolut (Spoiler: One of them would have saved me that £56).

Follow me on LinkedIn for next week’s experiment: I try to use AI to lower my car insurance.