“SAVE” I Asked ChatGPT to Audit My Credit Card Statement (It Found $930 a Month I Was Wasting)

Nobody actually knows what they're paying

West Monroe asked people to estimate their subscription spend, then checked the real total. 89% underestimated. Two out of three were off by more than $200 a month.

CNET's 2026 survey puts the average at $111 a month, up 23% in a single year, with $21 of that going to stuff nobody uses.

It's not because you're careless. 72% of people have everything on auto-pay and 74% say recurring charges are easy to forget. Nobody sends you a bill for the total. That's the whole problem.

What the audit actually found

My statement had 42 transactions on it. $4,395.65 for the month. About $2,787 of that was software for the business.

Found money: $930.98 a month. $11,172 a year.

But the total wasn't the useful part. The tiers were. It didn't dump 15 things on me and say "cancel these." It sorted them by how confident it was, which meant I could clear the safe pile in five minutes instead of arguing with myself about all of it.

The single most useful thing it showed me wasn't a subscription at all. It was an overlap map. Two tools doing the exact same job, both being paid for, both forgotten. My biggest line item in the safe pile was a $179/mo tool that did something another tool in my stack already did. I'd been paying both for months. Nobody was going to tell me.

That's the thing an app that just lists charges will never find. It takes something that can reason about what each tool is for.

How to run it in 30 seconds

Four steps.

The prompt

This is the whole thing. Paste it, then attach your file.

Build prompt: subscription audit and guided cancellation dashboard

Build a polished, single-page subscription audit dashboard that accepts a credit-card or bank-statement CSV upload and turns it into an actionable cancellation plan.

Data handling

* Parse transaction date, merchant, amount, currency, and account details where available.
* Exclude payments, refunds, duplicated foreign-currency reference rows, and non-purchase entries.
* Normalize merchant names and group recurring charges.
* Never invent financial data. Label uncertain merchant matches and explain limitations clearly.
* Identify likely subscriptions, duplicate tools, overlapping services, unexpected second charges, and unclear recurring payments.

Recommendation system
Group findings into:

* Cancel today: high-confidence, low-risk cancellations
* Needs review: recurring charges that may be useful but need the user's judgment
* Billing issue: possible duplicate, unexpected, or unclear charge
* Keep: likely necessary or clearly valuable

For every flagged item, show:

* Monthly charge and annualized cost
* Most recent charge date
* Clear reason for the recommendation
* Confidence label: High confidence, Needs review, or Billing issue
* Any overlapping service or alternative already in use

Dashboard design
Create an impressive dark financial-command-center interface suitable for a screen-recorded demo.

* Use a deep navy or charcoal background with layered gradients, subtle glow, restrained glass surfaces, and mint or electric-blue accents.
* Lead with a bold "Potential annual savings" hero number.
* Show total recurring spend, monthly recoverable spend, annual savings, subscriptions found, and items requiring review.
* Make the recovery plan the main focus, not a generic grid of cards.
* Use excellent hierarchy, contrast, spacing, responsive behavior, keyboard support, visible focus states, and reduced-motion support.
* Do not display card numbers or placeholder product branding.

Required dashboard sections

1. Hero and savings summary
   * Potential annual savings
   * Selected monthly and annual savings that update live
2. Your recovery plan
   * Grouped cancellation recommendations by confidence
   * Checkboxes for user selection
   * Clear explanations and confidence badges
3. Subscription overlap analysis
   * Show services that appear to perform the same job
   * Explain why a service may be redundant
4. Full statement trail
   * Every analyzed transaction
   * Selectable rows
   * Search by merchant or category
   * Quick filters for All, Cancel today, Needs review, and Billing issue
   * Selected statement rows must update the live savings total

Confirmation and cancellation flow

1. Let the user select recommended subscriptions or statement rows.
2. Provide a Review selected action.
3. Show a confirmation screen listing the selected services, monthly cost, and total savings.
4. After confirmation, open a guided cancellation queue.
5. Handle one service at a time:
   * Show the service name, selected monthly cost, and current status.
   * Provide an Open official account button that opens the service's official account or subscription page.
   * Tell the user to log in themselves. Never request, store, or handle passwords, payment details, or one-time codes.
   * Require the user to review the account details and complete cancellation on the official site.
   * Let the user mark the item as Cancelled or Skipped only after their own review.
6. Never claim a cancellation succeeded unless the user confirms it or the official account page visibly verifies it.

Persistent tracking
Persist progress locally on the user's computer so it survives refreshes and reopening the dashboard.
Persist:

* Selected subscriptions
* Cancelled status
* Skipped status
* Items still awaiting review

Make clear that this local progress is saved on the current computer and can be reset by the user.

Two lines in there are doing more work than the rest combined.

"Never invent financial data. Label uncertain merchant matches." Without this, a model will happily decide that some garbled payment-processor line is a gym membership. With it, the thing says "I don't know what this is" and puts it in your review pile. An audit that admits what it can't see is worth ten times one that sounds sure.

"Never claim a cancellation succeeded unless the user confirms it." More on that next.

The part everyone gets wrong

Most people would build this and stop at the list. A list is where good intentions go to die. You read it, you feel bad, you close the tab.

So the prompt keeps going past the list. You tick the boxes. You hit Review selected. You get a confirmation screen with the real total. Then it opens a queue and walks you through one service at a time, with the monthly cost, the reason, and a button straight to the official account page.

And then it stops and makes you do the actual work.

You log in. Not the AI. The prompt explicitly tells it to never ask for, store, or handle a password, a card number, or a one-time code. It cannot cancel anything for you and it should not pretend it can. It hands you the door. You walk through it.

It also won't mark anything cancelled on its own. You tick Cancelled or Skipped yourself, after you've seen it happen on the real site. An AI telling you "cancelled!" when it cancelled nothing is worse than no dashboard at all, because now you think you saved $900 and you're still paying it.

Progress saves to your own computer, so you can close it, cancel three today, and pick up the rest Saturday.

Before you upload anything

Your statement is one of the most revealing files you own. Where you go, what you eat, what you're treating, who you pay. Three rules.

Delete the account number column first. Open the CSV, delete that column, save. The analysis doesn't need it, and the cleanest way to protect something is to not send it.

Check your data controls. In ChatGPT: Settings → Data Controls → turn off "Improve the model for everyone." Do it before you upload, not after.

Business statements give bigger numbers. Mine was a software stack, which is why it came to $930 and not $30. If you're auditing personal spending, expect a smaller number and a much more personal file. Act accordingly.

Where to start

Don't try to cancel everything tonight. That's how people cancel nothing.

Run it. Look only at the Cancel today tier. Ignore the other two completely for now. Sort that tier by cost and cancel the single most expensive thing on it.

For me one cancellation was $179 a month. That's $2,148 a year from one click, and it was 72% of my entire safe tier. The big one is almost always a tool you replaced and forgot to turn off.

Then come back Saturday for the review pile.

All Resources