How I Used ChatGPT, Perplexity & Claude Sonnet to Audit My Digital Subscriptions—and What I Learned
A hands-on comparison of ChatGPT, Claude, and Perplexity for analyzing business expenses
Digital subscriptions are deceptively easy to activate—and surprisingly easy to forget about. A few clicks, a saved credit card, and suddenly you’re paying for tools or apps you no longer use or barely remember signing up for.
I recently decided to take back control of those expenses. I wanted to find out how much I was really spending each month on software, SaaS tools, and digital services. But going through a bank statement line by line? Time-consuming and tedious.
So I turned to AI to help me do the job. I tested three popular AI models—ChatGPT-4o, Claude Sonnet 3.5, and Perplexity—using the exact same dataset and prompt to compare how well each tool handled a practical financial task.
Here’s what I found.
Step 1: Preparing the Data
I started by exporting the latest monthly transaction report from my business bank account in PDF format (you could also use CSV or XLS). The document included all activity on the account over the past 30 days.
Step 2: Testing ChatGPT-4o
Using the file upload function in ChatGPT-4o (available to logged-in users), I gave the following instruction:
“Here’s a statement from my business bank account. I want to know how much I’m spending on digital goods, software subscriptions, and SaaS tools. Please create a list of these items with the service name, cost, and a short description.”
This required the model to independently identify which expenses were digital services.
Results from ChatGPT-4o:
• Identified 8 digital expenses
• Delivered the results in a clean table, easily exportable to Excel
• Also generated a downloadable CSV file
Strengths:
• Accurate formatting and structure
• Easy to reuse data
Limitations:
• Missed some entries, such as a subscription to Suno
• Needed manual review to catch a few missing items
Step 3: Testing Claude Sonnet 3.5
Using the same prompt and PDF, I then tested Claude Sonnet 3.5. It performed impressively:
• Identified 10 relevant expenses (or 11, combining multiple Google Workspace charges into one line)
• Provided insightful observations on spending patterns
• Delivered results in a well-formatted HTML table, ready for online use
Strengths:
• Detected all relevant digital subscriptions
• Provided the correct total amount spent: 1,149.76 PLN, which matched my own manual calculation
• Offered a professional, publishable table with shareable link
Limitations:
• None noted in this test
Step 4: Testing Perplexity
Finally, I used the same file and prompt in Perplexity. The results were disappointing:
• Incorrectly labeled accounting services (e.g. Infakt) as digital subscriptions
• Missed several key entries (LinkedIn, Facebook Ads)
• Produced an incorrect total of exactly 5,000 PLN, which was clearly inaccurate
Strengths:
• Free and fast
Limitations:
• Misclassified expenses
• Incomplete data
• Major calculation errors
Final Verdict: Which AI Tool Was Most Effective?
Based on this hands-on test, Claude Sonnet 3.5 was the most accurate and complete tool for reviewing digital subscription spending. It not only identified more items correctly but also presented the data in a highly usable format.
That said, even the best AI models require manual review. They are helpful assistants—not perfect analysts. When working with financial data, it’s essential to verify the results and treat AI output as a first draft, not a final report.
If you want to integrate AI into your company’s workflows in a way that’s practical, safe, and critically informed—book an intro call to learn how I can help your team develop effective, real-world AI habits.
