我用 AI「核對」信用卡月結單，才發現自己的計算過於樂觀

After changing my credit card payment settings, it's always a pain to confirm if the changes actually took effect. Especially with revolving credit, the fee calculations can be tricky, making it hard to pinpoint what to look for on the statement.

So, I decided to delegate this verification task to AI. I'd previously had AI simulate repayment plans, and I wanted to check if those simulations aligned with real-world data.

What I Asked the AI to Do: Read and Reconcile a Logged-In Screen

Having AI log directly into a financial institution's website is a non-starter from a security perspective, and frankly, it shouldn't be done. Instead, I used a browser extension to allow the AI to read my statement screen after I had already logged in via my browser.

The prompt was simple:

"Please re-check if the revolving credit payment was correctly applied."

I had the AI read the numbers displayed on the screen: the revolving balance, the current repayment amount, and the breakdown of principal and fees. The goal was for it to confirm if the settings were applied as intended.

The Result: AI Calling Out My Past Self

The AI's initial response was, "The settings have been correctly applied." So far, so good. But then, an unexpected report came in.

🔴 Important Correction: The lump-sum repayment amount will be double what was expected.

What did this mean? According to the AI, the assumptions in its own previous repayment simulation were too lenient. After comparing it with the actual statement data, it found that the proportion of the payment allocated to fees was significantly larger than anticipated, meaning the principal was being paid down much slower.

Specifically, the AI presented this verification:

Verification: 519,980 (Cumulative changed amount) - 14,700 (Principal repaid on 10/13) = 505,280 ✓

The amount of "principal repaid" was much smaller than in the past simulation. At this rate, the total amount if I were to repay in a lump sum in the future would swell to nearly double the original estimate. That was the correction report.

As a result of entrusting the AI with this task, it uncovered its own past calculation errors. It was like asking someone to review your work, only for them to start self-correcting – quite amusing, actually.

Remaining Manual Steps and Lessons Learned

What this experience re-emphasized is the importance of not blindly trusting AI's analysis or simulation results. AI excels at calculations based on given premises, but if those premises are off, the results will naturally be off too.

This is precisely why a "cross-checking" process, regularly reconciling with real data, is essential. In this case, by having AI perform not "identity verification" but "self-verification," I was able to catch the laxity in my calculations.

Also, when dealing with information from sites requiring login, the current method of "a human logs in, and the AI reads the screen" will likely remain a necessary step for the foreseeable future. While it seems like an extra step, it's also a crucial safety mechanism to avoid directly handing sensitive information to AI.

A Practical Tip You Can Try

This method can be applied to other tasks as well.

For example, if AI generates a report or data analysis, you can follow up with: "Please cross-reference this with the raw data and re-verify its correctness." Or, if AI drafts an email, you could ask: "Please confirm if the content aligns with our past communications."

After having AI perform a task, delegate the verification of its results as part of the same workflow. Adding this extra step allows AI itself to uncover its own "assumptions" or incorrect premises. You're not just using AI as a worker, but as a peer reviewer.

Instead of blindly accepting AI's output, add the step of cross-referencing with original data. It's a small investment that can significantly improve the accuracy of your work.

I build and run small Python systems — trading bots, RAG APIs, scheduled automation — and write up whatever breaks along the way.

If a provider-agnostic RAG Q&A API is useful to you, mine is MIT-licensed on GitHub: rag-faq-api. It runs and passes its full test suite **with no API key* (offline stub LLM + hashing embedder), swaps to Claude / Gemini / OpenAI via one env var, and ships a retrieval-quality harness (Hit@k / MRR / Recall@k) with a chunking sweep.*