I connected ChatGPT to my Amex accounts this week and asked it one question: where am I leaking money?

It handed me a "potential savings tracker." Four line items, each with a recommended action:

1. Bolt, $240 a year. Cancel if unused.
2. Experian, $324.96 a year. Downgrade if the paid benefits are unnecessary.
3. Hearst Connecticut Media, $431.52 a year. Cancel if unused.
4. An extra Amazon charge, $143.28 a year. Cancel only if it's a duplicate monthly membership.

Total: $1,139.76 a year. I've already addressed the first three, which is $996.48 back in my pocket.

Then it added a bonus: it suggested I verify my Amex digital entertainment credit first, because that credit might already offset several of my subscriptions.

Reader, I have two teenage boys and a teenage girl. We subscribe to every streaming service. Literally every one. Enrolling in that credit is roughly another $300 a year I was leaving on the table.

So call it $1,400 a year, found by a chatbot reading my statements.

What made this useful

Most people use AI to write things. This was a reading job. A month of card statements is exactly the kind of document no human wants to read line by line. I certainly didn't. ChatGPT read all of it, spotted the recurring charges and cross-referenced them against benefits I already pay for through the card.

That last step is the one I would never have done myself. I knew the entertainment credit existed somewhere in the fine print. I had never connected it to the streaming pileup on my statement.

Where I kept my hands on the wheel

Two things I appreciated about the output.

First, it hedged correctly. The tracker said, in plain words, that these were possibilities, not confirmed duplicates or verified savings. The annual figure also assumed every flagged charge was monthly. That's an assumption you want stated out loud, and it was.

Second, it gave me conditional actions. "Cancel if unused." "Cancel only if duplicate." It left the decision with me, which is where it belongs. I still logged into each account and confirmed before canceling anything.

That's the over-eager intern model working the way it should. The intern does the tedious reading and brings you a short list. You make the calls.

Try it this weekend

If you want to run your own version, the prompt doesn't need to be clever:

"Review my card transactions for the last 12 months. List every recurring charge, flag anything that looks unused or duplicated, and check whether any of my card's built-in credits or benefits already cover something I'm paying for. Show the annual cost of each and tell me what assumptions you made."

Ask it to show its assumptions and you'll catch the ones that would otherwise inflate the number.

For leaders, the same move applies to vendor spend, software seats and expense reports. The pattern is identical: a big pile of boring data, a model that reads it all, and a human who decides.

I'm curious what yours turns up. Reply and tell me the weirdest subscription it finds.

Alex

PS Bolt, I'm sorry. I forgot you existed.