Using AI for the Family Budget

On a Sunday morning I start the run, go get coffee, and come back to a finished month. Every transaction from every account is pulled in, checked, and written into our budget spreadsheet. Then my son and I sit down and talk about it.

The second part is why I built the first.

We used to spend the first half hour doing data entry, reading numbers off one screen and typing them into another. By the time we got to the real conversation we were both worn out. Him especially.

Pulling the Month Out of Rocket Money

Our transactions live in Rocket Money, a budgeting app that connects to our banks and cards and collects everything that moved. A few credit cards and two checking accounts.

Claude cleans up the month first, inside Rocket Money. It finds the names that arrived as machine gibberish, reads the rows around them to see what town we were in, searches the fragment along with the town, and comes back with the business. TST*GANSO becomes Ganso Ramen & Sushi. Then it sets a rule so the fix holds backward and forward. Anything it cannot identify, it asks me about.

That step is where an AI earns its place. None of the work is hard. It is ordinary judgment applied a few hundred times: this is a ramen place, that is a barber shop, this restaurant is not a utility bill. A regular program cannot make those calls, so a person had to.

Then a program takes over. It drives a Chrome browser I am already signed in to and lifts the transaction data straight out of the page rather than reading it off the screen. Every file it saves gets read back and compared byte for byte against what it meant to write. If one byte is off, it stops.

No AI touches the numbers. A language model is a strong writer and a mediocre calculator. It wrote the program that gets the numbers. It operates that program. But it is not reading things off the screen or putting them in a spreadsheet. No AI is up to that task, and these are numbers that matter.

Three Sets of Checks Before Anything Reaches the Spreadsheet

The first set asks whether the transactions are right. Does the file hold as many rows as it says it does. Do the date windows cover the month with no gaps. Are there duplicates. Does our spending total match Rocket Money's own total to the penny, and does every single category match, since two categories can be wrong in opposite directions and still add up to a total that looks fine.

The second set asks whether anything is missing, which is a different question and a harder one. Has an account gone quiet compared to its own normal rhythm. Has a category that always has spending suddenly read zero. Did every account actually sync in the last three days, and is its bank connection healthy.

The third set checks the balances against the transactions. Take what the bank said a checking account held before the period started, subtract everything we recorded since, and see whether it lands exactly on what the bank says today. Both checking accounts landed to the penny.

Only after all of that does anything get written into the sheet, and then the sheet gets read back out a different way and compared against the source. Last time, seventy checks, no failures.

Here I Am, the Human in the Loop

The process ran end to end and said everything was accounted for and my spreadsheet was fully up to date. So I went and checked it myself.

I checked the categories against Rocket Money. They matched. I checked the bank account totals. They lined up. Then I checked the credit cards, and the numbers in the sheet were off. They said we owed less than we did.

I still do not know exactly why. Probably something that had not fully synced, or a different idea of how a pending charge on a card gets counted. It does not really matter. What I care about is that my process let me spot it easily and fix it before we made decisions on bad numbers.

Whatever you have AI automating or supercharging for you, at some point you need a person in the loop to keep it tight and on the rails.

Asking About August in Plain English

Once the month's transactions are complete and checked, I can ask questions in plain English and trust the answers.

August came in well over budget, which looks like a catastrophe. So I asked what was in it, and named the things I already knew were one-time: a dentist, a vacation, a college move-in, a birthday, an annual backup renewal, a recorder I bought, a tuition payment, interest charges. Strip those out and the month lands about even against what we earned.

Then the finding I would never have gotten from a summary. Nearly everything still over budget was dining. Every other category combined came in under.

So it was not a disaster month. It was an ordinary month with a dining problem and a pile of scheduled large expenses that all landed at once. That is a completely different conversation to have with your family, and getting to it took about two minutes.

Other things become askable. What were the ten biggest things we bought last month. Which categories have been drifting up since spring. If we had skipped the vacation, where would we have landed. What are we paying for every month that we forgot we signed up for. Given what is in checking, what is the largest card payment we can make without going under.

A Smaller Version You Can Build First

I am not going to tell you this is a weekend project. The extraction piece takes real browser automation, and I could build it because I have spent twenty years around software. The honest framing is that this is the shape of what is reachable now, not a thing you download on Sunday and run on Monday.

You can get most of the way with far less. Pick the app your accounts already connect to and let an AI walk the list with you, cleaning up names and categories, so the numbers underneath deserve your trust. Export what it gives you rather than retyping it. Check that every account actually synced before you believe any total. Then ask the AI to talk you through the export, which it does well, and keep it away from arithmetic, which it does not.

Do that much and Sunday gets shorter and the conversation gets longer, which is why I started.

Thirty Minutes Back With My Son

The thirty minutes of typing is gone. What is left is why we sat down in the first place. My kid and I look at what our family actually did with its money and work out together what we would do differently. He is old enough to ask hard questions about it, and now he has more patience for them, and so do I. We get further in than we used to, which is what I built the machinery for.