Why kisir has no AI inside

What “no AI” means here

It means exactly this. Nothing in kisir is a model. When you add an expense, the app divides. When you open the overview, it adds. When you settle up, it runs a fixed procedure over the balances and returns the fewest transfers. Every number on every screen can be reproduced on paper by anyone in the group, and it will come out the same.

There is no receipt scanning, no automatic categories, no spending insights, no chat assistant, no predictions, no “smart” split. The only things the app fetches from a server are your group’s data and, for an expense in a foreign currency, today’s exchange rate, which is a table lookup you can overwrite.

I should also say what it does not mean. It doesn’t mean the app was made without AI tools. The code was written with an AI coding assistant from my descriptions, and I’ll come back to that at the end. The point of this post is what’s inside the thing you install, and the answer is: arithmetic.

Shared money is a checking problem, not a guessing problem

Most of what AI is good at is guessing well. Which email is spam, which word comes next, what’s in the photo. For those problems a system that is right most of the time is a great improvement over nothing.

A shared expense list is a different kind of problem. Four friends, a villa, groceries, a dinner. At the end, who owes whom? There is exactly one right answer, and the whole value of writing things down is that everyone can see it was reached honestly. The question isn’t “what’s the most likely split?” It’s “what did we agree, and does it add up?”

Here’s a small example. Ana paid €120.00 for the villa, for all four. Ben paid €40.00 for groceries, for all four. Cara paid €60.00 for a dinner that only Ana, Ben and Cara ate. Dan paid nothing.

The balances add up to zero, and the fewest transfers that settle them are two: Dan pays Ana €40.00 and Ben pays Ana €20.00. Cara is already even.

Nothing in that table was guessed. If Dan doubts it, he can redo it. If a model had produced it and Dan doubted it, what would he do? Ask the model again?

Where a model would sit, and what it would cost

I went through the places where an AI feature could go in a bill-splitting app, and each one trades a small convenience for something the group can’t check.

Reading the receipt. A scan saves you typing a number. It costs you checking the scan, and the person who paid already has the number in front of them. There is a full post on why kisir does not scan receipts.

Guessing the split. “Ben usually has the steak” is a prediction about Ben. A split between friends should be what they agreed, entered by the person who paid, visible to everyone. If it’s wrong, it’s wrong in a way you can see and fix. Equal, Shares or exact Amounts cover every real case we’ve met, and the editor refuses to save amounts that don’t add up to the total.

Sorting expenses into categories. Nice for a personal budget. For a group trip, nobody has ever asked me what percentage was Transport. They ask who owes whom.

Suggesting a settlement “strategy”. There is one correct plan: the fewest transfers that zero every balance. Offering three and asking you to pick is pretending there is a choice. The settle-up guide shows the procedure by hand, and the app gives the same answer.

Timing the reminders. A notification that arrives when you’re “most likely to pay” is a notification designed around you rather than around the group. kisir notifies you when something happened, and the group’s name is the title.

Each of those features would also need your data on a server where a model can see it. Which brings up the quieter reason.

Less inside means less to hold

A model needs something to work on. Receipt images, your spending history, your habits. Those have to be collected, stored and, usually, sent somewhere.

kisir asks for a name. It keeps your groups: who’s in them, what was paid, who it was for, who paid whom. There is no profile of you anywhere, and nothing is learning from your groups, because nothing in the app learns. The whole ledger exports to CSV or Excel whenever you want it, free. That is a comfortable position for a small app about money between friends, and adding a model would move us out of it.

It also works on a mountain

A model lives on a server, or needs a big download and a recent phone. kisir’s arithmetic runs on your phone. You can add, edit and delete expenses with no connection, in a villa with bad wifi or on a hiking trail with none, and they sync later. There’s a post about adding the expense now and syncing later. An app whose core feature phones home for every receipt would not have had that for free.

What’s actually clever in it

Saying “no AI” can sound like saying “no effort”. The effort went elsewhere.

All money arithmetic is done in integer cents, so sums never drift. When a split leaves a remainder, the leftover cents go to named people by a fixed rule, and the result is the same every time. Currencies that have no coins below one, like the dram, are settled in whole units so nobody is asked to pay 0.99 of something. The settlement procedure finds the fewest transfers. Every expense keeps a history of who changed what. Those are the problems a shared ledger actually has, and they are solved with rules you can read, not with a model you have to trust.

And the honest part

The app was built with an AI coding tool. I described the screens and the rules, it wrote the code, and I checked the result on my phone until it was right. That’s a tool used by the person making the thing. It’s different from a model sitting inside the thing, making decisions about your money with your friends. I wrote about that distinction in built with AI tools, shipped with no AI features.

So in a year when every app page promises intelligence, kisir’s promise is narrower. It divides correctly, it adds up, it shows its work, and your group can check all of it. If you’re choosing between apps, the best apps to split expenses guide compares them honestly, including the ones with AI features that some people genuinely want.