AI expense tracker: what actually works

Quick answer

AI solves three things well in expense tracking today: turning plain language into a structured entry, pulling the figures out of a photographed receipt, and answering questions about a history that is already in order. It still does not solve the part people believe it solves: it does not remove the need for data to get in, and an assistant that only reads what you paste is not keeping your books, it is commenting on the books you keep.

The distinction that decides whether this is useful to you

Almost everything marketed as "AI expense tracking" falls into one of two very different categories, and the advertising presents them identically.

In the first, the assistant talks about your spending. You paste a statement, a list or a summary, and it answers with analysis. It is useful for a while. The problem is that the boring work — keeping expenses logged, current and correctly categorised — remains entirely yours, and that work is ninety per cent of the job. A brilliant analysis of a month you have not finished logging is a wrong analysis, beautifully written.

In the second, the assistant operates the application. It logs, corrects, queries and settles inside the system where your data lives. Here the AI is not hovering above the problem describing it: it is inside, doing the part you do not feel like doing.

The technical difference between the two is whether a real integration exists — today, typically an MCP server — or whether there is only a chat window. It is worth learning to tell them apart before choosing a tool, because from the outside they look very similar.

What AI genuinely does well

Worth being concrete rather than spreading enthusiasm, so: three things that work today.

Turning a sentence into an entry. "Dinner with the flatmates, sixty-two euros, I paid, split four ways" contains everything a structured record needs: amount, description, payer, split. Current models handle this well, and it is where the most friction disappears — because the friction of logging is exactly why people stop logging in week three.

Reading a receipt. Extracting amount, date and merchant from a photo is a solved task. It is still worth checking what it understood, especially on long receipts, but the saving against typing it out is real.

Answering questions about an orderly history. "How much have I spent on transport this month?", "is that more than usual?". Mind the word order: about an orderly history. AI does not repair incomplete data; it interprets it with equal confidence whether it is complete or not.

What it does not do, however it is advertised

It does not make data appear. If nobody logs anything, there is nothing to analyse. This sounds obvious and it is why most attempts at keeping accounts fail: not for want of analysis, but for want of consistency. The only thing that attacks that at the root is driving the cost of logging down to nearly zero — which is precisely what saying it to an agent in one sentence does.

It does not magically connect to your bank. Bank aggregation is a regulated, separate problem with its own permissions and providers. An assistant connected to an expenses application does not see your current account.

It does not decide how you split. Whether a group pays equally, in proportion to income, or by category rules is a decision for the people involved, not the model. What the tool can do is apply the rule you agreed without rounding errors, which is where the genuinely silly arguments come from.

What this looks like when it works

An ordinary example: a couple sharing a flat on different salaries.

The split rule is agreed once — say, proportional to what each earns — and configured in the application. From then on, logging stops being a form: it is a sentence to the agent as you leave the till. At month end, instead of opening a spreadsheet, the question is "how are we doing?" and the answer comes back with real balances and who owes what to whom.

What changed there is not the analysis, which was always possible. What changed is that the record stays current with no effort, and that is why the analysis is correct.

What to check before picking a tool

Four questions that separate the wheat from the chaff.

Can your assistant write to the application, or only read what you paste? With no integration, it is a commentator.

What permissions does the integration grant, and can you narrow them? All-or-nothing access to your finances is a bad idea even when the tool is good.

Can access be revoked without dismantling the account? The answer should be a button.

What happens to your data? An agent operating your application should not mean your transactions feed anybody's training. It is a fair question and it deserves a clear answer in whoever's documentation.

Frequently asked questions

Can AI keep my accounts on its own?

It can handle the logging and the arithmetic if you give it access to an application where your data lives. What it cannot do is guess expenses nobody told it about, or decide the split rules for you.

Are ChatGPT or Claude any use for tracking spending?

They are useful for analysing whatever you paste, and that has a low ceiling. Connect them over MCP to an expenses application and they can log and query directly, which is where the saving in effort actually shows up.

Do I need to be technical?

Connecting an agent means pasting a URL and a token, about the same effort as setting up a new app on your phone. Using it afterwards takes nothing: you talk to it.

Can it read shopping receipts?

Yes — pulling amount, date and merchant from a photo is something current models handle well. Check the result on long receipts or poor handwriting.

Is it reliable for shared accounts?

The arithmetic is done by the application, not the model, so the split is deterministic and does not depend on the AI getting it right. Where it is worth checking is data entry: that what it understood from your sentence is what you meant.