Why not just paste the statement into ChatGPT?
A fair question, and it deserves a straight answer — starting with the part we won’t hide: we run on a large language model too. So this isn’t about the model. It’s about what a chat window can’t give you around it.
5 free conversions, no signup, no credit card.
We’re not selling you a smarter model. We run on one too.
Our extraction runs on a large language model — Google’s Gemini in production. We don’t run on GPT specifically, so this isn’t the old trick of secretly reselling the same thing you could use yourself. But it is the same class of tool, and that’s the honest point: a page telling you our model beats ChatGPT would be selling you something that isn’t the real difference.
The real difference isn’t model versus model. It’s a chat window — a capable model with nothing around it — versus the same class of model wrapped in a pipeline that checks its work, hands you a file, and does it the same way every time. ChatGPT reads a statement well. What it doesn’t do, on its own, is tell you when it got a number wrong.
What the pipeline adds around the model
A check the model cannot run on itself
Ask ChatGPT to read a statement and it returns numbers that look right, with no way to know whether they are. Our pipeline re-derives the statement’s running balance after extraction — balance plus the previous amount must equal the previous balance — and flags any row that fails as UNVERIFIED, in the filename and inside the file. This is deterministic arithmetic, not another model opinion, so it catches the model’s own mistakes. A chat has nothing equivalent, whichever model is behind it.
A file, not text to re-key
A chat gives you an answer to read. You still have to copy it into a spreadsheet — and hand-copying is exactly the transcription step you were trying to remove. We hand you a CSV or Excel file (JSON, OFX and a Fortnox-style CSV on Pro), columns already in place.
A fixed pipeline, not a conversation
A chat is conversational: reword the prompt and the result can shift, and a very long statement can get abbreviated in a long reply when you needed every row. The conversion here is a fixed pipeline — no prompt to engineer and nothing to phrase — and our production engine runs extraction at temperature zero, pinned to its most consistent output rather than a chattier one. Everything after extraction is fully deterministic: the same rows always reconcile and export to the same file.
Whole statements, scans included
A long statement can overflow a comfortable single paste. We split the text into whole lines so no transaction is ever cut across a boundary, and scanned or photographed statements go through the same reconciliation check — held to a stricter bar than digital PDFs, since a scan’s text layer can hide dropped rows.
A defined path for the data
A statement here is processed in memory, never written to disk, and dropped within minutes; the whole path — API, database and the AI step — runs in the EU. Pasting a statement into a general-purpose chat is a different thing: a consumer product handling your client’s financial data under whatever terms that product sets. We are not going to characterise those terms for you — only point out that they are not the same as a pipeline built for this.
Side by side
| ChatGPT chat | Bank Statement Sheet | |
|---|---|---|
| Reads a statement | Yes, and does it well | Yes — on a model of the same class |
| Checks the numbers are right | No built-in check | Reconciles every row against the statement’s balances |
| Output | Text in a message, to copy out | CSV / Excel file (JSON / OFX / Fortnox on Pro) |
| Repeatable | Varies with the prompt | Fixed pipeline, no prompt; temperature-zero extraction |
| Long statements | Can truncate or summarise | Chunked so no row is cut |
| Flags what it couldn’t verify | No | UNVERIFIED in the filename and in the file |
| Data path | A general-purpose chat | In memory, EU, dropped within minutes |
The point of the first column isn’t that a chat is bad at this — it’s good at reading. It’s that reading is the one step both sides share, and everything else in this table is the part a chat window leaves to you.
Questions
Hang on — don’t you use AI yourself?
We do, and we would rather say it plainly than pretend otherwise. Our extraction runs on a large language model — Google’s Gemini in our production setup. We do not run on GPT specifically, so this is not a case of “we secretly use the same thing”. But it is the same class of tool, which is the point: the model is not where the difference lives. What differs is everything built around it — a verification step the model can’t perform on itself, a structured file, a fixed process, and a defined data path.
So are you more accurate than ChatGPT at reading a statement?
Not a claim we make. We use a frontier model of the same class, so at the raw reading step the quality is comparable — that is the honest position. Our accuracy claim is narrower and more useful: every extracted row is checked against the statement’s own balances, so when the model does slip, you are told, rather than importing a wrong figure that looks fine. On our reference statement, a real 743-transaction Handelsbanken PDF, all 742 balance checks pass. That is one statement’s result, not an accuracy rate for every bank.
Honestly, when should I just use ChatGPT?
For a single statement you are eyeballing, where you can sanity-check the total yourself, ChatGPT may be all you need — and we would rather tell you that than oversell. This tool earns its place when the numbers have to be right without you re-checking each one, when you want a file rather than text to copy, when the statement is long, or when it is a client’s data and you would rather it went through a defined path than a paste into a general chat.
What file formats do I get?
Excel (.xlsx) and CSV on the free tier. JSON, OFX and a Fortnox-style CSV are Pro. A chat gives you text in a message; this gives you the file.
Does it connect to my accounting software?
No — there is no integration and no bank feed. You get a file and import it yourself into QuickBooks, Xero or whatever you use. Nothing we run touches your accounts.
Try it on a real statement
Five free conversions, no signup. Check the reconciliation preview before you download anything — the part a chat can’t show you.
Comparing with Claude instead? Bank Statement Sheet vs Claude