When a statement is too long for an assistant to handle in one go, split it by month, keep the column headers at the top of every piece, carry your category rules into each chat, and count the rows in each piece before and after. Then combine the results yourself. Splitting is easy; what protects you from lost rows is the count.
Why does a long statement cause problems?
Assistants can only read a limited amount of text at once, often called the context limit. If a file is bigger, the tool may cut off the end, skim the middle or quietly summarise instead of processing every row. You then get an answer that looks complete but covers only part of the year. The risk grows with a CSV of several thousand rows.
How do I split the file safely?
- Start with a CSV or Excel file if you can, as it is easier to split than a PDF.
- Sort by date so nothing sits out of order.
- Cut the file by calendar month, or by a fixed number of rows such as 200, whichever suits the tool.
- Copy the header row to the top of every piece, so each chunk is self-explanatory.
- Name the pieces clearly, such as April, May and June, and keep the original file untouched.
- Record the row count of each piece in a note before you upload it.
Splitting at a month boundary is neater than cutting at a row count, because each chunk matches a statement period, and you can reconcile each one against its own opening and closing balance.
How do I check that no rows were lost?
Make the assistant report back the number of rows it processed for each chunk, and compare it with your note. Say April has 142 rows, May 156 and June 131. The total should be 142 + 156 + 131 = 429. If the assistant says it handled 138 rows for April, four are missing, and you ask it to list the dates and amounts it saw so you can find which.
| Chunk | Rows in file | Rows reported by assistant | Difference |
|---|---|---|---|
| April | 142 | 142 | 0 |
| May | 156 | 151 | 5 missing |
| June | 131 | 131 | 0 |
| Total | 429 | 424 | 5 missing |
The numbers above are invented to show the method. A mismatch is the signal to re-run that chunk, not to carry on.
If you are working from a PDF, splitting is harder, because pages do not always break at a month end. In that case ask the assistant to process a fixed number of pages at a time, say five, and note which page each chunk starts and ends on. Overlap nothing, and write down the date of the last row in each chunk so the next one starts on the following row.
How do I keep categories consistent across chunks?
Each new chat starts with no memory of earlier ones, so repeat the same context every time. Paste your fixed category list and your mapping table of merchants and categories at the start of each chunk. After each month, add newly resolved merchants to the table, so later chunks inherit the earlier decisions.
How do I merge the results?
Ask for each chunk's output in the same column layout, then paste them one under another into a spreadsheet. Remove any repeated header rows, sort by date, and total by category yourself. Check that the number of rows equals your original count, and that opening balance plus all credits minus all debits equals the final closing balance.
If splitting and merging by hand starts to feel like the main job, a tracker that takes the whole file in one go is simpler. SpendixAI accepts statements as PDF, CSV or Excel, supports batch uploads and shows import history with status, so you can review the rows before they are saved.
Before you start, a short list
- Original file saved and left alone.
- Header row copied to every chunk.
- Row counts noted for each chunk.
- Same category list and mapping table pasted into each chat.
- Merged total compared with the original.
Useful ideas start with a clearer picture of your own spending.
Explore your reports


