PDF to CSV with AI: check rows, values and missing data
Prepared with AI assistance, identified sources and original fictional examples. This is not a comparison test of AI providers. Editorial method and corrections
A table converted to CSV can open normally and still contain an extraction error. A valid file means its structure can be read; it does not prove that the values match the PDF. This workflow separates reading, organization and checking so you can locate where a conversion went wrong.
Define the columns before extraction
Write down each column name and the data type you expect. An item code is text, even when it contains only digits. Quantity is an integer when the document says so. Money needs a clear decimal-separator convention. A blank observation means information is missing; it is not an instruction to fill in the field.
For this exercise, use the columns code, description, quantity, unit price and note. The document is entirely fictional: it is not a purchase, supplier quote or result from an AI service test. Download the reference CSV to compare your output.
| Code | Description | Quantity | Unit price | Note |
|---|---|---|---|---|
| 0018 | Notebook, blue cover | 10 | 12,50 | Under review |
| 0021 | A4 folder | 5 | 8,00 | Not provided |
| 0100 | Marker | 2 | 7,25 | Partial delivery |
Ask for fidelity and a report of uncertainties
Extract the table without filling in missing data. Keep codes as text and preserve decimal commas. Use a semicolon-delimited CSV. Do not mix comments into the CSV itself. After the table, list cells that could not be read, omitted rows and any interpretation you had to make. Do not calculate values unless asked.
A semicolon reduces visual confusion with decimal commas in this exercise, but the importer must also be set to that delimiter. Do not try to fix compatibility by replacing every comma with a period; that can alter descriptions and quoted fields.
Check four things before importing
Rows: are there three items in addition to the header? Was a footer mistakenly turned into a product?
Identifiers: did 0018, 0021 and 0100 keep their leading zeros?
Fields: is “Blue-cover notebook” still in one cell, even though the description contains punctuation?
Gaps: did a missing note stay blank instead of becoming a conclusion about delivery or approval?
The CSV validator checks quotation marks and consistent field counts. Select semicolon for this sample. Then compare the values with the PDF: the validator cannot see the original table. A row with a wrong value can still be structurally valid.
Reconcile totals independently
For this exercise, the subtotals are 10 × 12.50 = 125.00; 5 × 8.00 = 40.00; and 2 × 7.25 = 14.50. The sum is 179.50. These numbers are the answer key for the example, not a result from an external AI. If your spreadsheet shows another total, first check the decimal separator, cell type, row count and formulas.
If the real document contains its own total, compare it with your calculation. A difference can come from tax, a discount or rounding as well as a reading error. Do not change an arbitrary row just to make the totals match. Record the difference and return to the source passage that explains it.
JSON can make data types easier to review
The CSV-to-JSON converter keeps fields as strings, which can help preserve codes with leading zeros. The application that reads the JSON should decide which values to convert to numbers. Document that “12,50” means twelve and fifty hundredths in the notation used here instead of relying on automatic conversion.
When exporting JSON back to CSV, keep protection against formula-like text in mind if a field begins with characters a spreadsheet may interpret. Protection may add an apostrophe to an exported value; read the tool’s explanation and check how your importer handles it. Representation fidelity and safety both matter.
Keep a simple audit trail
Keep the original PDF, the extracted CSV and a note of corrections. Record who checked the data and which pages were included. For a large file, start with a few rows of different kinds before processing everything. A workflow that works for a simple table can fail on another with merged cells, multiple headers or scattered notes.
Sources and method
Sources checked on October 9, 2026. The procedures and exercises are original ConverteArq material; the documentation below supports specific observations about the technology cited.