How to convert PDF to Excel using AI without losing numbers

First, determine the PDF type

Try selecting a row with your mouse and copying it into plain text. If you get readable characters, start with text extraction. If you see only an image or gibberish, OCR is needed. Separately, check if the table continues on another page, if headers are carried over, and if there are summary rows.

via the OpenAI documentation, ChatGPT supports PDF and tabular formats, but accurate extraction from scans and complex images can be unreliable. Availability depends on your account and plan. This guide does not promise error-free recognition of any document.

Prepare a quick check before uploading

On a copy of the document, mark the number of items, currency, total sum, and two complex cells. Do not upload someone else's personal or commercial data without proper grounds; our fictional example below is sufficient for training.

In the training account, there are three items: notebook — 3 × 89.50 UAH; pen — 10 × 12.00 UAH; delivery — 1 × 65.00 UAH. Expected total: 453.50 UAH. Delivery is a separate item, not a note. This is intentionally a small task so that the result can be verified manually.

Upload source text · CSV for verification. Text can be saved as a PDF via print in your editor; this is training text, not a scan.

Prompt: extraction separate from guesswork

Upload a copy of the PDF if your tool supports files. Or paste the selected text. First, ask it to show the structure and all doubtful areas; then, to compile the table. Do not start by asking to fix or 'improve' the data: it makes it harder to understand what was in the original.

Convert the provided document into a table. Do not silently supplement or correct the source. 1. List the pages/fragments successfully read and the number of items. 2. Fields: item; quantity; unit; price_UAH; sum_UAH; page; doubt. 3. Leave empty cells empty. Mark ambiguous ones as "CHECK" and provide the source fragment. 4. Show a table for verification, and separately — the calculated total and the total in the source. 5. Only after my confirmation, prepare a CSV: UTF-8, separator ;, decimal comma. Mark text cells starting with = + - @ for safe manual import; do not create formulas from source text.

What to verify before opening in Excel

Compare the number of items and units of measurement. Check every price with a decimal comma, negative sums, discounts, and empty values. For a complex document, add "page" and "source fragment" columns: this way, an error can be found without re-reading the entire PDF.

In our example, the calculations are: 3 × 89.50 = 268.50; 10 × 12.00 = 120.00; 1 × 65.00 = 65.00; total 453.50. Compare the sum obtained from the rows with the printed total separately. If they differ, stop and find the reason instead of forcing the numbers to match.

If the CSV opens in a single column, use text/CSV import and explicitly select the ';' delimiter and the correct number locale. In our file, a comma is used for decimals and a semicolon between fields. Check that prices have become numbers, not text.

When a different path is better

For regular invoices of the same format, it is advisable to set up extraction and control rules once and leave the AI for ambiguous parts. For a single short document, manual entry with verification is sometimes faster than debugging a complex process. The goal is a correct table, not the maximum number of AI steps.

Do not call the result verified just because the model wrote "everything is correct." Your own verification here means re-calculating the sum and checking the rows. If the PDF contains formulas or legally significant data, keep the original and pass the result to a responsible specialist.

Next task: convert verified facts into a presentation. And to choose a service, start with testing on your own tasks.

Sources and method

• OpenAI — Data analysis with ChatGPT, verified 23.10.2026

Facts about the services have been verified against official materials. Examples, prompts, and the workflow were prepared by the editorial team. This is not a comparative test of model accuracy or speed. How we work.