Working with tables, figures and numbers in PDFs using AI
Updated 13 October 2026 · ChatPDF guides · 6 sources
AI can miss or misstate things. Summaries and answers from AI tools, ChatPDF included, can leave out important details or get them wrong. Check anything that matters against the document itself, especially legal, medical, financial and exam material.
- Text extraction keeps the characters of a table but often loses which number belongs to which row, column and unit. 1
- Language models handle tables less reliably than prose, and results change with how the table is presented to them. 2
- Charts and figures drawn as graphics carry no text, so a text-only tool cannot read the values in them.
- Ask for one number at a time with its label and unit, then check each one on the page.
Numbers are where AI answers about PDFs most often go wrong, and where mistakes cost the most: a wrong total in a budget, a misread dosage, the 2024 figure reported as 2025. The reasons are mostly mechanical and predictable, which means you can work around them.
Why tables break
The grid disappears during extraction
A PDF does not store a table as rows and columns. It stores characters and their positions on the page, and any software reading it has to reconstruct words, lines and reading order from those positions 1. For a table, that reconstruction often reads across or down in the wrong order, merges cells, or splits a cell that wraps onto two lines.
The numbers survive, but the links between them do not. When the AI model later reads that text, it has to guess which value belongs to which header.
Models are uneven with tables
Even when a table reaches a model intact, understanding it is a separate skill. A benchmark of table tasks such as looking up a cell or retrieving a row found that model performance changed with the table's input format and content order 2. On a benchmark of questions about company financial filings, one leading model paired with a retrieval system gave a wrong answer or no answer for 81% of a 150-question sample 3. Newer models do better, but the same failure modes remain.
Retrieval can split a table from its labels
Chat-with-PDF tools usually split documents into passages and pick a few for each question 4. A long table can be cut in two, leaving the numbers in one passage and the column headers or the "in thousands" note in another (see how chat-with-PDF tools work).
Charts are pictures
Bar charts, line graphs and diagrams are usually drawn as graphics. A tool that reads only the PDF's text sees the caption and axis labels at most, not the values. If the numbers matter, look for the table or text the chart is based on.
How to ask about numbers
- Ask for one value at a time, with its labels: "What was the South region's revenue in 2025, and in what unit?"
- Ask for the row, not a calculation: "List every value in the row labeled Operating costs, with its column header."
- Do the arithmetic yourself. Use a calculator for totals, differences and percentages instead of asking the AI to compute them.
- Ask about units and footnotes: "Does the table say whether these figures are in thousands or millions?"
- Name the table if it has a title or number: "In Table 3, what is the value for..."
Verify every number you use
- Search the PDF for the number itself. In the Preview app on iPhone, tap the search button and enter the number or a nearby label 5. See searching a PDF on iPhone.
- Read it in place on the page: check the row, the column header, the year and the unit.
- Check footnotes and notes below the table for restatements, exclusions or different units.
- Copy exact values rather than retyping them. Preview lets you select and copy text from a PDF 6, which avoids transcription slips.
If the search finds nothing, the table may be a picture of a table rather than text. Then neither search nor a text-only AI tool can read it; see scanned PDFs and OCR.
When to skip the AI
For financial statements, lab results, dosages, grades, invoices and anything with legal or money consequences, read the table yourself and use the AI only to find where it is ("Which section has the fee schedule?"). The general limits of AI summaries are covered in can you trust AI PDF summaries?
Where ChatPDF helps, and its limits
ChatPDF can help you locate a table, explain what its columns mean, or pull out a single labeled value to check. It works from the PDF's extracted text, so table layout can be lost, chart values are not read, and its answers do not show page numbers, so every figure still needs checking on the page.
Common questions
Why did the AI give me the right number for the wrong year?
Extraction often loses which column a value came from, so the model has to guess. Ask for the value together with its column header and check it on the page. 1,2
Can AI read values from charts in a PDF?
Not if the tool only reads the PDF’s text. Charts are usually drawn as graphics, so only captions and labels are available. Use the underlying table or read the chart yourself.
Should I let the AI do the math?
It is safer to ask for the individual values, verify them, and calculate totals or percentages yourself.
Sources and further reading
- Bast H, Korzen C (2017). A benchmark and evaluation for text extraction from PDF. ACM/IEEE Joint Conference on Digital Libraries (JCDL) 2017. doi:10.1109/JCDL.2017.7991564 ad-publications.cs.uni-freiburg.de
- Sui Y, Zhou M, Zhou M, Han S, Zhang D (2024). Table meets LLM: can large language models understand structured table data? A benchmark and empirical study. Proceedings of WSDM 2024. doi:10.1145/3616855.3635752 (arXiv:2305.13062). arxiv.org
- Islam P, Kannappan A, Kiela D, et al. (2023). FinanceBench: a new benchmark for financial question answering. arXiv:2311.11944. arxiv.org
- Anthropic (2024). Introducing Contextual Retrieval. anthropic.com
- Apple Support. Open and view PDFs and images in Preview on iPhone (iPhone User Guide). support.apple.com
- Apple Support. Annotate a PDF or image in Preview on iPhone (iPhone User Guide). support.apple.com