How to Verify AI PDF Summaries: Numbers and Sources
A fluent AI summary does not always mean it read the tables correctly, particularly with scanned PDFs, multi-year figures, and table footnotes. This article proposes checking key claims first, instead of rereading the entire answer.

Distinguish misread files, incomplete extraction, and summaries exceeding evidence
Check three stages of PDF summaries: what the system can read, what it selects, and how it concludes. Misread figures can underpin plausible but wrong explanations. Even correctly read numbers can mislead if notes are omitted.
Do not assume every service reads PDFs the same way. Some extract text layers, some use page images, and some combine methods. Gemini API documentation explains reading both text and visual elements in PDFs and recommends upright pages without blur. This API capability does not prove that all chatbots or plans read files equally.
To isolate problems, first ask AI to extract text or numbers from a specified location, then summarize. If extraction is wrong, improve the file or select clearer pages. If extraction is right but the summary is wrong, refine the task and review reasoning. Changing instructions alone cannot repair an unreadable original.
Prepare sources before sending them to AI
- Keep the original file: Work on copies, specifying the document name, version, date, and number of pages, especially for manuals that have been edited multiple times.
- Select a sample passage: If text cannot be selected, it may be scanned. If copied text is reordered, compare it to the page image. Selectable text does not guarantee complete table extraction.
- Verify important pages: Rotate tilted pages and find clearer images. Avoid compression that removes decimal points or minus signs.
- Match page numbers: The 12th PDF page may have printed page number 9. Specify both when necessary.
- Check permissions and personal information: Delete unnecessary data before sending it to external services. and read the policy of that service
Selecting relevant pages reduces distractions, but retain table-header pages, definitions, and explanatory notes. If using KrobClick PDF Editor to prepare files, check the guide and supported scope. This tool manages PDFs in-browser; it is not an AI document summarizer. Uploading to an AI service is a separate step.
If you need to solve the problem of searching or selecting text from a scanned image, read this. OCR Guide for PDF Scans first. This article focuses on checking AI summary claims after document reading, rather than performing OCR.
Hypothetical example: correct numbers, wrong units
The following table contains hypothetical data created by KrobClick to illustrate verification, not a real company report or model test. Suppose its heading reads “Sales, unit: thousand baht,” with these values: Suppose the table heading says “Sales, unit: thousand baht,” with the following data
| items | 2025 (B.E. 2568) | 2026 (B.E. 2569) |
|---|---|---|
| Total sales in thousand baht | 1,250 | 1,375 |
| Note | excluding VAT and using the same business scope in both years | |
Value2026 (B.E. 2569) equals 1,375 × 1,000 = 1,375,000 baht The rate of increase is (1,375 - 1,250) ÷ 1,250 × 100 = 10% The two annual figures suffice to calculate the change, but not to conclude “profit increased 10%,” because costs are missing. Do not describe amounts as tax-inclusive when the note says otherwise.
“Sales of 1,375 baht, profit up 10%” contains two distinct errors here: wrong units and the wrong metric. Matching 1,375 alone is insufficient; verify item labels, units, periods, and definitions together.
Points that should be additionally checked when documents are in Thai
- Year and period: B.E. 2569 corresponds to C.E. 2026 by subtracting 543. Check calendar versus fiscal years; converting the year does not automatically align reporting periods.
- Quantity and rate: A rise from 5% to 6% is 1 percentage point or 20% relative to the original value: two different descriptions
- Symbol: Parentheses may indicate negative values. A dash may mean no data or zero, depending on the document's notes. Do not assign a meaning yourself.
- Abbreviations and units: Baht, thousand baht, million baht, people, and times must match the row or table header. Do not borrow units from the previous table.
- Exceptions: “Excluding,” “only,” “estimated,” and footnotes often change an answer's scope
This is an editorial checklist, not a ranking of the most common AI errors. If a blank cell or symbol is unclear, state the uncertainty and ask the document owner rather than filling the gap by inference.
Instructions making answers traceable
Use only attached documents to extract “Total sales,” separating item, value, unit, period, and notes. Identify PDF pages and printed page numbers if present. Before summarizing, show short supporting text for every point. Write “not found” where absent. Separate source data from calculations and do not infer profit from sales.
Open each cited page, starting with figures that affect decisions most. Recalculate them and read the surrounding lines. A link or page number is a starting point for verification, not proof of accuracy. Asking the same AI “Are you sure?” uses the same source and mechanism rather than an independent check.
With a long document, if the system selects only portions, context may be lost before drafting begins. See the explanation of truncation and retrieval in Contextual Retrieval of Anthropic and read more in RAG articles to understand why having documents does not guarantee the correct passages are used
Check the evidence before sharing the summary
Before sharing the summary, identify the source page for each key conclusion, check units and periods, repeat calculations, and list missing information. Have a qualified reviewer check financial, legal, or health content before decisions. Put unresolved points in the summary rather than leaving them in chat history.
When the evidence is insufficient, identify what cannot yet be concluded, even if that makes the answer shorter. Use AI to help find information, then check the original document for support for each conclusion.
Sources
Sources reviewed October 14, 2026, Thailand time. Thai examples and prompts are KrobClick proposals, not product tests. Features and settings vary by version/account. Check originals and outputs before real use.