Can AI Read PDF and Excel Accurately? Recent Studies
A converted bank statement or AI-generated summary can look convincing while still containing missing rows, misplaced values or incorrect calculations. Research helps explain why a polished result is not enough.
Try AI Ready to inspect text and prepare a local analysis pack.
Recent studies examine different stages of document processing: table structure, Arabic text recognition and Excel formula repair. This article reviews three studies from 2025 and 2026 and separates their findings from our practical recommendations.
These studies did not evaluate HugPDF. Their findings apply to the datasets, tools and versions assessed by the researchers, not to every current service or user document.
1. August 2026: difficult tables expose hidden weaknesses
The TableParseMap paper, submitted to arXiv on August 10, 2026, introduces a diagnostic benchmark containing 916 real-world tables. It groups challenging scenarios and failure types rather than reporting only an aggregate score.
The strongest parser in the evaluation achieved 85.03 TEDS. TEDS measures similarity between the reconstructed table and a reference; it is not the percentage of files containing no errors. The researchers discuss large tables, ambiguous visual evidence and inconsistency between extracted tables and their source images. Read the original paper.
Our practical recommendation: after using PDF to Excel, compare headings, columns and the first and last rows with the original. Correct words can still be assigned to the wrong column. Inspect headings spanning multiple columns and tables continuing across pages.
2. A 2025 study: Arabic documents need dedicated evaluation
KITAB-Bench evaluates Arabic OCR and document understanding. The paper was submitted on February 20, 2025 and revised on June 27, 2025. Its benchmark contains 8,809 samples across nine domains, including tables, handwriting and charts.
The study discusses connected Arabic letters, right-to-left reading, typography, numerals and table structure. It reports progress for evaluated models alongside remaining limitations in document conversion. The results should not be treated as a ranking of every tool available today. Read KITAB-Bench.
Our practical recommendation: after OCR PDF, compare names, amounts and identifiers with the source image. For Arabic tables, check that each name remains associated with the correct amount, particularly when Arabic, English and numerals appear together.
3. A 2025 study: explaining an Excel error differs from fixing it
Benchmark Dataset Generation and Evaluation for Excel Formula Repair with LLMs was submitted on August 14, 2025. Its research page states acceptance at the KDD workshop on evaluation and trustworthiness of agentic and generative AI models.
The researchers constructed 618 evaluation samples for Excel formula repair. Their approach uses spreadsheet context and evaluates repairs through execution, alongside other validation procedures. Read the formula-repair study.
Our practical recommendation: test a proposed formula in Excel using known cases. Check its cell range rather than accepting a fluent explanation. A formula can execute without displaying an error while still including the wrong rows.
A simple example: double-counting a total
This is our educational example, not an experiment reported in the studies.
| Row | Amount in SAR | Row type |
|---|---|---|
| 2 | 100 | Invoice |
| 3 | 200 | Invoice |
| 4 | 50 | Invoice |
| 5 | 350 | Invoice total |
The formula =SUM(B2:B5) returns 700 because it counts the total again. =SUM(B2:B4) returns 350, the sum of the three invoices in this example. Ask which records entered a calculation and compare it with an independent result. Check the actual ranges in your workbook rather than copying this example blindly.
Checks before relying on a conversion or AI analysis
- Keep the original. Make changes to a copy you can compare with the source.
- Inspect text. Copy a sentence and a number. For scans, review OCR output before proceeding.
- Check table structure. Compare rows, columns and merged cells, especially across page breaks.
- Check data types. Amounts stored as text may need conversion; codes such as 0012 should retain their leading zeros.
- Verify calculations. Exclude total rows and duplicates, checking currencies, separators and dates.
- Request source references. Ask for page or row numbers and verify them yourself.
Use Extract PDF Pages to isolate relevant material while retaining necessary context. For structured prompts, read Prepare PDF and Excel Files for AI.
Accuracy and privacy are separate questions
An accurate result does not establish that sharing the file meets your workplace policy. Local processing does not eliminate the need to check output quality.
HugPDF prepares and processes documents locally in the browser. Sending the resulting file to a cloud AI service is a separate action governed by that service's policies. Read the Privacy Policy to distinguish document processing from site connections. Local processing alone is neither a compliance certification nor a guarantee of accuracy.
Original research sources
- TableParseMap: From Diagnosis to Correction — submitted August 10, 2026. An arXiv research paper; availability on arXiv alone does not establish peer review.
- KITAB-Bench — submitted February 20, 2025; revised June 27, 2025. The research page lists ACL 2025.
- Excel Formula Repair with LLMs — submitted August 14, 2025. The research page states acceptance at the KDD evaluation and trustworthiness workshop.
Frequently asked questions
Do errors make AI useless for documents?
No. AI can help summarize, organize and suggest formulas, with checks on important outputs. Results depend on the service, input quality and task.
Does a high benchmark score guarantee my file is correct?
No. A score describes a specific evaluation. Metrics such as TEDS do not directly represent the proportion of entirely error-free files.
Why can Arabic tables be difficult to extract?
Text and numeral recognition can interact with reading direction and table structure. Compare extracted text and cell placement with the original to identify the problem.
Did these studies test HugPDF?
No. They help explain general document-processing challenges. We do not attribute their scores to HugPDF or use them to claim superiority over another tool.
What should I check first after converting PDF to Excel?
Check headings, row and column counts, then identifiers, amounts and total rows. Test calculations on a known sample before relying on the complete workbook.