We Processed 500+ Documents — Here's What Breaks When You Convert PDF to Excel

We built CleanTably to convert any document to Excel. After running 500+ real documents through our AI pipeline (invoices, receipts, bank statements, handwritten forms, payroll sheets, and plenty of ugly scans), we now know where conversions work, where they break, and why.

Most articles about PDF to Excel conversion are written by people who haven't actually run hundreds of documents through these tools. This one isn't. The data below comes from our production logs, manual accuracy reviews, and the failure cases that landed in our debugging queue. We're publishing it because honest benchmarks are hard to find, and because we think the gaps in current tools deserve a straight answer.

Key Findings at a Glance

The 100% pipeline success rate means the system never crashes or hands you an empty file. Every document produces some output. But "produces output" and "produces accurate output" are two different things. The ~89% figure reflects the percentage of data fields extracted correctly on the first pass, no manual correction needed.

Accuracy by Document Type

Accuracy varies a lot depending on what you're converting. A clean printed invoice is a different challenge from a handwritten form on lined paper. Here's the breakdown from our test corpus:

That's roughly a 25-point gap between the best and worst categories. Invoices and typed forms? Genuinely production-ready. Complex multi-column layouts and very long documents? A solid first draft, but you'll want to review it. Handwriting? A starting point. Expect to correct things.

The 5 Most Common Conversion Errors

Five failure patterns show up again and again. If you know what they are, you can predict where to double-check your output:

1. Column mismatch in multi-column layouts Think payroll sheets: Income on the left, Deductions on the right, each with their own headers and rows. You glance at it and the structure is obvious. But the AI reads the document as a single grid, and the most common failure is rows from the left section getting merged with the right, producing records that mix unrelated data. We almost never see this on single-table invoices. It's a layout problem, not an intelligence problem.
2. Handwriting character confusion Handwritten text fails at specific, predictable points. The characters we see confused most often: 1 read as 7 (and vice versa), 6 as 8, lowercase a as o, cursive l as e. Each mistake is small on its own. But they compound fast in financial figures. A handwritten line item of $1,617.00 becomes $7,671.00 if just two characters get misread. If your source is handwritten, verify the numbers.
3. Silent data truncation on very long documents AI models have input and output size limits. When a document exceeds them, we found the model just stops producing output mid-document. No error message. The spreadsheet you get back looks complete. It isn't. The last several pages of data simply aren't there. We now flag this with a notice sheet in the output Excel file, but most tools don't. This is the most dangerous error on the list because you won't notice unless you check page counts. See the technical explanation of LLM context limits if you want the underlying mechanics.
4. Number format inconsistency Financial documents use number formats inconsistently, sometimes within the same page. A single invoice might have "$1,234.56" in the line items, "1234.56" in the subtotal, and "$23/hour" in the rate column. When those land in the same Excel column, some cells are numbers and some are text. Your SUM formula breaks silently. We handle normalization for common cases, but edge cases (mixed units, parenthetical negatives, embedded currencies) still show up in about 40% of financial documents. Even human data entry struggles with this one. See Excel number formats for how deep the rabbit hole goes.
5. LLM summarization instead of verbatim extraction This one surprised me. Language models are trained to be helpful, and sometimes "helpful" means paraphrasing instead of copying exactly. In rare cases, a document with long text fields (descriptions, notes, comments) comes back with slightly reworded content. Consulting invoices with detailed project descriptions and legal forms with explanatory paragraphs are the usual culprits. The meaning is preserved, but the exact wording differs. That matters if you need the output for compliance, auditing, or legal review.

What We Do About It

Once we mapped these failure modes, we built mitigations into the pipeline:

• Truncation warning: When a document hits the page cap (20 pages on the free tier), a notice sheet appears as the first tab in your Excel file telling you how many pages were processed and how many were skipped.
• Number normalization: The pipeline strips currency symbols and normalizes commas and decimals so values land in Excel as actual numbers, not text. Edge cases remain. We recommend a quick column-type check after import.
• Multi-sheet output: Documents with multiple distinct table sections get split across sheets, which reduces the column mismatch problem for side-by-side layouts.
• Handwriting flag: When the AI detects handwritten content, the output includes a note recommending you review numerical fields manually.

None of these fully eliminate the errors. They just make them visible so you can fix them faster.

Quotable Statistics

From our production data and test corpus. Cite with attribution to CleanTably:

Methodology

Transparency matters, so here's exactly where these numbers come from:

Test corpus (15 documents): A manually curated set covering all seven document types in the table above. We converted each one through the production pipeline and reviewed it field by field against the source. Accuracy = (correctly extracted fields) / (total extractable fields). We used this corpus to set baseline benchmarks before going live.

Production pipeline metrics: Aggregate success rate, processing time, and error category frequencies collected from the live system after launch in March 2026. We don't store individual document contents. Only anonymized outcome data (success/failure, document type, page count) is retained for monitoring.

The "500+" figure is total documents processed through production. The per-type accuracy numbers come from the 15-document test corpus, which is why they're approximate ranges, not precise percentages. We'll update them as the corpus grows.

For context on how OCR accuracy is typically measured in the industry, the Wikipedia article on OCR is a solid starting point. Our AI approach reads document structure holistically rather than character-by-character, but the measurement methodology is the same.

Related Articles

• How to Convert a Scanned Document to Excel (Free, No Software)
• Extract Data from PDF to Excel — Without Installing Software
• How to Convert a Scanned Invoice to Excel (Free, No Software)
• Image to Excel: Turn Any Photo into a Spreadsheet
• Why Copy-Pasting from PDF to Excel Breaks Your Data