Batch Data Extraction: One Folder In, One Spreadsheet to Review
Manually extracting data from a stack of documents takes roughly 3 minutes per page. Batch extraction does each page in 5-10 seconds and puts the whole folder into one spreadsheet, roughly 18x faster, with no templates and no pre-sorting.
5-10s per page · Any format (PDF/JPG/PNG) · Mixed batch · Up to 99% accuracy on printed text
What You Can Extract from a Batch of Documents
Name the columns you need once, and the AI pulls those values from every document in the batch, whatever their format or layout. Each document becomes one row in a single spreadsheet.
You type these column names once. The AI applies them to every document in the batch: PDFs, scans, photos, Word files, and text files alike.
Batch Data Extraction Sounds Simple Until You Try It on a Real Folder
Every tool promises "upload multiple files, get one spreadsheet." Most deliver that only when every document shares the same layout. Real folders don't arrive that way: vendors differ, formats mix, and a few files are always worse than the rest. And when the goal is the text itself rather than named columns, batch OCR that reads the whole folder is the lighter run on those same files.
Where Batch Processing Breaks Down on Real Folders
One template per document layout means one template per vendor. Template-based tools demand a parser per layout, so when every vendor's invoice is positioned differently, you build and maintain a separate configuration for each one. Even AI-powered tools like Nanonets or Rossum require per-document-type training or field mapping before they can process a batch. A user on Reddit described the situation bluntly: "We're manually copying data from PDFs into Excel every week and it's taking so much." The template solution doesn't remove the work; it shifts it from copying to configuring.
Real folders don't arrive sorted by format. A native PDF, a scanned image, and a phone photo each need a different processing pipeline in traditional OCR tools, so you end up running separate batches per format and merging the outputs by hand. One developer on Reddit put it starkly: "PDF is a human readable output format. It isn't a computer readable input format," which is why format-specific pipelines struggle the moment a folder mixes them.
Batch size and progress visibility are afterthoughts. Many tools process files one at a time behind a single progress bar. When one document fails, there's no way to know which one or why, so you re-check the entire batch by hand. The real bottleneck isn't AI speed; it's the lack of per-document status.
How One Batch Covers a Whole Mixed Folder
You define the output, not the input rules. Instead of building a parser for each layout, you type the column names you need (for example Document Date, Vendor, Total Amount, Tax, Due Date) and the AI finds those values on every document by understanding what they mean. "Invoice Date" on one vendor's PDF and "Transaction Date" on a scanned receipt from another all resolve to your "Document Date" column. The column names themselves are the only configuration, and they apply to every document in the batch regardless of vendor, format, or layout. This is Custom Column Extraction: the AI reads for meaning rather than enforcing positional rules.
All formats share one pipeline, so there's no pre-sorting step. Drop a folder of PDF invoices, scanned JPG receipts, mobile PNG photos, and the occasional Word or text file into the same upload. The visual language model reads every file the same way: it looks at the pixels and pulls out the columns you named. There's no "PDF pipeline" versus "image pipeline," just one batch, one column definition, and one output spreadsheet.
Per-document progress and independent processing mean one bad file doesn't sink the batch. Each document is processed independently with its own status tracked in real time. The batch view shows exactly which documents have completed, which are processing, and which encountered issues, instead of one progress bar hiding individual failures. Processing takes 5-10 seconds per page (vs ~3 minutes of manual entry per page), and the output is one consolidated XLSX or CSV file with the columns you defined.
Batch Data Extraction in Practice: From a Folder of Mixed Files to One Clean Spreadsheet
Here's how a mixed folder moves through one batch, and how you keep track of every file along the way.
Upload the Entire Batch, Formats and All
Your month-end folder contains 30 files: 18 PDF invoices from 10 different ERP systems, 7 scanned receipts as JPGs from field offices, and 5 mobile photos of handwritten delivery notes. Drag all 30 into the upload area. PDF, JPG, PNG, WebP, and even Word or text files go in together, with no pre-sorting and no "select document type" step. Batch upload handles roughly 30 files at once, each up to 10 MB.
Name Your Columns Once, They Apply to Every Document
Type Document Date, Vendor, Invoice Number, Total Amount, Tax Amount, Due Date. These become the exact headers in your output spreadsheet. The AI maps "Invoice Date" on digital PDFs, "Transaction Date" on scanned receipts, and "Date" on handwritten notes all to your "Document Date" column, because it understands what each field means rather than matching a document-type template. Optionally add a computed column like Line Total (Qty × Unit Price) and the AI performs the arithmetic during extraction.
Review the Batch File by File, Fix Only What Needs It
Processing takes 5-10 seconds per page. The output is a single XLSX file: 30 rows, one per document, with the columns you defined, and every row keeps its own status so you can jump straight to the files that need a look. If a vendor's invoice doesn't show tax, that cell is empty rather than a fabricated zero, and that one row can be corrected without touching the other 29. The AI also supports Inferred Columns: define a "Category" column with options like "Invoice / Receipt / Delivery Note / PO" and the AI classifies each document as it extracts, producing both the data and the document-type label in a single run.
When Batch Extraction Works Best, and When to Be Cautious
Batch extraction reliably handles mixed formats and varied layouts when the target fields are labeled. Understanding the boundaries keeps your batches running smoothly.
When It Works Best
Mixed-format batches sharing common field concepts. When every document carries a date, total, and vendor name at different positions, the AI maps them to your named columns without per-document configuration.
Clear printed labels near values. Up to 99% accuracy when fields are labeled, such as "Invoice Date:" and "Total:" at any page position.
~30 documents per batch with per-file status visibility. Each document's status (processing / completed / error) visible in real time. Larger volumes split into multiple batches, each producing an independent spreadsheet.
When to Be Cautious
Heavily degraded source quality. Faded thermal receipts, scans under 200 DPI, or compressed images with artifacts. The AI compensates using context better than traditional OCR, but poor source quality is the biggest accuracy bottleneck regardless of document type.
Highly document-specific fields with no shared meaning to anchor on. If one document contains "Container Seal Number" and nothing else in the batch references something similar, the AI has no cross-document signal to work from. Type-specific fields extract better when batched with similar documents.
Unlabeled numeric values without nearby field names. A dollar figure sitting alone in a paragraph without "Total" or "Amount" nearby gives the AI no label to match against. Label-value pairs extract reliably; isolated numbers in narrative text may not.
Frequently Asked Questions
Can I batch-extract data from PDFs, scanned images, and phone photos in the same upload, or do I need to separate them by format?
All formats go into the same batch. Upload PDFs, JPGs, PNGs, WebP, AVIF, Word documents, and text files together. The visual language model processes every file through one pipeline: it reads the pixels, understands the content, and pulls out the columns you named. A Document Date found in a native PDF comes through the same way whether it sits in a scanned JPG or a phone photo, because the AI reads for meaning rather than checking the file format. No separate batches per format, no manual format pre-sorting.
What happens if a field I defined, like Tax Amount, doesn't exist on some documents in the batch?
The AI leaves that cell empty for that document rather than inventing a value or halting the batch. A purchase order without a tax line will show a blank cell in the Tax Amount column, while invoices in the same batch with tax will have the value filled in. No error stops the batch, and no made-up value silently enters your spreadsheet. This is especially useful when your batch contains different document types or vendor formats that carry overlapping but not identical fields, because each document is processed independently within the same column framework.
Can I extract line items in batch mode when each document has a different number of rows, from a single item on one invoice to 50 rows on a purchase order?
Yes. The AI reads each document's structure independently and generates as many output rows as that document contains. A single-line invoice produces one row with the header fields (Document Date, Vendor, Total Amount) and one line item row (Quantity, Unit Price, Line Total). A 50-line purchase order produces 50 rows, all sharing the same column headers. You don't pre-define row counts and you don't get a broken table from variable-length content. Computed Columns like "Line Total (Qty × Unit Price)" also work across variable-length batches.
Can I add calculated columns to a batch extraction, for example Line Total computed as Qty × Unit Price during the batch run?
Yes. Computed Columns let you define calculations that the AI performs during extraction rather than requiring a separate Excel formula step. Type "Line Total (Qty × Unit Price)" as a column name, and the AI multiplies the values it finds for Quantity and Unit Price on each document in the batch, outputting the result directly. This works across every document regardless of how many line items each one has. For logged-in users, Rule Format supports more complex multi-step calculations including cross-row aggregation and conditional logic.
Does one poor-quality document in the batch affect the others?
No. Each document is processed independently with its own status tracked in the batch view. If one file in a batch of 30 is a faded scan, that document's extraction accuracy will be lower, but the other 29 files are unaffected. The poor-quality document still appears in the output with whatever values the AI could extract, and its row is easy to identify for review. For best results, scan at 300 DPI or higher, and use the per-document status in the batch view to find the files that need attention without re-running the entire batch.