Extracting text from a PDF is one thing. Keeping its tables intact is another. Columns merge, rows split across pages, and scanned documents may not contain selectable text at all.
The right document processing tool for PDFs with tables depends on three things: whether the PDF has a text layer, how complex its tables are, and whether you need a one-time conversion or a recurring workflow that sends rows into a spreadsheet or business system.
This guide compares tools by approach, from open-source libraries to intelligent document processing (IDP) platforms, and shows how to test them on your own PDFs.
Key Takeaways
- Scanned PDFs need more than text-based extraction. Tabula and Camelot rely on a text layer, so scans need OCR or vision-based extraction.
- Table structure is the hard part. Multi-page tables, merged cells, multi-level headers, borderless layouts and variable row counts cause missing or misgrouped data.
- Cheapest per page is not cheapest overall. Cloud APIs charge about a cent per page for tables, but someone still builds intake, validation and delivery around them.
- Choose by workflow, not by file. Open-source and desktop tools suit one-off conversions. For recurring PDFs from many senders, an IDP platform such as Parseur extracts the rows and delivers them to Excel, Google Sheets or your business system.
- Test with your hardest PDFs. Vendor accuracy figures are not comparable without the same documents and the same method.
Which Service Extracts Table Data From Scanned Documents Without High Monthly Fees?
For scanned documents, the lowest per-page prices come from cloud APIs: Amazon Textract charges $0.015 per page for table analysis and Google Document AI's Layout Parser $10 per 1,000 pages at list prices. Teams without developers usually spend less overall on an IDP platform with a small plan, such as Parseur from $49 a month for 100 pages, because the subscription includes the intake, validation and spreadsheet delivery that an API leaves to you.
| Option | Published price | Scanned tables | What you still build or do |
|---|---|---|---|
| Amazon Textract (Tables) | $0.015 per page, US West (Oregon), first million pages | Yes | Intake, merging tables across pages, mapping, validation, delivery |
| Google Document AI (Layout Parser) | $10 per 1,000 pages | Yes | Intake, page merging, mapping, validation, delivery |
| Google Document AI (Form Parser) | $30 per 1,000 pages, first million pages | Yes | Same as above |
| Parseur (IDP platform) | $49 a month for 100 pages, $129 for 1,000, $499 for 10,000 | Yes, Vision AI | Define fields once, connect Excel, Sheets or your system |
| Desktop OCR (ABBYY, Acrobat) | Per-seat software license | Yes | Open, convert and clean each file by hand |
Sources: Amazon Textract pricing, Google Document AI pricing, Parseur pricing.
The per-page price only tells half the story. A Textract or Document AI call returns cells and coordinates, so someone still writes code to merge tables that cross pages, map columns, catch bad rows and push results to a spreadsheet. If that someone is you on a Friday afternoon, the "cheap" option is the expensive one. For what that pipeline involves, see what a production Textract pipeline takes.
Why Tables Are the Hard Part of PDF Extraction
PDFs store text and its position on the page, not rows, columns or the relationships between them, so every extraction tool has to rebuild the table structure itself.
When that rebuild goes wrong, you see:
- Merged columns: values from neighbouring columns land in one field.
- Split rows: a long description becomes several records.
- Lost headers: multi-page tables drop their headers or repeat them as data rows.
- Misread totals: subtotals and grand totals come out as line items.
Structure changes results sharply. On the PDF table subset of the OmniDocBench benchmark presented at CVPR 2025, PaddleOCR scored 70.6 on tables with merged cells versus 75.2 without them, and 23.3 on rotated tables versus 74.6 on upright ones. These are table-similarity (TEDS) scores for one model on one dataset, not general accuracy rates, but they show how much a rotated scan or a merged header can cost.
The people who feel this are rarely developers. It is the AP clerk at a distributor whose biggest supplier sends a 30-page invoice every week, the logistics coordinator copying packing lists into a shipment tracker, and the bookkeeper rebuilding bank statement transactions in QuickBooks. Their goal is not a pretty table. It is every row, in the right column, in the system they work in. For how vision models handle these layouts, see vision AI table extraction.
First Check - Does Your PDF Have a Text Layer?
Before choosing a tool, check whether your PDF contains selectable text, because that decides which extraction methods can work at all.
Born-digital PDFs have a text layer, so tools can read each word and its position. Scanned PDFs are images of pages and need OCR or vision-based extraction first.
How to check in 30 seconds: open the PDF and try to select a number inside the table.
- Text is selectable: the PDF likely has a text layer. Tabula, Camelot and pdfplumber may work.
- Text is not selectable: the PDF is an image. Use an OCR-enabled application, a cloud document AI service or an IDP platform.
- Some pages are, some are not: use a tool that handles both.
A text layer does not guarantee clean tables. Multi-page tables, merged cells and irregular layouts still need structural analysis. For image-based files, see scanned PDF extraction.
Table Types and What Breaks Them
Different table layouts fail in different ways, so test the structures your documents actually contain.
| Table type | Typical documents | What to test |
|---|---|---|
| Simple single-page grid | Price lists, reports | Values stay in the right columns |
| Multi-page table, repeated headers | Bank statements, long invoices | No rows lost at page breaks, headers not treated as data |
| Variable row counts | Invoices, purchase orders | Every line item captured, whatever the length |
| Merged cells, multi-level headers | Financial reports | Values assigned to the right header |
| Borderless tables | Statements, packing lists | Column boundaries found without visible lines |
| Wrapped or nested rows | Line items with multi-line descriptions | Each item stays one record |
| Rotated or skewed scans | Faxed or photographed documents | Rows and columns come out in order |
| Tables mixed with fields | Invoices with header details and lines | Header fields and rows extracted and linked |
Missing rows, misplaced values or duplicated headers all mean cleanup before the data is usable.
What to Look for in a Tool for PDF Tables
The right tool preserves table structure, handles your document formats, and delivers usable rows to where you work.
Most teams start with what is free or already installed: Tabula, Acrobat's export, or pasting a PDF into ChatGPT. That works for the first file. It stops working when the same supplier invoice, statement or packing list arrives every week and someone spends Monday morning fixing merged columns. That is the moment to compare tools on cleanup time per month, not just price per page.
- Scanned PDF support: OCR or vision-based extraction.
- Row-level extraction: each table row becomes its own record.
- Multi-page tables: all rows captured across pages, whatever the row count.
- Layout flexibility: different sender layouts without a template each.
- Structured output: Excel, CSV, JSON, Google Sheets or business integrations.
- Recurring intake: documents arrive by email, folder or API.
- Review and correction: users can check and fix data before export.
- Security and compliance: data handling, retention, access controls and compliance documentation.
Structure deserves its own test. In a 2025 ACL study comparing table extraction methods, TATR-OCR reached a 98.2% structural layout F1 score against 94.9% for GPT-4o on 1,000 simpler PubTables-1M tables without merged cells or projected rows. The specialized approach led on structure, GPT-4o scored higher on cell content, and neither result covers every document type.
Tool details come from each vendor's public documentation reviewed on October 27, 2026.
Where Extracted Table Data Goes
Table data is only useful in the app where the work happens. For most teams that is Excel or Google Sheets first, then an accounting system such as QuickBooks, Xero or NetSuite, a tracker such as Airtable or Smartsheet, or a BI tool. Here is how Parseur gets extracted rows into each.
| Destination | How Parseur sends table rows |
|---|---|
| Excel | Native PDF to Excel export, one row per line item |
| Google Sheets | Native PDF to Google Sheets export |
| QuickBooks and Xero | Zapier (QuickBooks, Xero) |
| Airtable and Smartsheet | Zapier (Airtable, Smartsheet) |
| Tableau and BI tools | Zapier (Tableau) or a Sheets or Excel source |
| NetSuite, ERPs, databases | Webhooks, the REST API, or Make and Power Automate |
Tools by Approach
Document processing tools for PDFs with tables fall into six categories: IDP platforms, open-source libraries, desktop OCR applications, cloud document AI APIs, parsing APIs for AI pipelines, and template-based parsers. The right one depends on your PDFs, table complexity, technical resources, and whether this is a recurring workflow.
Parseur publishes this guide and lists its own platform first. The tools are not ranked.
IDP Platforms
IDP platforms extract fields and table rows, validate them and deliver them to downstream systems, which suits invoices, purchase orders, statements and delivery notes arriving every week from different senders.
Parseur
Parseur is an IDP platform for teams that receive table-heavy PDFs every week and need each row in a spreadsheet or business system. It reads scanned and text-based PDFs with its Vision AI and Text AI engines, guided by plain-English field instructions, so there is no template per sender.
Parseur table fields capture repeating data such as invoice line items, and each extracted row becomes its own record. It supports tables spanning multiple pages and offers row-grouping controls for descriptions that wrap over several lines. Bundled PDFs can be split by page, page range, keyword or AI before extraction.
Documents arrive by email, upload or API. Parseur exports to Excel, Google Sheets, CSV and JSON, and delivers through webhooks, Zapier, Make and Power Automate. It has run in production since 2016, processed more than 100 million documents, completed a SOC 2 Type II examination in September 2026, is GDPR compliant and hosts data in the EU. See Parseur security and PDF table extraction.
Best for: teams processing recurring, table-heavy PDFs from many senders who need rows delivered to spreadsheets or business systems.
Not ideal for: a single one-off conversion. Parseur does not provide per-field confidence scores. It validates each result against the extraction schema instead, and a failing document appears in the app, triggers an email notification and fires a webhook.
Nanonets
Nanonets is an IDP platform that extracts fields and tables from scanned and text-based PDFs such as invoices and purchase orders. It supports line-item extraction, validation rules, human review and exports or integrations, with configuration for business-specific rules.
Best for: teams that want configurable validation and exception review across invoices and similar documents.
Not ideal for: occasional conversions. Plan time for rules and integrations. See Parseur vs Nanonets.
Docsumo
Docsumo is an IDP platform for financial and operational documents such as invoices, bank statements and bills of lading. It processes scanned and text-based PDFs, extracts line items and transaction tables, including tables that continue across pages, and delivers through APIs, webhooks or Excel exports.
Best for: finance teams processing statements, invoices and bills of lading with transaction tables.
Not ideal for: occasional conversions. Custom outputs need integration work. See Parseur vs Docsumo.
Lido
Lido extracts structured data from scanned and text-based PDFs into spreadsheets and automated workflows. It supports multi-page tables, merged headers and variable line items, lets users define columns and review results in a spreadsheet, and processes documents from Google Drive, OneDrive, email or API.
Best for: finance and operations teams that want PDF table data in spreadsheets on a recurring basis.
Not ideal for: real-time, sub-second extraction. Lido's documentation estimates 10 to 30 seconds per document. See Parseur vs Lido.
Open-Source Table Libraries (Text-Layer PDFs Only)
Open-source libraries suit selectable-text PDFs and developers who want direct control. Most do not perform OCR, so scans need an OCR step first.
Tabula
Tabula is an open-source desktop application for text-based PDFs. You draw a box around the table and export CSV or spreadsheet-compatible data.
Best for: analysts converting occasional text-based PDF tables without code.
Not ideal for: scans, complex layouts or recurring workflows with many senders.
Camelot
Camelot is a Python library for text-based PDFs, with separate methods for tables that have ruling lines and tables separated by whitespace. It exports to CSV, Excel, JSON and more.
Best for: Python developers scripting extraction for consistent layouts.
Not ideal for: scans without preprocessing, or teams that do not want to maintain code.
pdfplumber
pdfplumber is a Python library for inspecting PDF characters, lines and tables, with configurable table detection.
Best for: developers who need fine control over layout analysis.
Not ideal for: non-technical users or scans without a usable text layer.
Open-source libraries are flexible, but OCR, validation, layout changes and delivery to business systems stay with the developer.
Desktop OCR and PDF Applications
Desktop applications convert scanned and text-based PDFs into editable formats such as Excel, for people who will inspect and fix each result.
ABBYY FineReader PDF
ABBYY FineReader PDF is a PDF editor and OCR application that converts scanned documents into Word, Excel and other formats, individually or in batches, with an OCR Editor and a verification tool for misread characters. FineReader PDF Corporate adds scheduled Hot Folder processing.
Best for: teams that need OCR, editing and reviewed PDF-to-Excel conversion.
Not ideal for: extracting invoice line items as records and delivering them into business systems without extra processing.
Adobe Acrobat
Adobe Acrobat converts text-based and scanned PDFs to Excel, using OCR for scans, and Acrobat Pro's Action Wizard can batch OCR across files. Adobe notes that tables can flatten into unstructured text during export, separating values from their rows or columns, so results need checking.
Best for: converting PDF tables to Excel while keeping PDF editing tools.
Not ideal for: recurring line-item extraction with validated delivery to business systems.
Cloud Document AI APIs
Cloud APIs extract text, tables and structure from scanned and text-based PDFs and return machine-readable results for developers to build on. Validation, exceptions and delivery are yours to build.
Amazon Textract
Amazon Textract detects tables, cells, merged cells, headers, titles and footers, and returns them as structured objects with relationships. It supports PDF, JPEG, PNG and TIFF, and analyzes tables page by page, so multi-page tables need merging in your code.
Best for: development teams building custom workflows on AWS.
Not ideal for: business teams wanting validated rows in a spreadsheet without development.
Google Document AI
Google Document AI's Form Parser extracts key-value pairs, tables, checkboxes and generic fields, and its Layout Parser extracts text, tables and lists. Google documents that a table spanning several pages can come back as two separate tables.
Best for: development teams on Google Cloud building custom pipelines.
Not ideal for: teams expecting multi-page tables as complete records without extra processing.
Azure AI Document Intelligence
Azure AI Document Intelligence's Layout model returns paragraphs and tables, with row counts, column counts, row spans and column spans in its JSON. It extracts tables per page, and custom models support dynamic tables with variable row counts.
Best for: development teams in the Microsoft Azure ecosystem.
Not ideal for: teams that need a complete document-to-system workflow without building validation and integration. For the trade-offs, see build vs buy invoice data extraction.
Document Parsing APIs for AI and LLM Pipelines
Parsing APIs prepare PDFs for large language model (LLM) applications and retrieval-augmented generation (RAG) by preserving text, tables and page structure. They also return table data, but connecting it to business processes is up to you.
LlamaParse
LlamaParse, from LlamaIndex, converts PDFs, scans and Office files to Markdown or JSON, and its Extract API returns schema-shaped data. Tables come out with cell structure as Markdown, HTML, rows or CSV, optionally written to an Excel workbook, and a setting merges tables that continue across pages.
Best for: developers building AI or retrieval applications that need layout-aware parsing.
Not ideal for: operations teams wanting ready-made intake, validation and delivery.
Reducto
Reducto's Parse API turns documents into structured JSON in one call, combining OCR, layout detection and table reconstruction, with HTML (including merged-cell spans), Markdown, JSON rows and CSV outputs. Its Extract API returns schema-typed fields.
Best for: engineering teams building AI agents, RAG systems or custom pipelines.
Not ideal for: teams that do not want to build intake, validation and delivery around an API.
Template-Based PDF Parsers
Template-based parsers extract fields and tables with rules tied to a layout, which works for consistent formats and needs adjusting when layouts change.
Docparser
Docparser uses parsing rules, including Table Data and Line Items rules with defined column boundaries, plus built-in OCR. It accepts documents by email and exports CSV, Excel, JSON or XML. Saved Smart Tables rules lock in a layout, so changed tables may need new rules. AI-assisted rule creation is available.
Best for: teams with consistent table layouts that want configurable rules.
Not ideal for: many senders with changing layouts. See Parseur vs Docparser and AI vs rule-based PDF parsers.
Comparison Table - Document Processing Tools for PDFs With Tables
The table compares 15 tools on scanned PDFs, multi-page tables, variable rows, template needs, outputs and recurring workflows, based on documented features rather than independent accuracy tests.
| Tool / category | Scanned PDFs | Multi-page tables | Variable rows | No template needed | Output | Recurring workflow | Best for |
|---|---|---|---|---|---|---|---|
| Parseur / IDP | Built-in OCR and Vision AI | Tables spanning multiple pages | Yes | AI with plain-English instructions, templates optional | Excel, Sheets, CSV, JSON, webhooks | Email, API, uploads | Recurring table-heavy PDFs from many senders |
| Nanonets / IDP | OCR-enabled | Documented, test page breaks | Line items | AI, workflow configuration may be needed | CSV, XML, XLSX, integrations | Approval and export workflows | Finance and operations documents |
| Docsumo / IDP | Scanned and native PDFs | Cross-page tables documented | Line items and transactions | Pretrained models for supported documents | CSV, JSON, Excel, API, webhooks | Configurable workflows | Financial documents with transaction tables |
| Lido / document processing | Yes, test document quality | Multi-page, verify continuity | Configured columns | AI with configurable columns | Excel, CSV, API | Google Drive, OneDrive, email, API | Spreadsheet-centered recurring workflows |
| Tabula / open source | No built-in OCR | Page by page, manual combining | Depends on detection | Select table area | CSV, TSV, JSON | Manual or scripted | Occasional text-based tables |
| Camelot / open source | No built-in OCR | Developer handles continuity | Detected rows | Settings may need tuning | CSV, Excel, JSON, HTML, Markdown, SQLite | Python scripts | Developer-controlled extraction |
| pdfplumber / open source | No built-in OCR | Developer handles continuity | Detected rows and cells | Configurable detection | Python data structures | Python scripts | Custom layout analysis |
| ABBYY FineReader PDF / desktop OCR | OCR-enabled | Multi-page conversion, verify continuity | Depends on recognition | No template for conversion | Excel and other editable formats | Batch, Hot Folder in Corporate | Reviewed PDF-to-Excel conversion |
| Adobe Acrobat / desktop PDF | OCR-enabled | Multi-page conversion, batch OCR | Depends on recognition | No template for PDF-to-Excel | Excel workbook | Batch actions in Acrobat Pro | PDF conversion and editing |
| Amazon Textract / cloud API | OCR-enabled | Per page, no automatic merging | Rows and cells | Prebuilt table analysis | Block objects in API responses | Developer-built AWS workflow | Custom AWS processing |
| Google Document AI / cloud API | OCR-enabled | May split into two tables | Rows and cells | Prebuilt processors | Structured API output | Developer-built Google Cloud workflow | Custom processing and RAG |
| Azure AI Document Intelligence / cloud API | OCR-enabled | Per page, combine in post-processing | Rows and cells, dynamic tables in custom models | Prebuilt Layout, optional custom models | JSON, Markdown | Developer-built Azure workflow | Custom Microsoft-based processing |
| LlamaParse / AI parsing API | Supports scans | Configurable continued-table merging | Structured table output | No positional template | Markdown, HTML, JSON, CSV, XLSX | API-based implementation | AI applications and retrieval |
| Reducto / AI parsing API | OCR-enabled | Configurable merging, validate results | Rows and schema extraction | No positional template | HTML, Markdown, JSON, CSV | API-based implementation | AI agents and custom pipelines |
| Docparser / template-based | Built-in OCR | Rule-dependent, test page breaks | Repeating tables | Rules per layout, AI-assisted creation | CSV, Excel, JSON, XML | Email intake, API, webhooks | Consistent layouts with configurable rules |
Notes: some features depend on editions, tiers or configured workflows. For multi-page tables, test whether a tool returns one continuous table or one per page. "No template needed" does not mean no setup.
Which Tool Fits Which Job
The best tool depends on whether you are converting an occasional file, building your own pipeline, or processing recurring documents from many senders.
| Your job | Best fit |
|---|---|
| Convert one text-based PDF table to Excel | Tabula or a desktop PDF application |
| Convert one scanned PDF table to Excel | ABBYY FineReader PDF or Adobe Acrobat, then check rows and columns |
| Extract scanned tables at low cost, with developers available | Amazon Textract or Google Document AI, paid per page |
| Build table extraction into your own product | Amazon Textract, Google Document AI or Azure AI Document Intelligence |
| Feed PDF tables into an AI or retrieval application | LlamaParse or Reducto |
| Process the same supplier layout every month | Docparser, or an IDP platform if layouts or requirements change |
| Process table-heavy PDFs from many senders every week into Sheets, ERP or CRM | An IDP platform such as Parseur, Nanonets, Docsumo or Lido |
Which should you choose? Choose Parseur when table-heavy PDFs arrive regularly from different senders and the rows must reach your spreadsheet or business system without building a pipeline. Use Tabula or desktop tools for one-off conversions, cloud APIs for developer-owned workflows, and parsing APIs for AI applications. For a broader list, see the best PDF parser and the best invoice OCR tools.
How to Check Table Extraction on Your Own PDFs
Test each tool on 10 to 20 of your hardest documents, including scanned invoices, multi-page tables, merged headers and different supplier layouts. One clean PDF proves nothing about your weekly workload.
- Row accuracy: are line items missing, duplicated or split?
- Multi-page tables: are rows kept across page breaks without repeated headers?
- Column alignment: do descriptions, quantities, prices and totals land in the right fields?
- Data formatting: are dates, currencies, decimals and negative values preserved?
- Layout changes and output: does a new supplier format need reconfiguration, and does the data reach your spreadsheet or system correctly?
Count errors, manual corrections and setup time for each tool. The best choice delivers usable rows from your real PDFs with a review and maintenance load your team can live with.
Last updated on
