Understanding how extraction approaches differ for structured forms, semi-structured business documents, and unstructured text — and why modern AI handles all three
The document extraction landscape splits into three categories that determine how hard it is to get data out of a document and into a spreadsheet or database. Structured documents have a fixed, predictable layout where every piece of data appears in the same position on every document. Tax form W-2 is a perfect example: box 1 is always wages, box 2 is always federal tax withheld, and every W-2 from every employer follows the same template. Extracting data from structured documents is a solved problem — you map field coordinates once, and the same template works for every document of that type.
Unstructured documents sit at the opposite end of the spectrum. A legal contract, a business letter, a medical report, or a research paper contains information embedded in natural language paragraphs with no fixed layout. The relevant data — a contract value, a diagnosis, a key finding — could appear anywhere in the document, phrased in any number of ways. Traditional OCR extracts the text but cannot identify which text contains the data you need. That requires language understanding, which is a fundamentally different capability than character recognition. Between these two extremes sits the semi-structured category, where most real-world business documents actually fall.
Lido uses AI-powered extraction that handles all three document categories without requiring you to pre-sort your documents or configure templates. The AI automatically detects whether a document is structured, semi-structured, or unstructured and applies the appropriate extraction approach. Upload a batch containing invoices, contracts, and forms together, and the system processes each one correctly. For a deeper look at how extraction accuracy varies across these document types, see our accuracy guide. Start with 50 free pages.
Structured documents are forms with fixed fields in fixed positions. Government tax forms, standardized applications, and regulated reporting documents are structured. Every IRS Form 1040 is identical in layout. Every FDA 510(k) submission follows the same template. Every OSHA incident report has the same fields in the same positions. Structured extraction is the easiest category because the problem reduces to coordinate mapping: field X is always at position (row, column) on the page. Template-based OCR tools handle structured documents well, and a single template serves for all documents of the same type. The limitation is that each new document type requires a new template, and the template breaks if the form design changes even slightly — a field that moves half an inch in a new form revision can cause the entire extraction to fail.
Semi-structured documents contain consistent fields in variable layouts. Invoices are the most common example. Every invoice includes a vendor name, invoice number, date, line items, and a total amount. But Vendor A places the invoice number in the upper right corner, Vendor B puts it in a header bar across the top, and Vendor C buries it in a text line below the company logo. A company receiving invoices from 200 different vendors deals with 200 different layouts, each containing the same fields in different positions. Template-based extraction fails at this scale because creating and maintaining 200 templates is impractical, and new vendors arrive constantly with layouts the system has never seen. AI-powered extraction handles semi-structured documents by understanding what fields to look for based on their labels and context rather than their coordinates.
Unstructured documents contain information in natural language without designated fields. Contracts, correspondence, reports, and narrative documents have no field labels, no input boxes, and no consistent layout. The data exists within sentences and paragraphs, and the AI must comprehend the text to identify the relevant information. Extracting the effective date from a contract means finding a sentence like "This agreement shall be effective as of January 15, 2025" somewhere within a multi-page document. Extracting the total contract value might require identifying and summing multiple payment terms scattered across different sections. This is natural language processing, not optical character recognition, and it requires a fundamentally different extraction approach.
Real-world document batches are always a mix. The practical challenge is that organizations rarely process just one document type. An accounts payable department receives structured purchase orders, semi-structured invoices, and unstructured delivery confirmations. A legal department processes structured court filings, semi-structured contracts, and unstructured correspondence. A human resources department handles structured tax forms, semi-structured resumes, and unstructured reference letters. Any extraction system that can handle only one category forces the organization to pre-sort documents before processing, which adds a manual step that negates much of the automation benefit.
Modern AI extraction unifies the three document categories under a single processing pipeline by combining visual layout analysis with natural language understanding. The first step is document classification: the AI examines the document and determines whether it has a form-like layout (structured), a business document layout with identifiable but variable fields (semi-structured), or a text-heavy layout without clear field boundaries (unstructured). This classification happens automatically and determines which extraction strategy the system applies.
For structured documents, the AI identifies the form fields by their visual boundaries — boxes, lines, and shaded areas — and extracts the content within each field. It reads the field labels to understand what each field represents, which makes the extraction robust to minor form layout changes that would break coordinate-based templates. A field that moves from one position to another is still correctly extracted because the AI follows the label, not the coordinates.
For semi-structured documents, the AI uses a combination of layout analysis and field detection. It identifies the key-value pairs on the document — "Invoice Number: 12345", "Date: 03/15/2025", "Total: $4,500.00" — by recognizing the label text and associating it with the adjacent value. For tables and line items, the AI detects the tabular structure visually and extracts each cell with its column context. This approach works regardless of where the fields appear on the page, which is why it handles invoices from hundreds of different vendors without requiring a template for each one.
For unstructured documents, the AI reads the full text and applies natural language understanding to identify the requested information. If you ask for contract dates, payment terms, and party names, the AI locates the relevant sentences and extracts the values from within the natural language context. It handles variations in phrasing — "the agreement is effective from" and "this contract commences on" both indicate the same field — and resolves ambiguity when multiple dates or amounts appear in the document by using the surrounding context to determine which is the effective date versus a reference date.
When your documents are fully structured, extraction is straightforward. If you process a single standardized form type — the same tax form, the same application, the same regulatory filing — any OCR tool with template mapping will handle it. The AI advantage in this category is not accuracy but flexibility: you can change the form design or add a new form variant without rebuilding templates. For organizations that process a small number of highly standardized forms, template-based tools are adequate. For organizations where form designs change periodically or where multiple form versions coexist, AI-powered extraction eliminates the template maintenance burden.
When your documents are semi-structured, AI is essential. The semi-structured category is where most organizations find their biggest extraction pain. Invoices from multiple vendors, purchase orders from different systems, bank statements from various institutions, and shipping documents from international carriers all fall into this category. Template-based tools require a new template for every new vendor or format, creating a maintenance burden that grows with every new business relationship. AI-powered extraction processes these documents immediately without any template configuration. This is the category where the ROI of AI extraction is highest because the alternative — building and maintaining hundreds of templates or relying on manual data entry — is the most expensive.
When your documents are unstructured, AI is the only automated option. Contracts, legal documents, medical narratives, research reports, and business correspondence cannot be processed by template-based extraction at all because there are no fixed fields to map. Before AI, the only option was manual reading and data entry. AI-powered extraction makes it possible to extract specific data points from unstructured documents automatically: pulling contract values and dates from legal agreements, extracting diagnoses and procedures from medical reports, or identifying key findings and recommendations from research documents. The accuracy is lower than structured extraction because natural language is inherently ambiguous, but the time savings over manual reading are substantial.
When you process a mix of all three, a unified tool eliminates the sorting step. The most common real-world scenario is a document inbox that contains a mix of structured forms, semi-structured business documents, and unstructured correspondence. A single AI extraction platform that handles all three categories eliminates the manual pre-sorting step that would otherwise be required. Upload the entire batch, define the fields you want extracted, and the AI classifies each document, applies the appropriate extraction strategy, and outputs a consistent data format regardless of the source document type. This unified approach is what makes document automation practical for organizations that process diverse document types rather than a single standardized form.
Upload structured forms, semi-structured business documents, or unstructured text and get structured spreadsheet data
Structured data extraction processes documents that have a consistent, predictable layout where specific data always appears in the same location — such as invoices, tax forms, and bank statements. The extraction system knows exactly where to find each field. Unstructured data extraction processes documents with no fixed layout — such as contracts, emails, reports, and letters — where the information is embedded in natural language paragraphs and its position varies from document to document. Structured extraction maps coordinates to fields; unstructured extraction requires the AI to read and understand the text to identify the relevant data.
Unstructured documents are significantly harder to extract because the target data has no fixed position and must be identified through language understanding rather than spatial coordinates. In a structured invoice, the total amount is always near a "Total" label in a predictable location. In an unstructured contract, a payment amount might appear anywhere in a multi-page document, embedded in a sentence. Extracting that figure requires the AI to understand the sentence meaning, not just the character positions. However, structured extraction becomes harder when dealing with many different templates.
Traditionally, yes — template-based OCR tools handled structured documents while separate NLP tools handled unstructured text. Modern AI-powered extraction tools handle both. The AI automatically detects whether a document is structured, semi-structured, or unstructured and applies the appropriate extraction approach. This means a single tool can process a mixed batch containing invoices, contracts, and correspondence without pre-sorting by document type.
Semi-structured documents fall between structured and unstructured. They contain the same types of information in every document but the layout varies between sources. Invoices are the classic example: every invoice has a vendor name, invoice number, date, line items, and total, but each vendor uses a different layout, font, and field arrangement. Purchase orders, receipts, shipping labels, and medical claims are also semi-structured. Extracting semi-structured documents requires the AI to understand what fields to look for while being flexible about where those fields appear on the page.
50 free pages. All features included. No credit card required.