AWS Textract

Extract text, forms, tables, and query answers from documents with Amazon Textract. Runs synchronous OCR for single pages and asynchronous jobs for multi-page PDFs in S3.

8 actions available

Vendor invoice PDF lands in the intake folder

Agent runs Analyze Document to pull fields, tables, and totals

Agent checks the confidence scores on each extracted field

Agent matches the vendor and total against the purchase order

Agent creates the bill record with the extracted line items

AP team gets a summary of the processed invoice

Low-confidence fields or a PO mismatch route to AP before the bill posts

What This Integration Enables

Textract reads documents structurally, not just as a wall of text. It pulls fields and their values, keeps tables as rows and columns, and can answer targeted questions like the invoice total or the due date. For multi-page PDFs it runs asynchronous jobs against S3 and returns the result when the job completes. Extraction is only half the job. The other half is knowing when the extraction is good enough to trust. Textract returns a confidence score per field, and an orchestration layer uses those scores to decide what posts automatically and what a person checks first. FlowRunner is built for that layer, so document processing scales without quietly writing a misread total into your books.

Without FlowRunner

Invoices typed by hand Staff key vendor, amount, and line items from every PDF

Tables lost in OCR Line-item tables collapse into unusable text

No confidence signal A misread digit posts to the ledger unnoticed

With FlowRunner

Fields extracted automatically Vendor, totals, and line items pulled from the PDF

Tables kept intact Line-item tables come back as structured rows

Low-confidence flagged Uncertain fields are routed to a person before posting

Use Case Scenarios

Invoice Processing

Vendor invoices arrive as PDFs. The agent runs Analyze Document to extract the vendor, totals, and line-item table, checks the confidence on each field, and matches the invoice against its purchase order. Clean, matched invoices create a bill record automatically. Anything with a low-confidence field or a PO mismatch is routed to AP. The team handles exceptions instead of typing every invoice.

Form Digitization

Paper forms are scanned into multi-page PDFs. The agent starts an asynchronous Textract job against the file in S3, waits for completion, and maps the extracted fields to a record. Forms that extract cleanly flow straight into the system. Forms with unreadable sections are held for a person, so no record is created from a bad scan.

Document Query at Intake

Rather than extract an entire contract, the agent asks Textract specific questions: the effective date, the renewal term, the total value. It records the answers and their confidence. High-confidence answers populate the record; anything uncertain is surfaced to the person who owns the document, so the flow never guesses on a key term.

Human-in-Loop Highlight

The danger in document automation is not a failed extraction; it is a confident-looking wrong one. Textract returns a confidence score with every field, and FlowRunner uses it to draw the line. When a field falls below the threshold or an extracted total does not match its purchase order, the agent routes the document through a [human-in-the-loop](/concepts/human-in-the-loop/) step: it pauses, shows the person the original document alongside the extracted value, and waits. They correct or confirm. The clean documents post themselves; the uncertain ones get a human read before they touch the ledger.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via preferred channel

Human decides

Agent resumes with decision

Agent Capabilities

8 actions

Text Detection

Detect Document Text Runs synchronous OCR on a single-page image (JPEG, PNG, or TIFF) or single-page PDF and returns every detected line and word as Textract Block objects.

Document Analysis

Analyze Document Runs synchronous document analysis to extract structured data using one or more feature types: FORMS (key-value pairs), TABLES (rows and cells), QUERIES (natural-language questions answered from the document), SIGNATURES (signature locations), and LAYOUT (reading order and layout elements).
Analyze Expense Synchronously analyzes an invoice or receipt and extracts financially relevant data. Supply the document as a FlowRunner file URL (single page, up to ~5 MB) or an S3 object.
Analyze ID Synchronously analyzes identity documents such as U. S.

Asynchronous

Start Document Text Detection Starts an asynchronous OCR job over a document stored in Amazon S3 (used for multi-page PDFs and TIFFs, up to 500 MB / 3000 pages).
Get Document Text Detection Retrieves the results of an asynchronous OCR job started by Start Document Text Detection.
Start Document Analysis Starts an asynchronous document analysis job over a document stored in Amazon S3 (used for multi-page PDFs and TIFFs, up to 500 MB / 3000 pages).
Get Document Analysis Retrieves the results of an asynchronous document analysis job started by Start Document Analysis.

Start building with AWS Textract

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us