AWS AI Services for Computer Vision — Rekognition, Textract, Bedrock
The AWS AI Practitioner (AIF-C01) exam expects you to match a computer vision scenario to the right managed AWS service. This page walks through a representative practice question, then covers when to use Rekognition, Textract, Bedrock multimodal, or SageMaker.
Short answer
The correct answer is B. Amazon Textract.
Amazon Textract is the AWS managed service purpose-built for extracting text, key-value pairs, and tables from scanned documents, PDFs, and images. For invoice processing, Textract offers the AnalyzeDocument API with FORMS and TABLES features, plus a specialized AnalyzeExpense API that returns normalized fields such as vendor name, invoice number, totals, and line items out of the box. It preserves the structure of the document — which plain OCR does not — so downstream systems can load values into an ERP without fragile regex parsing. The scenario has three decisive clues: scanned documents as input, structured extraction (key-value pairs and tables) as output, and a fully managed service with no custom training. Those three together uniquely point to Textract. Textract is serverless, scales automatically for batches of thousands of documents, and integrates with S3, Lambda, and Step Functions for end-to-end pipelines. For the AIF-C01 exam, treat any scenario that mentions scanned forms, invoices, receipts, tax documents, contracts with key-value fields, or tables inside PDFs as a Textract signal — even if the word OCR is not used.
The Question
A finance team receives thousands of scanned supplier invoices each week as PDFs and images. They need to automatically extract structured data — including vendor name, invoice number, line items, and totals — so it can be loaded into an ERP system. They want a fully managed AWS service that returns key-value pairs and table structure without training a custom model. Which service best fits this requirement?
Why B is correct
Amazon Textract is the AWS managed service purpose-built for extracting text, key-value pairs, and tables from scanned documents, PDFs, and images. For invoice processing, Textract offers the AnalyzeDocument API with FORMS and TABLES features, plus a specialized AnalyzeExpense API that returns normalized fields such as vendor name, invoice number, totals, and line items out of the box. It preserves the structure of the document — which plain OCR does not — so downstream systems can load values into an ERP without fragile regex parsing. The scenario has three decisive clues: scanned documents as input, structured extraction (key-value pairs and tables) as output, and a fully managed service with no custom training. Those three together uniquely point to Textract. Textract is serverless, scales automatically for batches of thousands of documents, and integrates with S3, Lambda, and Step Functions for end-to-end pipelines. For the AIF-C01 exam, treat any scenario that mentions scanned forms, invoices, receipts, tax documents, contracts with key-value fields, or tables inside PDFs as a Textract signal — even if the word OCR is not used.
Why the other options are wrong
Amazon Rekognition analyzes images and video for objects, scenes, faces, celebrities, text in images, and unsafe content. It can detect text via DetectText but does not return structured key-value pairs or preserve table layout, so it is not the right fit for invoice data extraction.
SageMaker with a custom OCR model would work technically, but it violates the explicit "fully managed, no custom training" requirement. Building and hosting a custom document AI model is the opposite of the serverless, pre-built Textract path the scenario describes.
A text-only foundation model in Bedrock cannot ingest a scanned image or PDF directly. You would still need an upstream OCR step to turn the document into text, which brings you back to Textract. Text-only models also do not return structured forms and tables natively.
Choosing the right AWS managed CV service
AWS splits computer vision into managed services for specific jobs, plus SageMaker for anything custom. For the AIF-C01 exam, memorize the primary job each service is built for. Amazon Rekognition handles image and video analysis. Core features include object and scene detection, face detection and comparison, face search against a collection, celebrity recognition, text in images (DetectText), personal protective equipment detection, and content moderation for unsafe or inappropriate imagery. If the scenario mentions detecting people, faces, objects, activities, or moderating user-uploaded content, Rekognition is the answer. Amazon Textract handles document understanding. It extracts raw text (OCR), key-value pairs from forms, and table structure from scanned documents and PDFs. Specialized APIs exist for invoices and receipts (AnalyzeExpense), identity documents (AnalyzeID), and lending documents. If the scenario mentions invoices, forms, receipts, tax documents, contracts, or extracting fields from scans, Textract is the answer. Amazon Bedrock offers multimodal foundation models — for example Anthropic Claude and Amazon Nova — that can take an image plus a text prompt and reason about the image in natural language. Bedrock fits open-ended visual question answering, image captioning, or flexible extraction where schemas change often. It is more expensive per call and less deterministic than Rekognition or Textract for their specialized jobs. Amazon SageMaker is for building, training, or hosting custom CV models when the managed services cannot handle the task — for example, detecting a proprietary defect on a factory line or classifying medical images against a bespoke taxonomy. The heuristic: prefer Rekognition or Textract for common CV jobs, reach for Bedrock multimodal when the task is open-ended or conversational, and only choose SageMaker when the scenario explicitly requires a custom model.
Ready to see how you'd score?
Take the free practice quiz and find out which AWS Certified AI Practitioner domains you need to focus on. No signup required.
Practice 5 similar questions
Same cert, same or adjacent domain. Use these after reviewing the explanation.