AWS offers a broad and changing set of AI and machine-learning services. Start with the workload: what the system must do, which data it may use, how it will be reviewed, and what your team can operate.
This guide compares Amazon Bedrock, SageMaker, Kendra, OpenSearch, Comprehend, Textract, and Rekognition. Features, model support, Regions, quotas, and prices change, so confirm current AWS documentation and test likely candidates with representative data.
Amazon Bedrock: Managed foundation models
Bedrock provides managed access to supported foundation models and related capabilities through AWS APIs. Check the current catalog and Region support for the models, customization, retrieval, evaluation, and guardrail features your workload needs.
Common fit
- Generative AI applications - chatbots, content generation, summarization, code generation
- RAG pipelines with built-in Knowledge Bases
- Multi-model strategies where you want to switch between providers without changing code
- Teams that prefer managed model access and can operate the surrounding application
Items to verify
- Model, feature, and Region availability for the selected architecture
- Whether supported customization options meet the task's evaluation threshold
- Whether the managed inference choices fit the required latency, throughput, cost, and operational control
Compare Bedrock with the other viable options for your task. Use representative prompts and data to measure quality, latency, cost, security fit, and operating effort before choosing an architecture. Our AWS AI consulting practice can help structure that comparison.
Amazon SageMaker: ML development and operations
SageMaker provides capabilities for data preparation, model development, training, tuning, deployment, and monitoring. Compare its supported workflows and infrastructure controls with Bedrock's managed model access based on the model and operating responsibility you need.
Common fit
- Custom model training on your proprietary data
- Supported fine-tuning workflows that require configurable training parameters
- Deploying models to dedicated inference endpoints with custom instance types
- Traditional ML workloads - classification, regression, forecasting, anomaly detection
- Teams with ML engineering capacity that need more infrastructure choice
Items to verify
- The engineering and operating effort for training, deployment, and monitoring
- Cost management requires careful attention to instance selection and endpoint scaling
- Delivery time compared with managed alternatives for the same task
When to use SageMaker over Bedrock
Evaluate SageMaker when the workload needs custom training, a model or deployment option unavailable through Bedrock, configurable inference infrastructure, or traditional ML. Test the candidate architecture and map where data is processed across the account, Region, network, and services.
Amazon Kendra: Enterprise search
Kendra is a managed intelligent-search service. It uses natural-language processing and ranking to search unstructured sources beyond simple keyword matching. Compare its current connectors, features, Regions, quotas, and access-control behavior with the needs of your search experience.
Common fit
- Enterprise search across diverse data sources - documents, wikis, ticketing systems, CRMs
- FAQ-style question answering where users expect direct answers, not document links
- Use cases that require access control inheritance from source systems (SharePoint permissions, Salesforce visibility rules)
Kendra vs Bedrock Knowledge Bases
Bedrock Knowledge Bases provides retrieval for foundation-model applications. Kendra provides managed enterprise search and ranking. Compare them using your sources, identity requirements, retrieval tests, Region, latency, operating model, and current pricing. Kendra can also serve as the retrieval layer for a Bedrock-powered application when that combination fits the architecture.
Amazon OpenSearch: Vector and full-text search
OpenSearch Service and OpenSearch Serverless provide search and analytics capabilities, including supported vector and full-text search. Evaluate them for retrieval when you need configurable indexing, query, or relevance behavior.
Common fit
- Vector storage and similarity search for RAG pipelines
- Hybrid search combining semantic (vector) and lexical (keyword) retrieval
- High-volume, low-latency search workloads with custom relevance tuning
- Teams that need direct control over indexing, sharding, and query optimization
OpenSearch vs Bedrock Knowledge Bases
Bedrock Knowledge Bases supports selected vector-store integrations and manages parts of ingestion and retrieval. Direct OpenSearch exposes more index and query configuration. Compare supported stores, chunking, embeddings, synchronization, access control, relevance, latency, cost, and operating effort with your own documents and queries.
Amazon Comprehend, Textract, and Rekognition: Purpose-built AI
AWS also offers task-specific AI services. Compare them with foundation-model and custom-model approaches when their supported APIs match the work you need to perform.
Amazon Comprehend
Natural-language processing service for supported tasks such as sentiment analysis, entity extraction, language detection, topic modeling, and PII detection. Compare Comprehend with eligible foundation or custom models using representative data, current pricing, latency, Region support, quotas, and the review the output requires.
Amazon Textract
Document-intelligence service for extracting supported text, tables, forms, and structured fields from documents and images. Textract includes specialized APIs for several document types. Validate extraction quality with representative, rights-cleared documents and route low-confidence or consequential fields for human review before using a foundation model or downstream workflow.
Amazon Rekognition
Computer-vision service with supported APIs for image and video analysis, including labels, faces, text in images, content moderation, and custom labels. Compare Rekognition with other eligible vision models using representative, rights-cleared data and measure task quality, latency, Region support, review needs, and current cost.
How these services fit together
Some workflows combine services. For example, a document pipeline might evaluate Textract for extraction, a task-specific or custom model for classification, OpenSearch for retrieval, and Bedrock for language interaction. An agentic AI system may also use separate orchestration, model, and search services when each addition earns its operating cost.
Start with the simplest architecture that meets the test criteria. If the workflow needs several services, define clear interfaces so each component can be evaluated, operated, and changed independently.
Decision framework
When evaluating AWS AI services for a new project, work through these questions in order:
- Is there a task-specific service? Test eligible APIs against a foundation model or custom model on the same representative dataset.
- Do you need generative output? Compare supported Bedrock models with any other options your architecture permits on quality, latency, cost, and controls.
- Do you need custom training? Compare the available Bedrock customization and SageMaker training paths with the data, skills, and infrastructure you have.
- What retrieval behavior do you need? Test Bedrock Knowledge Bases, Kendra, OpenSearch, or a simpler approach with your sources, permissions, and labeled queries.
- What can your team operate? Include deployment, monitoring, incident response, upgrades, and cost ownership in the service comparison.
Test before committing
Service choices affect implementation, data movement, testing, and ongoing operations. Define your requirements, compare viable services with actual data, and record why the selected option met the workload's quality, security, latency, cost, and support needs.
Our AWS AI consulting team helps customers define the test, compare suitable AWS services, and implement the selected workflow with agreed operating controls. For generative AI consulting beyond AWS, we apply the same workload-based evaluation to multi-cloud and hybrid options.
Need help building AI on AWS?
Explore Our AWS AI Consulting Services