AWS AI Services Explained: SageMaker, Bedrock, Rekognition, Comprehend, Polly, Transcribe and Lex

AWS has more AI services than most teams will ever use, and the names do not always explain themselves. This guide describes the ones that matter for application teams, grouped by the question they answer: do you want to build your own model, call a hosted foundation model, or plug in a prebuilt capability such as image recognition or speech? For the account and cost side, see working on AWS.

Build and deploy your own models: Amazon SageMaker #

SageMaker is the fully managed platform for building, training and deploying machine learning models at scale. It provides hosted notebooks, managed training jobs on CPU or GPU instances, experiment tracking, a model registry and hosted inference endpoints with autoscaling. Use it when you have your own data and a task that benefits from a custom model: demand forecasting, fraud scoring, recommendation, or fine-tuning an open model on proprietary text.

Call foundation models: Amazon Bedrock #

Bedrock gives access to foundation models from several AI companies through a single API, so you can build generative AI applications without managing model infrastructure. Teams use it to experiment across models, fine-tune for specific use cases, build agents and retrieval workflows, and scale from prototype to production on the same API. It suits chat assistants, summarization, content generation and code assistance. Our Bedrock overview covers the workflow in more detail. Model access and request quotas are enabled per account and per region and grow with usage.

Prebuilt services for vision, language and speech #

These services need no training data and no model expertise: you send an image, a document or audio and get structured results back. They are the fastest route to an AI feature when the task is one they already cover.

Which one do I need? #

• The task is standard (moderate images, transcribe audio, read sentiment): start with the prebuilt service.
• The task is generative (answer questions over your documents, draft text, build an assistant): Bedrock.
• The task needs your own model (predictions from tabular data, a domain-specific classifier, a fine-tuned open model you want to own): SageMaker.
• You are not sure: prototype on Bedrock or a prebuilt service first; move to SageMaker only if quality or cost forces it.

Supporting services #

AI applications on AWS typically also use S3 for datasets and documents, EC2 GPU instances or SageMaker training instances for compute, Lambda and API Gateway to expose features, and IAM to scope what the application may call. CloudWatch and CloudTrail provide the monitoring and audit trail.

Cost and quota notes #

• Bedrock is billed per token or per provisioned throughput; prebuilt services are billed per request or per unit processed; SageMaker bills for the instances behind notebooks, training and endpoints.
• Stop notebooks and delete endpoints when idle; they are the most common source of surprise AI bills.
• Model access and request-per-minute quotas start low on new accounts and are raised through the Service Quotas console with usage history. Quotas belong to the account they were granted to.

A worked example: support ticket triage #

Suppose a team wants incoming support tickets summarized, classified by sentiment and routed. A pragmatic AWS design uses Comprehend to score sentiment and extract entities, Bedrock to produce a two-sentence summary and suggested reply, Lambda to orchestrate the calls, and S3 plus DynamoDB to store results. Nothing is trained; the whole feature is API calls. If, after a few months, the classification quality is not good enough on the company’s own categories, a small classifier trained in SageMaker on the accumulated tickets can replace that one step without changing the rest.

Getting started in order #

1. Create a dedicated development account or environment with a budget alert.
2. Enable model access in the Bedrock console for the region you will use.
3. Create an IAM role for the application with permissions limited to the specific services and actions it calls.
4. Prototype with the prebuilt services and Bedrock from a notebook or a small Lambda function.
5. Add CloudWatch metrics and logs from the start so you can see latency and cost per request.
6. Move to SageMaker only for the steps where a custom model demonstrably improves results.

Security notes for AI applications #

• Scope IAM permissions to the exact model IDs and services in use.
• Do not send secrets or regulated data to models unless your data handling review allows it; Bedrock and the prebuilt services document their data usage policies.
• Log prompts and outputs with care; they often contain customer data.
• Rate-limit your own endpoints so a bug cannot burn through model quotas and budget.

Official references #

The AWS documentation has developer guides for SageMaker, Bedrock and each prebuilt service. Our AWS platform overview covers the platform as a whole.

Frequently asked questions #

Is Bedrock a replacement for SageMaker?

No. Bedrock hosts foundation models you call; SageMaker is where you train and host your own. Many teams use both.

Do I need a data scientist to use these services?

Not for Bedrock or the prebuilt services. SageMaker rewards ML experience but also offers low-code tooling.

Why does a Bedrock model show as unavailable?

Model access is enabled per account and per region, and some models require a request. Check the model access page in the Bedrock console for your region.