The discriminator across the AI catalogue is whether anybody is training anything. A scenario describing a team with labelled data, a model to train and an endpoint to deploy is SageMaker AI. A scenario describing a task with a name (transcribe, translate, extract, moderate, recommend) and no mention of training is one of the pre-trained services, and choosing SageMaker AI there means building something that already exists.
The pairs worth separating: Textract reads documents and forms, while Comprehend reads meaning out of text that is already text. Transcribe turns speech into words, Polly turns words into speech. Lex builds the conversation; Connect Customer, the contact centre product that AWS used to call simply Amazon Connect, is where it might sit.
On the analytics side the separation is where the data lives and what is being asked of it. Athena queries objects in S3 in place with nothing to provision, charged by the data scanned, which makes file format and partitioning part of the cost. Redshift loads data into a columnar warehouse for repeated analytical work. Kinesis and Firehose carry data in motion, Glue transforms it and catalogues it, and Quick Sight is what a business user finally looks at.
Flash card
The AI/ML and analytics service catalogues: recognise each service by the single job it does, and tell the pre-trained task services from the platforms for building your own.
- Build your own: Amazon SageMaker AI builds, trains and deploys models. Use a foundation model: Amazon Bedrock gives API access to foundation models for generative AI, and Amazon Q Developer is the assistant built on them. Amazon Q Business is closed to new customers; Amazon Quick replaces it.
- Vision and speech: Rekognition for images and video (objects, faces, moderation), Transcribe for speech to text, Polly for text to speech, Translate for languages, Textract for text and structure out of scanned documents and forms.
- Language and prediction: Comprehend for sentiment, entities and key phrases, Lex for chatbots and voice bots, Personalize for recommendations. Fraud Detector (fraudulent transactions and accounts) and Kendra (enterprise search) are both closed to new customers, so neither is a live pick.
- Query: Athena runs SQL directly against data in S3 with no cluster, charged per data scanned. Redshift is the columnar data warehouse for analytical queries over large volumes. OpenSearch Service handles search and log analytics.
- Move and prepare: AWS Glue is serverless ETL and a data catalogue, Amazon Kinesis handles streaming data, Amazon Data Firehose loads streams into S3, Redshift and OpenSearch, and Amazon MSK is managed Kafka.
- Process and present: Amazon EMR runs managed Hadoop, Spark, Hive, Presto and Trino; Amazon Quick Sight, now a feature of Amazon Quick, builds business intelligence dashboards; AWS Lake Formation governs and secures a data lake on S3.
Pick it when
Pick a pre-trained AI service when a scenario describes a named task (read this document, transcribe this call, detect sentiment) and nobody is training a model. Pick SageMaker AI when they are. Pick Bedrock when the scenario is generative.
It's the wrong answer when
SageMaker AI is the wrong pick whenever a pre-trained service already does the job: a standard task with no data science team is not a training project. Athena is wrong for a transactional workload, Redshift for a high-volume single-row lookup.