Ready-Made AI Services

In part twenty-three we got answers from data that's already organized. But a lot of your data isn't: photos, voice calls and scanned invoices. Understanding them used to need a machine learning team. Today, much of it is one API call.

AWS offers three levels, and most people should start from the top:

Images: Rekognition #

Rekognition analyzes images and video without any training on your part:

• Labels: what's in the image, such as a car, a beach or a crowd.
• Text: reading words in images, such as signs.
• Content moderation: catching inappropriate images before they're published, and you can train it further on your own examples.
• Faces: detecting, comparing and searching for them.

Note that the Streaming Video and Batch Image Content Moderation features closed to new customers on April 30, 2026. Other features aren't affected, and you can process frames or batches yourself through the image API.

Speech: Transcribe and Polly #

Transcribe turns speech into text, from a file in S3 or a live stream. It can tell speakers apart, add your own vocabulary such as product names, and hide personal data on supported languages.

Polly does the opposite, turning text into speech. You can control pronunciation in two ways:

• Lexicons: teach it to read AWS as "Amazon Web Services".
• SSML: tags in the text that add pauses, emphasis or even a whisper.

Translate #

Amazon Translate translates between 75 languages, in real time or in batches from S3. Add custom terminology to keep your brand names and terms as they are. Pairing it with Polly and Transcribe gets you a whole translated conversation.

Conversation: Lex and Connect #

Lex builds chatbots and voice bots. You define what the user may want, an intent, and the details you need to carry it out, called slots.

Connect is a contact center in the cloud: phone numbers, queues, flows and agents, with no equipment, and you pay by usage. Together they might work like this:

1. A customer calls the number on Connect.
2. They say "I want to change my appointment", and Lex understands the intent and asks for the date.
3. A Lambda function updates the appointment in your system.
4. If the bot gets stuck, Connect transfers the call to an agent.

Text: Comprehend #

Comprehend reads text and extracts its meaning: language, sentiment, entities such as people and places, key phrases and personal data. And you can train it to classify your own categories, such as the type of support ticket.

Comprehend Medical does this for clinical text: conditions, medications, doses and tests, and it detects protected health information.

Its Topic Modeling, Event Detection and Prompt Safety Classification features closed to new customers on April 30, 2026.

Documents: Textract #

Textract goes beyond reading text from a scanned file: it understands structure, so it returns forms as field-and-value pairs, and tables as rows and columns. It also has dedicated paths for invoices, receipts and identity documents.

Recommendations: Personalize #

Personalize builds "you might also like" from your users' behavior: you give it who viewed or bought what, and you get an API that returns recommendations in real time, or in batches for email campaigns. You don't need to design a model yourself.

Search over your documents: what about Kendra? #

You may read about Amazon Kendra, an intelligent search service for internal documents. It stopped accepting new customers on July 30, 2026, and AWS points to Bedrock Knowledge Bases as the alternative, which brings us to the next section.

Generative models: Amazon Bedrock #

Bedrock gives you foundation models from many providers, such as Anthropic, Amazon, Meta and Mistral, through one API, with no servers. You pay per token, meaning by the amount of text going in and out. The Converse API has a single form across models, so changing a model is a matter of changing one parameter.

You no longer need to enable models one at a time, since access is now controlled with IAM. The exception is Anthropic's models, which need a one-time form before first use.

Knowledge Bases: answers from your documents #

A model doesn't know your company's policies. Instead of training it, you find the relevant passages and hand them to it with the question. This is called RAG, and Knowledge Bases handles it for you:

The rest of the toolbox #

Bedrock Agents Classic closed to new customers in favor of AgentCore, so don't start with it if you see it in an older article.

Your own model: SageMaker AI #

When ready-made isn't enough, such as predicting risks from data only you have, Amazon SageMaker AI (formerly SageMaker) covers the journey: notebooks, data preparation, training, then an endpoint serving predictions. JumpStart gives you hundreds of ready models to start from.

Don't confuse it with the new Amazon SageMaker, an umbrella that gathers data, analytics and AI in one studio. Also, several of its add-ons, such as Ground Truth, Clarify and Model Monitor, closed to new customers in mid-2026, so check a feature's status before relying on it.

Bedrock or SageMaker AI? #

AWS's own advice: start with Bedrock using prompts and Knowledge Bases, and move to SageMaker AI when you need to change the model's behavior.

And Arabic? #

If your application serves Arabic speakers, check this first, because not all services are equal:

Chaining services #

The real strength lies in connecting them with what you've already learned. Imagine call recordings landing in S3, where each step triggers the next through events or Step Functions:

Without any machine learning, a manager now reads a summary of every call and sees which ones ended badly.

Hands-on: a review passes through four services #

We'll take a customer review, analyze it, translate it, voice it and reply to it. Everything is a direct call with nothing to set up. These services aren't in every Region, so if a command reports that it can't connect, add --region us-east-1 to it. From CloudShell:

REVIEW="The app is great, but delivery was very slow and support never replied."

aws comprehend detect-dominant-language --text "$REVIEW" \
  --query "Languages[0]"

aws comprehend detect-sentiment --language-code en --text "$REVIEW" \
  --query "[Sentiment,SentimentScore]"

aws comprehend detect-key-phrases --language-code en --text "$REVIEW" \
  --query "KeyPhrases[].Text"

You'll see the language, the sentiment with its scores, and the key phrases. Now translate it to Arabic, then check that Comprehend understands the translation too:

ARABIC=$(aws translate translate-text --text "$REVIEW" \
  --source-language-code en --target-language-code ar \
  --query TranslatedText --output text)

echo "$ARABIC"

aws comprehend detect-sentiment --language-code ar --text "$ARABIC" \
  --query Sentiment --output text

Voice the translation with Polly:

aws polly synthesize-speech --voice-id Zeina --output-format mp3 \
  --text "$ARABIC" review.mp3

ls -lh review.mp3

To listen, from the CloudShell Actions menu choose Download file, and enter review.mp3.

Finally, ask a model on Bedrock to reply to the customer. The first command looks up the identifier of a low-cost Amazon Nova Lite model in your Region. If it prints None, add --region us-east-1 to both commands:

MODEL=$(aws bedrock list-inference-profiles \
  --query "inferenceProfileSummaries[?contains(inferenceProfileId,'nova') && contains(inferenceProfileId,'lite')].inferenceProfileId | [0]" \
  --output text)
echo $MODEL

aws bedrock-runtime converse --model-id $MODEL \
  --messages "[{\"role\":\"user\",\"content\":[{\"text\":\"Reply to this customer in two short sentences: $REVIEW\"}]}]" \
  --query "output.message.content[0].text" --output text

Got a photo? Upload it from Actions, then Upload file, and ask Rekognition what's in it:

aws rekognition detect-labels --image-bytes fileb://photo.jpg \
  --max-labels 5 --query "Labels[].[Name,Confidence]" --output table

Summary #

• Start from the ready-made services, then Bedrock, and only then SageMaker AI.
• Rekognition for images, Transcribe and Polly for speech, Translate for translation, and Comprehend for text.
• Textract reads structured documents, but it doesn't support Arabic.
• Knowledge Bases answers from your documents without training a model, and it's the alternative to Kendra.
• Connecting these services with events is where the real value is.

Test yourself #

Before you move on, try these questions to check you've got this part. Each one is a practical scenario, and the explanation appears once you answer.

This completes the data and AI section. Your application is big now, and it has a lot running in it. How do you know it's healthy, and who changed what? In part twenty-five we start the final section, with monitoring and auditing through CloudWatch, CloudTrail and Config.