Lesson 1: The AWS AI and ML Map on One Page

Why you need a map first

AWS sells more than twenty AI and machine learning services. Several have changed names, and since July 2026 a few no longer accept new customers. Pick the wrong one and you can spend a sprint wiring up something your new account isn't allowed to create. So before you compare costs or write any code, you need a one-page map of what exists, what each service is called today, what it used to be called, and whether you can still sign up for it.

Think of it like a hardware shop. Some aisles sell finished appliances (call an API and get a result). Some sell kits (adapt a model). Some sell raw timber and power tools (train your own). Before you shop, you need to know which aisle you're in and which shelves now say "existing customers only".

The six groups

Some services that used to be in the ready-made aisle have already gone. Amazon Forecast is closed to new customers, and Amazon Lookout for Vision was discontinued on 31 October 2025. If an older architecture diagram shows one of these, treat it as legacy.

Gotcha: One service can span groups. Bedrock appears under both generative AI and search (Knowledge Bases), and AgentCore can call Bedrock models or other providers' models. Put each service where its main job is and add a note for anything else it does.

Current vs former names: reading old docs and Terraform

Renames are the most common cause of "this tutorial doesn't match the console". The ones you'll meet most often:

Marketing names change, but the identifiers in code usually don't. For Agents Classic, AWS states that the bedrock-agent API namespace, SDK clients, CloudFormation resource types and IAM action prefixes are all unchanged. Terraform for SageMaker AI still uses aws_sagemaker_* resources. So when you go through a repository, search for the identifiers, not the product names:

# Inventory AI/ML resources in an existing Terraform codebase
# -o prints only the match; -h hides filenames; sort|uniq -c counts each type
grep -rhoE 'resource "aws_(sagemaker|bedrockagent|bedrockagentcore|kendra|qbusiness|rekognition|comprehend|transcribe)[a-z_]*"' . \
  | sort | uniq -c

# Example output
#   3 resource "aws_bedrockagent_agent"            <- Agents Classic: closed to new accounts
#   1 resource "aws_bedrockagentcore_agent_runtime" <- AgentCore: the current agent platform
#   2 resource "aws_kendra_index"                  <- Kendra: maintenance mode
#   4 resource "aws_sagemaker_endpoint"            <- SageMaker AI, old prefix kept

Every line of that output belongs in your map. The aws_bedrockagent_agent and aws_kendra_index lines are the important ones: that code will run in this account but may fail in a brand-new one.

Lifecycle status: what "closed to new customers" really means

AWS publishes lifecycle states on its AWS Lifecycle Changes page (the old aws.amazon.com/products/lifecycle/ link now redirects to it). It uses three terms:

• Maintenance: new customers can't sign up. Existing customers keep using the service with bug fixes and security updates, but no new features.
• Sunset: existing users should migrate. AWS gives an end date, typically about 12 months out.
• Full shutdown: the service is gone.

The exact rule differs from service to service, so read each one's availability note. Bedrock Agents Classic is a good example:

# In an account WITHOUT Bedrock Agents activity in the past 12 months:
aws bedrock-agent create-agent --agent-name demo \
  --foundation-model <model-id> --agent-resource-role-arn <role-arn>
# An error occurred (AccessDeniedException): Bedrock Agents is in Maintenance Mode.
# New agent creation is not available for accounts without prior service usage.

The check is made per account, and there's no exception process. A pipeline that deploys fine to your old production account can fail with a 403 in a newly created sandbox account. Listing, invoking and updating existing agents still works.

The recommended alternatives:

• Bedrock Agents Classic → Bedrock AgentCore (the managed harness for config-style agents, or code-defined agents)
• Kendra → Bedrock Managed Knowledge Base. It has fewer native connectors (7 against Kendra's 32+) and no faceted search, synonyms or spell checking.
• Q Business → Amazon Quick, which can reuse an existing Q index

Best practice: For each service on your map, record four things: status, the date it took effect, the recommended alternative, and a link to the source. In six months, "I think Kendra is deprecated" is no use to anyone. "Kendra: maintenance, closed to new customers 30 Jul 2026, alternative Bedrock Managed Knowledge Base, link" is something you can act on.

Your map entry, worked example

- service: Amazon Kendra
  group: search-and-retrieval
  job: enterprise search over many connected sources
  former_names: []
  status: maintenance            # no new features since 2026-06-30
  closed_to_new: 2026-07-30
  alternative: Amazon Bedrock Managed Knowledge Base
  iac_identifiers: [aws_kendra_index, AWS::Kendra::Index]
  source: https://docs.aws.amazon.com/kendra/latest/dg/kendra-availability-change.html

Key takeaways

• Six groups: generative AI platforms and models, agents, search and retrieval, ready-made AI APIs, ML platforms, accelerators. File each service under its main job.
• SageMaker became SageMaker AI in December 2024. Bedrock Agents is now Bedrock Agents Classic. CodeWhisperer joined Q Developer in April 2024.
• Code identifiers (aws_sagemaker_*, bedrock-agent) didn't change with the names, so search code by identifier.
• From 30 July 2026, Kendra, Q Business and Bedrock Agents Classic are closed to new customers. Their alternatives are Bedrock Managed Knowledge Base, Amazon Quick and AgentCore.
• Maintenance doesn't mean it stops working, but the check is per account, so new accounts can be blocked.
• Record the status, date, alternative and source link for every service.