AWS AI services are Amazon Web Services’ managed portfolio of artificial-intelligence and machine-learning tools (pre-trained AI APIs, the SageMaker ML platform, and generative AI services such as Amazon Bedrock) that let organizations build, train, and run AI without managing the infrastructure underneath.
Knowing that is the easy part. The harder part is that AWS AI is several different things sold under one name, and picking the wrong one means paying for complexity you don’t need, or hitting a ceiling you didn’t see coming. This guide sorts the portfolio by the job each service does, tells you when to use it, and shows what happened when we used these services on real enterprise problems.
It’s also been a busy year. AWS has retired or frozen several services that older “AWS AI services” lists still recommend, so we’ve flagged those too.
What are AWS AI services?
AWS AI services are AI and machine-learning capabilities that AWS runs for you. You pay for what you use, and AWS handles servers, scaling, and patching. That’s what people mean by AWS managed AI services, and it’s why AI services in AWS can go from idea to working prototype in days rather than quarters.
The useful way to read the AWS AI cloud services catalog is as a ladder of ownership. The higher you climb, the more control you get, and the more work lands on your team:
- Pre-trained AI services. AWS built the model. You call an API.
- Generative AI services. AWS hosts foundation models. You build the application, connect your data, and set the rules.
- The ML platform. You build and run your own models, with AWS providing the tooling.
- AI infrastructure. You bring the workload and AWS provides the compute.
Most enterprises end up on more than one rung. The skill is knowing which problem belongs on which one.
AWS AI and machine learning services
Here’s the AWS AI services list, sorted by tier, with what each service does and when it’s the right pick. Think of it as a quick reference for AWS AI/ML services, or AWS AI ML services if that’s how you search for them. Where AWS has changed a service’s status in 2026, we say so.
Generative AI services
- Amazon Bedrock: managed access to foundation models from AWS and third-party providers through one API, with tooling for RAG, guardrails, and agents. Use it when you’re building an assistant, a natural-language interface, or a content workflow and don’t want to train a model.
- Amazon Q: AWS’s family of AI assistants. Use it when you want a ready-made assistant instead of building one. One change to know about: Amazon Q Business stopped accepting new customers on July 30, 2026, and AWS points new users to Amazon Quick for similar capabilities.
- Amazon Q Developer: AI help for coding, debugging, and optimization. Use it when your engineering teams want faster delivery without changing their tools.
The ML platform
- Amazon SageMaker: the AWS machine learning platform for building, training, deploying, and monitoring custom models. Use it when you have a prediction, classification, or specialized data science problem that your own data can solve better than a general model.
- SageMaker Studio: an integrated workspace for data scientists and ML engineers. Use it when your team wants one place for the full workflow.
- SageMaker JumpStart: a catalog of pre-built models and solution templates. Use it when you’d rather start from something that already works.
Pre-trained AI services (AI APIs)
These cover computer vision, document processing, conversational AI, forecasting, and personalization. You get a result back without training anything.
- Amazon Rekognition: analyzes images and video. Use it for object detection, face analysis, and content moderation.
- Amazon Textract: extracts text and structured data from documents. Use it for forms, invoices, contracts, and financial records.
- Amazon Comprehend: natural language processing. Use it for pulling topics, entities, and sentiment out of large volumes of text.
- Amazon Transcribe: speech to text. Use it for call analytics, meeting transcripts, and voice features.
- Amazon Polly: text to speech. Use it when your product needs to talk.
- Amazon Translate: language translation. Use it for multilingual products and support.
- Amazon Lex: text and voice chatbots. Use it for structured conversational interfaces.
- Amazon Personalize: recommendations. Use it for tailoring products or content to each user.
- Amazon Forecast: time-series forecasting. AWS no longer offers it to new customers, though existing users can carry on. For new forecasting work, use SageMaker.
- Amazon Kendra: ML-powered enterprise search. Points new projects to Amazon Bedrock’s managed Knowledge Base.
AI infrastructure and chips
- AWS Trainium: chips built for training models. Use it when you’re training large models and compute cost is a line item someone watches.
- AWS Inferentia: chips built for running trained models. Use it when you’re serving models at scale.
- EC2 accelerated instances: GPU and specialized hardware on demand. Use it when you need flexible capacity for development or deployment.
AWS generative AI services in depth
Amazon Bedrock (AWS Bedrock to anyone searching for it) is where most enterprise generative AI work on AWS starts. You pick a foundation model, connect it to your data, and add controls. It’s a practical fit for knowledge assistants, document analysis, natural-language interfaces, customer service automation, and content generation.
Teams spend a lot of energy on which model to choose. In our experience, that’s rarely what decides the outcome. What matters more is whether the model can reach the right information, and whether you can tell when it’s wrong.
That’s what retrieval-augmented generation (RAG) is for. It connects a foundation model to your own sources, so answers come from retrieved material instead of the model’s general memory. Bedrock supports it with guardrails, which set limits on model behavior, and knowledge bases, which handle retrieval. In 2026, AWS added a managed Knowledge Base on Bedrock aimed at enterprise RAG pipelines, which is the replacement it recommends for Kendra.
Sometimes plain RAG isn’t enough because the relationships between facts matter as much as the facts. A global mobility company had that problem: map data had to be validated against more than 5,000 pages of technical documentation. Intellias built a GraphRAG solution that structured the knowledge so the system could reason across it, and validation time dropped by up to 60%.
The same pattern works on the data side. For an identity and location technology company, we replaced siloed data with a platform centered on SageMaker Unified Studio, then added a generative AI assistant on Bedrock. People who used to wait hours for an answer now ask a question in plain language and get one in eight seconds or less.
Amazon Bedrock vs. Amazon SageMaker: Which should you use?
Both sit at the center of AI on AWS, and they’re easy to confuse. The difference is who builds the model.
Reach for Bedrock if the model you need already exists and your work is wrapping it in an application: an assistant, a RAG system, a content tool. You’ll need less ML infrastructure experience.
Reach for SageMaker if the model has to be yours: trained on your data, tuned for your task, and owned by your team. You get more control, and your team takes on data preparation, monitoring, and ML engineering.
A quick test: if you can describe your problem as “answer, summarize, or generate,” start with Bedrock. If you’d describe it as “predict, score, or classify,” start with SageMaker.
You also don’t have to choose. The identity and location platform above uses both. Match each service to the part of the workload it’s best at.
AWS agentic AI services
A chat assistant answers. An agent finishes a job. It breaks a goal into steps, calls your systems through APIs, makes decisions within limits you set, and hands back a result.
AWS agentic AI services have been the fastest-moving part of the portfolio in 2026. Amazon Bedrock AgentCore is AWS’s platform for building and running agents, and it works with any framework and any model, not just those hosted on Bedrock. At the AWS Summit in New York, AWS made its AgentCore harness generally available, so teams can define an agent’s model, tools, and instructions in configuration instead of hand-coding the orchestration loop. The original Bedrock Agents service is now Amazon Bedrock Agents Classic.
More capability means more design work. An agent that touches operational systems needs clear access rules, audit trails, and limits on what it can do alone. That’s worth it only when the workflow justifies it.
Here’s one that did. An agribusiness organization built each campaign brief by hand, pulling from six disconnected data sources, and it took two to three days. Intellias developed a multi-agent AI platform that unifies those sources, and the same briefs now take less than an hour. The goal wasn’t to “use agents.” It was to remove a slow, manual, cross-system process, and agents were the right way to do it.
Before building one, ask whether there’s a multi-step workflow across systems whose speed you can measure. If there isn’t, you probably want an assistant.
How to choose the right AWS AI service
Start from the problem. Work through these in order and most of the list falls away.
- What kind of problem is it? Generating or summarizing points to Bedrock. Predicting or classifying points to SageMaker. A standard task like extraction, transcription, or translation points to a pre-trained API first.
- Is customization worth paying for? Building your own model buys control and costs development and operating time. If a managed service already covers the need, building needs a business case. If your edge is proprietary data and a specialist use case, it probably has one.
- What data does it need? An enterprise application has to answer from the right company information. That means retrieval, access controls, data architecture, and evaluation, which should be designed alongside the model, not bolted on later.
- Who will run it? A sophisticated architecture is a liability if your team can’t operate it. Match the service to the skills you have.
- How sensitive is the data? If security and compliance rule, look for services that support governance and access control, and plan your guardrails early.
For many enterprises the answer ends up being a combination. That’s fine, and it’s usually healthier than forcing everything onto one platform.
What’s new in 2026
Three changes affect real decisions:
- Agents grew up. Bedrock AgentCore, with the harness generally available, is now AWS’s route to production agents.
- Retrieval got a managed option. Bedrock’s managed Knowledge Base takes over where Kendra left off for new projects.
- Some older services froze. Amazon Q Business (Amazon Quick is the successor), Kendra, Bedrock Agents (now Classic), and Forecast are closed to new customers, though existing users can keep using them.
If you’re comparing options from an older article, check the status of every service on it.
Pick the simplest thing that works
AWS gives you several ways into AI. The best one is rarely the service with the longest feature list. It’s the simplest architecture that solves your business problem and fits your data, your team, and your production requirements.
Intellias is an AWS AI Services Competency partner. Our AWS AI consulting work covers generative AI, machine learning, RAG, agentic systems, and data platforms, with production results in mobility, property intelligence, agribusiness, and identity and location technology. If you’re weighing AWS AI solutions for a specific use case, we’re happy to think it through with you.
A practical view of AWS AI/ML services, from first use case to production scale
Choosing an AWS AI service is easier when you start with the problem you need to solve. Our extended guide takes a closer look at the options and the decisions enterprise teams need to make when moving AI into production.
Download the whitepaper to explore:
- A practical map of the AWS AI/ML landscape
See how generative AI, machine learning platforms, pre-trained AI services, and AI infrastructure fit together, with guidance on what each service does and when to use it. - A deeper comparison of Amazon Bedrock and Amazon SageMaker
Compare the two platforms by purpose, use cases, model approach, and technical expertise to understand which better fits your requirements. - Six common enterprise AI scenarios
Explore options for generative AI applications, custom ML models, business process automation, early-stage AI adoption, security- and compliance-sensitive use cases, and long-term scalability. - How generative and agentic AI fit into enterprise workflows
See where generative AI applications and multi-step agentic systems can support different business needs, from knowledge assistants to workflow automation. - Real-world examples with measurable outcomes
See how AWS-based solutions have helped reduce validation time by up to 60%, cut total cost of ownership tenfold, bring time to insight down to eight seconds or less, and reduce campaign brief creation from days to under an hour. - What to look for in an AWS AI implementation partner
Understand what the AWS AI Services Competency represents and why technical validation, delivery experience, and demonstrated customer outcomes matter when taking AI into production.
Use the whitepaper as a practical reference for evaluating AWS AI/ML options, aligning technology choices with business needs, and planning a clearer route from use case to production.
