Amazon Bedrock: A Practical Guide to Generative AI on AWS

Jonas Jovaišas Senior Technical Product Manager

September 6, 2026

AWS partner dedicated to startups

• 2000+ Clients
• 5+ Years of Experience
• $10M+ saved on AWS

Amazon Bedrock is AWS’s fully managed platform for building generative AI applications with foundation models from Amazon and leading AI providers.

For Bedrock’s serverless foundation models, AWS manages the underlying model-serving infrastructure, allowing developers to invoke models through managed APIs rather than provisioning GPU infrastructure themselves. Bedrock also provides tools for Retrieval-Augmented Generation (RAG), agents, guardrails, model evaluation, customization, and production AI workloads. Bedrock Marketplace further expands the model catalog with models that can be deployed on managed endpoints.

In 2026, Bedrock is considerably more than an API for accessing large language models. AWS now supports 100+ foundation models from providers including Amazon, Anthropic, DeepSeek, Moonshot AI, MiniMax, OpenAI, xAI, and others, with additional models available through Amazon Bedrock Marketplace.

At Cloudvisor, we work with Amazon Bedrock in real AWS environments. We have implemented Bedrock as part of production AI architectures for customers such as MavenPay and We Love Joe, and Cloudvisor is also an Anthropic Authorized Reseller for Amazon Bedrock.

What is Amazon Bedrock?

Amazon Bedrock is a fully managed AWS service for developing and operating generative AI applications.

It provides access to foundation models without requiring teams to provision model-hosting infrastructure themselves. Developers can experiment with models, integrate them into applications through APIs, connect models to private company data, build AI agents, apply safety controls, and evaluate model performance from within the AWS ecosystem.

This abstraction is one of the main reasons teams choose Bedrock. You can focus on the application, workflow, and customer experience rather than maintaining the infrastructure required to serve large foundation models.

Bedrock is particularly relevant when a company already uses AWS and wants its AI workloads to fit within its existing cloud architecture, identity controls, security processes, monitoring, and governance practices.

How does Amazon Bedrock work?

At its simplest, your application sends a request to Amazon Bedrock, Bedrock invokes the selected foundation model, and the model response is returned to your application.

A basic architecture might look like:

Application → Amazon Bedrock API → Foundation model → Response

Production systems often add several other layers:

Application → authentication → Bedrock → model → Knowledge Base or tools → Guardrails → application response

This lets teams build considerably more sophisticated applications without operating the underlying model infrastructure themselves.

Foundation models and unified model access

One of Bedrock’s biggest advantages is model choice.

AWS currently supports more than 100 foundation models from multiple providers. Teams can select models based on reasoning ability, latency, modality, cost, context requirements, or the specific problem being solved.

Bedrock supports AWS-native runtime APIs as well as model-specific and compatible interfaces depending on the model. AWS recommends the bedrock-runtime endpoint for new applications using its native model access.

That flexibility also reduces the need to architect an entire application around one AI provider from day one.

Amazon Bedrock Knowledge Bases and RAG

A foundation model only knows what is available in its training data and current prompt context. Businesses often need AI applications to answer questions using private or frequently changing information.

Amazon Bedrock Knowledge Bases supports Retrieval-Augmented Generation, commonly known as RAG.

With RAG, relevant information is retrieved from an organization’s own data and passed to the model as context when a question is asked.

This can be used to build applications such as:

• internal knowledge assistants
• customer support systems
• document search
• product information assistants
• compliance or policy search
• technical documentation tools
• company-specific AI copilots

Bedrock Knowledge Bases can handle embedding, retrieval, reranking, and integration with different data sources and models, reducing the amount of custom RAG infrastructure teams need to build themselves.

Amazon Bedrock Agents and AgentCore

Bedrock can also support AI systems that do more than answer questions.

Agents can reason about a request, determine what actions are required, use tools or external systems, retrieve information, and complete multi-step workflows.

For teams building more sophisticated agentic applications, Amazon Bedrock AgentCore provides modular capabilities for running and scaling agents, memory, tool integration, identity and access, browser and code execution, observability, evaluation, and governance. These capabilities can be used together or independently depending on the agent architecture.

We have a separate guide covering Amazon Bedrock AgentCore in more depth for teams evaluating production AI agents.

Amazon Bedrock Guardrails

Generative AI applications also need controls around what users can submit and what models are allowed to return.

Amazon Bedrock Guardrails can apply policies to prompts and model responses. Depending on the configuration, these can include content filters, denied topics, sensitive-information filtering, word filters, prompt-attack protection, and grounding checks.

Guardrails can be used with direct model inference, Knowledge Bases, agents, and other Bedrock workflows.

They are useful for reducing risk, but guardrails should be viewed as one part of a broader AI security and governance architecture rather than a replacement for application-level controls.

Which AI models are available on Amazon Bedrock?

The Bedrock model catalog changes quickly, which is why we would not recommend maintaining a static list of every model on this page.

AWS currently documents 100+ supported foundation models, with additional models available through Amazon Bedrock Marketplace.

Representative providers and model families include:

Model availability can differ by AWS Region, API, and inference option, so we recommend checking AWS’s current model catalog when making an architectural decision.

The important point is not having access to the largest possible number of models. It is choosing the model that delivers the right combination of quality, latency, security, and cost for your actual use case.

What can you build with Amazon Bedrock?

Bedrock can support a broad range of generative AI applications.

AI assistants and customer support

Companies can create conversational assistants that answer questions, summarize information, retrieve company-specific knowledge, and guide users through workflows.

Combining a foundation model with a Bedrock Knowledge Base allows responses to be grounded in internal documentation rather than relying only on general model knowledge.

AI agents and workflow automation

Agents can be used for multi-step business processes where the AI needs to reason, retrieve information, call APIs, and take actions.

Examples could include processing requests, managing inventory tasks, supporting internal operations, or orchestrating SaaS workflows.

Document analysis and information extraction

Bedrock models can analyze and reason over business information such as documents, reports, images, and other content.

For some highly structured extraction tasks, however, a purpose-built AWS service may be more appropriate than a generative model. The correct architecture depends on how deterministic the output needs to be.

Software development

Modern foundation models can support code generation, debugging, documentation, technical research, and developer-assistance workflows.

Bedrock lets AWS-based companies integrate these capabilities while keeping model access within their wider AWS environment.

Generative AI for existing SaaS products

One of the most practical uses of Bedrock is adding AI features to an existing application without turning the company into a machine learning infrastructure team.

This could include summarization, intelligent search, recommendations, conversational interfaces, classification, content generation, image analysis, or agentic functionality.

How much does Amazon Bedrock cost?

There is no single Amazon Bedrock price.

Pricing depends on the model, provider, modality, service tier, inference method, and additional Bedrock capabilities used.

For many models, on-demand inference means paying according to actual usage. Other workload patterns can use alternative service tiers or batch processing.

AWS currently offers batch inference at 50% lower pricing than on-demand inference for select foundation models, which can be useful for asynchronous workloads that do not require immediate responses.

Additional Bedrock features can have their own charges. For example, Knowledge Bases, Guardrails, model evaluation, Data Automation, and agent infrastructure can introduce additional usage-based costs depending on the architecture.

The real Bedrock cost question

For production applications, we normally recommend looking beyond the published token price.

The real cost is influenced by:

• how many requests users make
• average input and output size
• which model is selected
• whether every request requires the most capable model
• RAG and retrieval architecture
• agent tool usage
• retries and failed requests
• prompt design
• caching
• traffic growth
• observability and related AWS services

A more expensive model can sometimes produce a lower total application cost if it completes a task in fewer calls. Likewise, using the highest-performing model for every request can be unnecessarily expensive when a smaller model would deliver the same business result.

Model selection should therefore be treated as both a technical and FinOps decision.

Is Amazon Bedrock secure?

Security is one of the main reasons AWS customers consider Bedrock instead of connecting an application directly to multiple external model APIs.

AWS states that customer prompts and model outputs in Amazon Bedrock are not shared with third-party model providers and are not used to train or improve the underlying base models.

Data is encrypted in transit and at rest, and AWS KMS can be used where customer-managed encryption controls are required.

AWS PrivateLink can also establish private connectivity between a VPC and Amazon Bedrock without routing that traffic over the public internet.

Bedrock also integrates with AWS security and governance capabilities, allowing AI workloads to be incorporated into the same broader cloud-security model as other AWS applications.

Is Amazon Bedrock GDPR compliant?

AWS states that customers can use Amazon Bedrock in compliance with GDPR and that Bedrock is within the scope of several AWS compliance programs and certifications.

That does not mean using Bedrock automatically makes an AI application GDPR compliant.

Compliance depends on the complete system: what data is collected, the legal basis for processing, Region selection, retention, identity and access controls, application architecture, subprocessors, logging, and organizational processes.

For European workloads, these requirements should therefore be considered during architecture design rather than after the application has already been deployed.

How we use Amazon Bedrock with Cloudvisor customers

Generic Bedrock examples are useful, but the most important question is what the service looks like when it becomes part of a real production architecture.

We have used Amazon Bedrock in customer environments where AI needed to operate alongside networking, containers, databases, security controls, CI/CD, infrastructure as code, and cost management.

MavenPay: a production-ready AI bookkeeping agent

MavenPay was developing an AI-powered bookkeeping SaaS platform and needed to move from a local prototype to production-grade AWS infrastructure within a 60–90 day launch window.

We built a scalable AWS foundation using Terraform, VPC networking, ECS with Fargate, RDS PostgreSQL, CI/CD automation, Secrets Manager and other AWS services.

Amazon Bedrock was integrated to provide the application’s AI agents with direct access to Amazon Nova Pro within the AWS environment.

The key lesson from this project is that implementing Bedrock was only one part of the solution.

A production AI application also needs repeatable infrastructure, secure connectivity, monitoring, deployment automation, budgets, and operational governance.

We Love Joe: Bedrock and SageMaker in the same architecture

Not every AI workload should use the same service.

For We Love Joe, we implemented a hybrid AI architecture using Amazon Bedrock for managed foundation-model access and image analysis while using Amazon SageMaker for proprietary model training and inference.

The broader environment used AWS services including ECS Fargate, S3, SQS, CloudFront, GuardDuty, CloudTrail, and other AWS infrastructure while keeping the architecture aligned with EU data-residency requirements.

This is a good example of why we do not view the Bedrock decision as simply “use Bedrock or don’t.”

Different parts of the same application can benefit from different AWS AI services.

Amazon Bedrock vs SageMaker AI

Bedrock and SageMaker AI both support AI workloads, but they solve different problems.

Bedrock is often a strong choice when the priority is to ship a generative AI capability quickly without operating the model infrastructure.

SageMaker becomes more relevant when a company needs deeper control over training, deployment, model lifecycle, infrastructure, or unit economics at scale.

The two can also be used together, as we did in the We Love Joe architecture.

We cover this decision in more detail in our Bedrock vs SageMaker AI guide.

When should you use Amazon Bedrock?

Amazon Bedrock is particularly attractive when:

• you want to add generative AI to an AWS application quickly
• you want access to multiple model providers
• you do not want to manage model-serving infrastructure
• you need RAG over private business data
• you are building AI agents or automated workflows
• AWS security and identity integration matters
• you want to evaluate different models before committing to one
• AI is an application capability rather than proprietary model infrastructure

For many startups and SaaS companies, this makes Bedrock a practical starting point for validating AI features before taking on more infrastructure complexity.

When might Amazon Bedrock not be the right choice?

Bedrock is not automatically the best solution just because a project involves AI.

If your competitive advantage depends on owning and training proprietary models, SageMaker AI or another ML platform may make more sense.

If the task is highly structured and deterministic—for example, basic OCR, transcription, or a narrow classification workflow—a purpose-built AI service may be cheaper and easier to operate.

And sometimes the correct decision is not to add generative AI at all.

A technically impressive LLM workflow that increases costs, latency, and unpredictability without improving the user experience is not a successful AI implementation.

The right question is therefore:

What is the simplest architecture that creates measurable business value?

How to get started with Amazon Bedrock

The easiest way to explore Bedrock is through the AWS Management Console.

Developers can select a model from the model catalog, experiment in a playground, and then integrate model inference into an application through an AWS SDK or API.

A simplified Python pattern using the Bedrock runtime looks like this:

import boto3

bedrock = boto3.client(

    “bedrock-runtime”,

    region_name=”eu-west-1″

)

response = bedrock.converse(

    modelId=”YOUR_MODEL_ID”,

    messages=[

        {

            “role”: “user”,

            “content”: [

                {“text”: “Summarize the key benefits of Amazon Bedrock.”}

            ]

        }

    ]

)

print(response[“output”][“message”][“content”][0][“text”])

In a production environment, the model call is usually only the beginning.

You will also need to consider IAM permissions, application authentication, logging, error handling, quotas, model fallback, data access, cost controls, guardrails, deployment automation, and monitoring.

Why trust Cloudvisor on Amazon Bedrock?

Our Amazon Bedrock guidance is based on more than product documentation.

Cloudvisor is an AWS Advanced Tier Partner, and our team has implemented Amazon Bedrock as part of production customer environments involving AI agents, application infrastructure, security, networking, CI/CD, databases, containers, and cost management.

Our Bedrock work includes the MavenPay AI bookkeeping platform, where we implemented Amazon Nova Pro through Bedrock, and We Love Joe, where we combined Amazon Bedrock with SageMaker AI in a hybrid machine learning architecture.

Cloudvisor is also an Anthropic Authorized Reseller for Amazon Bedrock, giving us direct experience supporting companies that want to use Claude models within their AWS environment.

We use that implementation experience when evaluating whether Bedrock is the right service, which model or architecture fits the workload, and what is required to move an AI project from proof of concept into production.

How Cloudvisor can help with Amazon Bedrock

Cloudvisor is an AWS Advanced Tier Partner, and our team works with companies building and scaling AI applications on AWS.

We are also an Anthropic Authorized Reseller for Amazon Bedrock, allowing us to support startups and SMBs that want to use Claude models as part of their AWS infrastructure.

Our work can extend beyond selecting a model.

We can help assess:

• whether Bedrock is the right solution at all
• model and architecture selection
• Bedrock vs SageMaker decisions
• RAG and Knowledge Base architecture
• AI agent infrastructure
• AWS security and networking
• application infrastructure
• DevOps and CI/CD
• cloud cost management
• production readiness and scaling

For companies that are still deciding where AI can create meaningful value, our AI Readiness Assessment evaluates your existing AWS environment, potential AI use cases, and a practical implementation roadmap covering services such as Amazon Bedrock, Amazon Q, and SageMaker AI.

Talk to our AWS AI experts to evaluate your use case and build an Amazon Bedrock architecture designed for production, not just a proof of concept.

Frequently asked questions about Amazon Bedrock

1. Is Amazon Bedrock an LLM?

No.

Amazon Bedrock is an AWS service that provides managed access to many foundation models, including large language models.

Models such as Claude or Amazon Nova can run through Bedrock, but Bedrock itself is the platform and service layer around those models.

2. How many models does Amazon Bedrock support?

AWS currently documents 100+ foundation models from providers including Amazon, Anthropic, DeepSeek, Moonshot AI, MiniMax, OpenAI, xAI, and others.

Amazon Bedrock Marketplace provides access to more than 100 additional popular, emerging, and specialized foundation models.

Because the catalog changes frequently, always check current model and Region availability before designing around a particular model.

3. Does Amazon Bedrock train on my data?

AWS says no.

Inputs and model outputs are not shared with third-party model providers and are not used by AWS or those providers to train the underlying base models.

This is an important distinction for businesses working with proprietary or sensitive information.

4. Is Amazon Bedrock free?

Bedrock is generally usage-based rather than a single fixed subscription.

The amount you pay depends on the models and capabilities you use. Different models have different inference pricing, and features such as Guardrails, Knowledge Bases, agents, or other Bedrock components may introduce additional charges.

For a proof of concept, model cost may be small. At production scale, model selection, prompt size, traffic, retries, RAG architecture, and agent behavior can materially change the total cost.

5. What is the difference between Amazon Bedrock and ChatGPT or Claude?

ChatGPT and Claude are end-user AI products and interfaces provided by OpenAI and Anthropic respectively.

Amazon Bedrock is an AWS platform developers use to build their own applications using foundation models from multiple providers.

For example, a company might use Claude through Amazon Bedrock as the reasoning layer inside its own SaaS application rather than asking employees or customers to use the standalone Claude interface.

6. Does Amazon Bedrock support Claude?

Yes.

Anthropic Claude models are available through Amazon Bedrock, subject to current model and Region availability.

Cloudvisor became an Anthropic Authorized Reseller for Amazon Bedrock in May 2026, strengthening our ability to support AWS customers adopting Claude within their existing AWS environments.

7. What is Amazon Bedrock Knowledge Bases?

Amazon Bedrock Knowledge Bases is AWS’s managed capability for connecting generative AI applications to external or private data.

It is commonly used for RAG, where relevant information is retrieved before the model generates its answer.

This allows applications to answer questions based on business-specific information rather than relying only on a model’s pre-existing knowledge.

8. Should I use Amazon Bedrock or SageMaker AI?

Use Bedrock when the priority is fast access to managed foundation models, RAG, agents, and generative AI capabilities without operating the underlying model infrastructure.

Consider SageMaker AI when you need greater ownership over training, model deployment, MLOps, or specialized machine learning workloads.

In practice, the answer can also be both. Our We Love Joe implementation combines Amazon Bedrock and SageMaker because the two parts of the AI architecture have different requirements.

Ready to build with Amazon Bedrock?

Amazon Bedrock has evolved from a convenient foundation-model API into a broad platform for building production generative AI applications on AWS.

The model is only one part of a successful implementation. Architecture, data, security, observability, cost management, deployment, and operational ownership matter just as much once an AI proof of concept reaches real users.

At Cloudvisor, we help teams make those decisions based on the application they actually need to build.

Contact our team to discuss your Amazon Bedrock use case, or start with an AI Readiness Assessment to understand where AI can create the most value in your AWS environment.

Additional Resources

1. Amazon Bedrock Pricing
   • Learn about the pricing details and options for Amazon Bedrock, helping you understand the cost implications of using this service.
2. Amazon Bedrock Documentation – What Is Bedrock?
   • Access the official documentation to get an in-depth understanding of what Amazon Bedrock is and how it works. This resource provides valuable insights into the platform’s capabilities.
3. Amazon Bedrock API Reference
   • Explore the API reference documentation for Amazon Bedrock, which can be invaluable for developers looking to integrate and interact with Bedrock’s features and functionalities programmatically.

Last updated: August 2026
Written by: Jonas Jovaišas, Senior Technical Product Manager, Cloudvisor

Expert review: This guide has been prepared using Cloudvisor’s hands-on experience implementing Amazon Bedrock in production AWS environments. Our team has deployed Bedrock as part of real customer architectures, including MavenPay and We Love Joe. Cloudvisor is an AWS Advanced Tier Partner and an Anthropic Authorized Reseller for Amazon Bedrock.