Amazon Bedrock is a fully managed cloud service that gives you access to powerful pre-trained AI models from companies like Anthropic, Meta, and Stability AI, all through a single set of tools. For someone studying for the AIF-C01 exam, understanding Bedrock is crucial because it represents the easiest way for businesses to add AI capabilities without needing to build or train their own models from scratch. It solves the problem of complexity and cost, letting you focus on using AI rather than managing infrastructure.
Jump to a section
A simple way to picture Amazon Bedrock Fundamentals
Standing in front of a massive buffet line at a busy food hall, the Holiday Feast Hall. You see dozens of steaming trays of food from different cuisines — Italian pasta, Japanese sushi, American barbecue, Indian curries. Each tray is a 'foundation model', a pre-prepared recipe cooked by a world-class chef (like Anthropic, Meta, or Stability AI), kept warm and ready to serve.
Instead of spending days in your own kitchen cooking from scratch, you just grab a plate and walk down the line. You point at what you want: 'I'll have some of that text-generation tray, please.' The chef behind the counter doesn't hand you the entire pot — they give you a small serving on your plate, tailored to your request. You don't need to know how the chef chopped the vegetables or simmered the broth. You just need to know which tray best suits your hunger (your use-case, like chatbots or image creation).
If you want, you can customise your portion — ask for extra spice (fine-tune the model with your own data) or request a specific garnish (add a knowledge base). The buffet line never runs out because the kitchen keeps restocking from a giant fridge (AWS's infrastructure). And you only pay for what you scoop onto your plate. Amazon Bedrock is that buffet line: a managed service that gives you access to multiple pre-built AI models without the hassle of building, hosting, or maintaining them yourself. The Holiday Feast Hall is the AWS cloud, and the chefs are the model providers.
Let's start with the most basic question: what is Amazon Bedrock? It's a cloud service offered by Amazon Web Services (AWS) that acts as a single, central place where you can discover, test, and use a variety of Foundation Models (FMs). A Foundation Model is a huge, general-purpose AI model that has been pre-trained on a massive amount of text, images, or code. Think of it as a brain that already knows a lot about language, concepts, and patterns. Companies like Anthropic build models such as Claude, Meta builds Llama, and Stability AI builds models for generating images. Bedrock gives you access to all these different brains through one common interface.
Why does this service exist? Before Bedrock, if you wanted to use a powerful AI model, you had two difficult options. First, you could build and train your own model from scratch, which requires an enormous amount of data, specialised engineering talent, and expensive computing power — something only the biggest tech companies could afford. Second, you could connect directly to a model provider's API, but then you'd have to manage security, scaling, and costs yourself for each different provider. Bedrock solves both problems. AWS hosts the models on its own servers, handles the infrastructure, and provides a standard way to interact with all of them. This means you can switch between different models without rewriting your application.
So how does it work in practice? You start by logging into the AWS Management Console (a web dashboard for managing AWS services). From the Bedrock section, you can browse a catalogue of available models. Each model has different strengths: some are better at conversation, some at summarising long documents, some at generating images from text descriptions. Once you find a model you like, you can test it directly in the console using a 'playground' — a chat interface where you type prompts and see the model's response immediately. This lets you experiment without writing any code.
When you're ready to use the model in a real application, you integrate Bedrock using an API, which is a set of rules that allows one piece of software to talk to another. Bedrock offers a single API that works the same way no matter which underlying model you choose. You send a request to the API with your input text (called a prompt) and some settings (like how creative you want the response to be, controlled by a parameter called temperature). Bedrock forwards your request to the chosen model and returns the generated output to you. You don't need to know where the model is running or how it's scaled to handle millions of users.
One of the most powerful features of Bedrock is the ability to customise a model without retraining it from scratch. This is called fine-tuning. You can take a pre-trained model and give it additional training on your own private data, like your own company's emails, product descriptions, or customer service transcripts. This makes the model better at tasks specific to your business, while still keeping the general knowledge it already has. Bedrock also supports Retrieval Augmented Generation (RAG), a technique where the model accesses a separate knowledge base (like your company's documents) when answering a question, allowing it to give answers based on information you provide without needing to retrain the model.
Finally, Bedrock handles all the security and access control. You use AWS Identity and Access Management (IAM) to decide exactly which people or applications in your organisation are allowed to use which models. All data sent to Bedrock is encrypted (scrambled into a secret code) both when it's travelling across the internet and when it's stored. AWS also promises that your data is never used to improve the underlying models for other customers — your private data stays private. This combination of access, customisation, security, and simplicity makes Bedrock a foundational tool for any business wanting to use AI in a safe, scalable, and cost-effective way.
Model Discovery and Selection
You log into the AWS Management Console, navigate to Amazon Bedrock, and browse the model catalogue. You review the descriptions, capabilities, and pricing for each Foundation Model. This step is where you decide which model (e.g., Anthropic Claude vs. Meta Llama) best fits your use case, such as conversation, summarisation, or image generation.
Testing in the Playground
Before committing to a model, you use the Bedrock playground to type sample prompts and see real responses. This lets you evaluate the model's tone, accuracy, and behaviour. You can adjust parameters like 'temperature' (creativity) and 'max tokens' (response length) to see how they affect output. This step is critical for avoiding costly mistakes in production.
Customisation (Fine-tuning or RAG)
If the base model's performance is not good enough for your specific task, you can customise it. For fine-tuning, you upload a small dataset of labelled examples, and Bedrock trains the model further on that data. For RAG, you create a Knowledge Base by uploading your documents to Amazon S3, then link that knowledge base to the model so it can retrieve facts during inference.
Integration via API
You write your application (e.g., a chatbot or content generator) to call the Bedrock API. You use the AWS SDKs (pre-built code libraries) to send prompts and receive responses. This step involves configuring IAM roles to give your application permission to invoke the specific model. You also decide on the runtime settings, such as how many tokens to generate per request.
Monitoring and Cost Management
Once your application is live, you monitor usage through AWS CloudWatch (a monitoring service) and AWS Cost Explorer to track how many tokens are being processed and how much it costs. You can set budgets and alerts to avoid unexpected charges. This step ensures the service remains performant and cost-effective as usage scales.
Let's walk through a realistic scenario: a mid-sized online retail company called 'ShopSmart' wants to add an AI-powered chatbot to help customers track orders, handle returns, and answer product questions. The IT team at ShopSmart has never built an AI model before, and they don't have a data science team. They are AWS customers and need a solution that can be live within a few weeks.
The IT lead, Priya, starts by logging into the AWS Management Console and opening Amazon Bedrock. She browses the model catalogue and tests three different models in the playground: Anthropic's Claude for conversation, Meta's Llama for text summarisation, and Amazon's Titan for question-answering. After testing, she finds that Claude gives the most natural-sounding customer service responses. She selects Claude as the base model for the chatbot.
Next, Priya customises the model. ShopSmart has a database of thousands of past customer service conversations and a list of all their current products. She uses Bedrock's RAG feature (Retrieval Augmented Generation) to connect the model to this internal knowledge base. Now when a customer asks 'What is the return policy for shoes?', the model doesn't guess — it pulls the exact policy from ShopSmart's own documents and generates a response based on that real data. This means the model always gives accurate, up-to-date information.
To make the model even better, Priya performs a small fine-tuning job. She takes a subset of the best past conversations and uses Bedrock's fine-tuning tool to train Claude on ShopSmart's specific tone and common scenarios. The fine-tuning takes a few hours and costs a few hundred dollars — a fraction of what it would cost to train a model from scratch. Now the chatbot not only knows the facts, but also uses the same polite, helpful language that ShopSmart's best human agents use.
Priya then builds a simple web application using AWS Lambda (a serverless compute service that runs code on demand) and connects it to Bedrock's API. She uses AWS IAM to set permissions so that only the chatbot application can invoke the model, and she enables logging to track every request for auditing purposes. When a customer types a message, here is the step-by-step flow:
The customer types 'Where is my order?' in the chatbot on ShopSmart's website.
The chatbot application sends this question as a prompt to the Bedrock API.
Bedrock retrieves the prompt, checks IAM permissions, and forwards it to the fine-tuned Claude model.
The model, also using RAG, looks up the customer's order from ShopSmart's database (connected via Bedrock Knowledge Bases).
Claude generates a response: 'Your order #12345 is out for delivery and should arrive by 5 PM today.'
Bedrock sends this response back to the chatbot application, which displays it to the customer.
The entire process takes a few seconds. Priya monitors the cost using AWS Cost Explorer; she pays only for the number of tokens (pieces of words) processed by the model. If usage grows, Bedrock automatically scales up to handle thousands of simultaneous conversations without Priya having to manage servers. The project goes live in three weeks, and ShopSmart sees a 40% reduction in customer service call volume. This scenario shows exactly what an IT professional does with Bedrock: they browse, test, customise, integrate, secure, and monitor — all without ever building a model from scratch.
The AIF-C01 exam tests your understanding of Amazon Bedrock in several specific ways. First, you need to know that Bedrock is a fully managed service that provides access to Foundation Models (FMs) from multiple providers, not just Amazon's own models. A common exam trap is a question that implies Bedrock only contains Amazon's Titan models. The correct answer will always mention that you can also access models from Anthropic, Meta, Stability AI, and other third-party providers. Memorise this list: Anthropic (Claude), Meta (Llama), Stability AI (Stable Diffusion for images), and Amazon (Titan).
Second, the exam tests your understanding of the difference between using a model 'as-is' (called Inference) versus customising it. Inference is sending a prompt to a pre-trained model and getting a response. Customisation includes fine-tuning (further training the model on your data) and RAG (connecting the model to a knowledge base to retrieve facts during a query). The exam will try to confuse you by mixing these terms — for example, a question might ask 'Which method allows a model to use your company's documents without retraining?' The correct answer is RAG, not fine-tuning.
Third, the exam focuses on security and data privacy. Key facts to remember: data is encrypted in transit and at rest. Your data is not used to train the underlying models for other customers. Access control is managed through AWS IAM — you assign permissions to users or applications to invoke specific models. A trap question might say 'Data sent to Bedrock is used to improve the model for everyone.' That is false; AWS explicitly states that your data stays private to your account.
Fourth, be prepared for questions about the Bedrock Knowledge Bases feature. This is how RAG is implemented. A knowledge base is a collection of documents (stored in Amazon S3, a cloud storage service) that the model can query when generating responses. The exam might ask what service you would use to store the source documents — the answer is Amazon S3.
Key definitions to memorise for the exam:
Foundation Model (FM): a large, pre-trained AI model.
Amazon Bedrock: a managed service to access and deploy FMs.
Inference: using a model to generate output from input.
Fine-tuning: training a pre-trained model further on custom data.
RAG (Retrieval Augmented Generation): augmenting a model's response with real-time data retrieval from a knowledge base.
Token: a unit of text (e.g., a word or part of a word) used for billing and processing.
Finally, the exam may ask about the 'playground' feature. This is the console interface where you test models before deploying them. It's important for experimentation and evaluation. A question might say 'How would you quickly compare two models to see which one gives better responses?' The answer is to use the Bedrock playground. No inference is needed — the playground is built into the management console. Understanding these exact concepts and the common traps will dramatically increase your chances of answering correctly on exam day.
Amazon Bedrock is a managed service that gives you access to multiple Foundation Models from different providers through a single API.
You can use pre-trained models as-is for inference, or customise them with fine-tuning and RAG without training from scratch.
Your data sent to Bedrock is encrypted and is not used to train the underlying models for other customers.
Access to Bedrock models is controlled using AWS IAM policies, allowing you to restrict which users or applications can invoke each model.
The Bedrock playground is a no-code interface for testing and comparing models directly in the AWS Management Console.
Retrieval Augmented Generation (RAG) connects a model to your own knowledge base so it can answer questions using your data without retraining.
These come up on the exam all the time. Here's how to tell them apart.
Amazon Bedrock
Provides pre-trained Foundation Models from multiple providers that you use as-is or customise.
Requires minimal machine learning expertise to get started.
Managed service that handles infrastructure scaling automatically.
Amazon SageMaker
A platform for building, training, and deploying your own machine learning models from scratch.
Requires strong data science and machine learning skills to use effectively.
You have full control over the training infrastructure, including GPU instances.
Fine-tuning
Changes the model's internal parameters by training on your custom dataset.
Requires a dataset of labelled examples (pairs of input and desired output).
The model becomes specialised for your task but cannot access new information after training.
Retrieval Augmented Generation (RAG)
Does not change the model's internal parameters; it adds a retrieval step to fetch external information.
Requires a knowledge base (e.g., documents in Amazon S3) but does not need labelled examples.
The model can access real-time or updated information without needing retraining.
Inference
Using a model to generate output from an input prompt (e.g., answering a question).
Does not change the model's knowledge or parameters.
Typically costs less per token than training.
Training
The process of teaching a model from data, adjusting billions of internal parameters.
Permanently changes the model's behaviour and knowledge.
Requires significant computational resources and time, costing much more than inference.
Mistake
Amazon Bedrock only gives you access to Amazon's own Titan models.
Correct
Amazon Bedrock is a marketplace that includes models from multiple providers, including Anthropic (Claude), Meta (Llama), Stability AI (Stable Diffusion), and Amazon Titan.
The name 'Amazon Bedrock' sounds like it's an Amazon-only product, leading beginners to assume it only features Amazon's own creations.
Mistake
Using Bedrock means you have to train a new model from scratch for every use case.
Correct
Bedrock gives you access to pre-trained models. You can use them as-is or optionally customise them with fine-tuning or RAG, but you never need to train from scratch.
Many beginners think 'AI service' means they must build a model themselves, not realising that Bedrock is about consuming ready-made intelligence.
Mistake
Fine-tuning and Retrieval Augmented Generation (RAG) are the same thing.
Correct
Fine-tuning changes the model's internal parameters by training it on your data, while RAG leaves the model unchanged and adds a retrieval step to pull external information into the prompt.
Both are customisation techniques that improve model accuracy, so beginners easily conflate them without understanding the fundamental difference in how they work.
Mistake
You must write complex code in Python or another language just to test a model in Bedrock.
Correct
Bedrock provides a visual 'playground' in the AWS Management Console where you can type prompts and see responses without writing any code.
Many people associate AI with programming, so they assume all interaction requires code, overlooking the user-friendly testing interface designed for non-developers.
Reveal each answer, then mark whether you got it right. Score 60%+ to unlock the next chapter.
No. Bedrock is designed for developers and IT professionals who have general cloud skills but may not have deep machine learning expertise. The playground lets you test models without coding, and the API is standardised.
No. Bedrock is for accessing and customising existing Foundation Models, not for training new ones from scratch. If you need to train a custom model from the ground up, you would use Amazon SageMaker instead.
Yes. Data is encrypted in transit and at rest. AWS explicitly states that your data is not used to improve the underlying models for other customers, and you retain full ownership of your data.
Amazon Bedrock gives you ready-to-use Foundation Models via an API with optional customisation. Amazon SageMaker is a platform for building, training, and deploying your own machine learning models from scratch or from templates.
No. You can use multiple models from different providers for different tasks within the same AWS account. For example, you could use Anthropic Claude for chatting and Stability AI for image generation.
You are charged based on the number of tokens processed (both input and output), plus any additional costs for fine-tuning or storing your Knowledge Base. Each model has its own per-token pricing, which you can see in the Bedrock pricing page.
You've finished Amazon Bedrock Fundamentals. Continue through the AIF-C01 study guide to build a complete picture of the exam.
Done with this chapter?