🛠️ Build an AI Support Assistant with Amazon Bedrock

Build an AI Support Assistant with Amazon Bedrock by testing prompts with Nova Lite and invoking the model through Python.

Series: 🚀 AI on AWS: From Zero to Building (7 articles)

1
🚀 New to AI on AWS? Start Here: A Beginner’s Roadmap to Artificial Intelligence
…
3
🛠️ Build an AI Support Assistant with Amazon Bedrock This article
4
🧠 Give Your AI Its Own Knowledge: RAG with Amazon Bedrock

Swarali Chine

Published Sep 29, 2026

Last modified Sep 29, 2026

In the first article, we learned the basics of AI and generative AI. Then, we explored the main AI services available on AWS.

Now, it is time to build something.

In this guide, we will create a simple AI Cloud Support Assistant using Amazon Bedrock and Amazon Nova Lite.

The assistant will take an AWS problem and return possible causes and troubleshooting steps.

What Are We Building?

Suppose a user asks:

My application receives an AccessDenied error when uploading a file to Amazon S3.

Our assistant will provide:

A summary of the problem
Possible causes
Troubleshooting steps
AWS settings to inspect
A security reminder

The application flow is simple:

1
User → Python application → Amazon Bedrock → Amazon Nova Lite → Response

Amazon Bedrock manages the foundation model for us, so we do not need to deploy or maintain any model infrastructure.

Step 1: Open Amazon Bedrock

1
Amazon Bedrock

Select this Region:

1
US East (N. Virginia) — us-east-1

Model availability can vary by Region, so we will use us-east-1 throughout this project.

Step 2: Select Amazon Nova Lite

From the left menu:

Open Model catalog
Select Amazon as the provider
Search for Nova Lite
Open the model

Nova Lite is a fast, low-cost model suitable for simple text-generation applications.

We do not always need the largest model. We need a model that can perform our task at an acceptable cost and speed.

Step 3: Test a Basic Prompt

Open Playgrounds → Chat/Text.

Choose Select model, select Amazon Nova Lite, and then choose Apply.

Enter this prompt:

1
2
My application receives an AccessDenied error when uploading
an object to Amazon S3. What could be causing it?

Choose Run.

The model may identify causes such as:

Missing s3:PutObject permission
An explicit deny in an IAM or bucket policy
An incorrect S3 object ARN
Missing AWS KMS permissions
A restrictive service control policy

The answer may be useful, but its structure might not always be consistent.

Step 4: Improve the Prompt

Add this system instruction:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
You are an AWS Cloud Support Assistant.

For every issue:

1. Summarize the problem.
2. List the most likely causes.
3. Provide numbered troubleshooting steps.
4. Mention the AWS settings that should be inspected.
5. Include a security note when relevant.
6. Ask for more information if the cause cannot be confirmed.

Do not present assumptions as confirmed facts.
Do not invent AWS commands or features.
Never ask the user to share credentials.

Run the same S3 question again.

The response should now be clearer and easier to follow.

The system instruction gives the model:

A role
A specific task
A response format
Safety boundaries

Set the inference values to:

1
2
3
Temperature: 0.2
Top P: 0.9
Maximum output tokens: 600

A low temperature helps produce more focused responses, which is useful for technical troubleshooting.

Step 5: Test More Questions

Try a few different AWS problems:

1
I cannot connect to my EC2 instance using SSH.

1
My Lambda function times out while connecting to Amazon RDS.

1
2
My application receives AccessDeniedException when invoking
an Amazon Bedrock model.

Check whether the assistant:

Follows the requested format
Gives reasonable troubleshooting steps
Separates possible causes from confirmed facts
Requests missing information
Avoids asking for credentials

Testing different questions helps us identify weaknesses before placing the model inside an application.

Step 6: Build the Python Application

Open AWS CloudShell and create a file:

1
nano support_assistant.py

Paste this code:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
import boto3
from botocore.exceptions import ClientError

MODEL_ID = "amazon.nova-lite-v1:0"

SYSTEM_PROMPT = """
You are an AWS Cloud Support Assistant.

For every issue:
1. Summarize the problem.
2. List the most likely causes.
3. Provide numbered troubleshooting steps.
4. Mention the AWS settings to inspect.
5. Include a security note when relevant.

Do not present assumptions as confirmed facts.
Never ask the user to share credentials.
"""

bedrock = boto3.client(
    "bedrock-runtime",
    region_name="us-east-1"
)

def troubleshoot(issue):
    response = bedrock.converse(
        modelId=MODEL_ID,
        system=[{"text": SYSTEM_PROMPT}],
        messages=[{
            "role": "user",
            "content": [{"text": issue}]
        }],
        inferenceConfig={
            "temperature": 0.2,
            "topP": 0.9,
            "maxTokens": 600
        }
    )

    return response["output"]["message"]["content"][0]["text"]

try:
    issue = input("Describe your AWS issue: ")
    print("\n" + troubleshoot(issue))

except ClientError as error:
    print(error.response["Error"]["Message"])

Save the file and run it:

1
python3 support_assistant.py

Enter an AWS issue when prompted.

The application sends the question to Amazon Bedrock, Nova Lite generates the response, and Python displays the result.

What Is Happening Behind the Scenes?

When we enter a question:

Python receives the question.
Boto3 sends it to Amazon Bedrock.
IAM verifies permission to invoke the model.
Nova Lite processes the system instruction and question.
Bedrock returns the generated response.
Python displays it.

The application needs this IAM permission:

1
bedrock:InvokeModel

Important Limitations

Our assistant cannot see the user’s actual:

IAM policies
S3 bucket policies
CloudWatch logs
VPC configuration
CloudTrail events
AWS resources

It can suggest possible causes, but it cannot confirm the root cause without evidence.

1
AI guidance is not the same as a verified diagnosis.

Always review the response before making changes to an AWS environment.

Security and Cost

For a real application:

Use IAM roles instead of hard-coded credentials
Follow least privilege
Never include passwords or access keys in prompts
Avoid sending sensitive customer data
Protect stored prompts and responses
Validate recommendations before executing them

Amazon Bedrock uses usage-based pricing. The cost depends mainly on the selected model and the number of input and output tokens.

This project should remain inexpensive because we are using Nova Lite, short prompts, and limited responses.

Common Errors

`AccessDeniedException`

Confirm that your IAM identity has:

1
bedrock:InvokeModel

Invalid model identifier

Confirm the model ID and selected Region.

On-demand throughput is not supported

Try the cross-Region inference-profile ID:

1
MODEL_ID = "us.amazon.nova-lite-v1:0"

The `converse` operation is unavailable

Upgrade Boto3:

1
pip3 install --user --upgrade boto3

What We Learned

In this project, we:

Selected a foundation model
Tested it in the Bedrock playground
Improved its response with a system instruction
Configured inference settings
Tested multiple support questions
Invoked the model from Python
Considered security, permissions, and cost

We have moved from learning about AI services to building with them.

Our assistant can answer general AWS questions, but it does not yet know our private documentation or support procedures.

In the next article, we will add our own trusted knowledge using Retrieval Augmented Generation, or RAG.

Our journey now looks like this:

1
Learn → Explore → Build → Add Knowledge

Let’s keep building.

AWS References

# amazon-bedrock # gen-ai # prompt-engineering

Series: 🚀 AI on AWS: From Zero to Building (7 articles)

1
🚀 New to AI on AWS? Start Here: A Beginner’s Roadmap to Artificial Intelligence
…
3
🛠️ Build an AI Support Assistant with Amazon Bedrock This article
4
🧠 Give Your AI Its Own Knowledge: RAG with Amazon Bedrock

Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.

Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article

Builder Center

🛠️ Build an AI Support Assistant with Amazon Bedrock

Series: 🚀 AI on AWS: From Zero to Building (7 articles)

Swarali Chine

What Are We Building?

Step 1: Open Amazon Bedrock

Step 2: Select Amazon Nova Lite

Step 3: Test a Basic Prompt

Step 4: Improve the Prompt

Step 5: Test More Questions

Step 6: Build the Python Application

What Is Happening Behind the Scenes?

Important Limitations

Security and Cost

Common Errors

AccessDeniedException

Invalid model identifier

On-demand throughput is not supported

The converse operation is unavailable

What We Learned

AWS References

Series: 🚀 AI on AWS: From Zero to Building (7 articles)

`AccessDeniedException`

The `converse` operation is unavailable