Sep 5, 2025
AI agents: How complex and costly is the development?
From idea to implementation with Amazon Bedrock and AgentCore.
At the end, a practical example of an agent that responds via WhatsApp.
AI adoption is now a strategic priority for large enterprises, but the path forward is often unclear. Information about how to develop agents, implement them, and understand the associated costs remains confusing, which slows progress.
The key lies in understanding what’s required: data quality, integration with existing systems, and features that ensure security and scalability. This article explains those elements and how Amazon Bedrock and AgentCore can simplify the journey toward automation.
Amazon Bedrock and AgentCore: Leveraging Large Language Models Securely
Can you imagine ChatGPT responding with your company’s data? That’s exactly what Amazon Bedrock enables. It connects to large language models (LLMs) and allows your agent to reason through them securely. Your data never leaves the protected AWS environment and is not used to train the LLMs.
If Bedrock is the brain that thinks, AgentCore is the structure that brings order. Thanks to AgentCore, the agent stays focused on relevant data and applies clear rules about who can access what information. For example, a user from the Risk department won’t be able to query the CEO’s salary, while the Accounting team will have access to that information to fulfill their responsibilities.
Now you may say: Sounds great. But, what do I need to get started?
Before implementing AI agents, companies need to establish a few foundational elements:
- Structured databases: Information must be organized in a way that machines can understand. A structured database contains clear tables, labeled columns, and clean data. Without this order, AI cannot deliver reliable results.
- An AWS account with access to Bedrock: This is the platform where the model runs.
- Clear requirements, security rules, and data governance: Define what the agent is meant to do, what internal data it will access, who can access it, and how usage will be monitored.
- Impact on human roles and adoption practices: Consider how the agent will affect existing roles and what strategies will support its effective adoption.
These elements are essential for accurately estimating development effort, implementation timelines, and training costs.
And the development effort?
The effort required varies depending on the role of the agent you want to implement:
- Junior Agent
Responds to basic questions through a web or intranet channel (e.g., “How many credit requests were approved last month?”).
Translates the query into SQL over a single main table with up to 10 fields.
Estimated development time: 80–120 expert hours.
Limited to simple questions over a subset of data and lacks advanced security handling.
- Senior Agent
Supports more varied queries via intranet or internal email, accessing 3–5 tables with dozens of fields.
Integrates AgentCore for role-based security and permissions, and can retain memory of previous queries.
Estimated development time: 200–300 expert hours.
Capable of processing more data and responding to more complicated questions. Requires well-structured databases and data modelling.
- Executive Agent
Connects multiple channels (e.g., WhatsApp, Teams, internal apps) and accesses 10+ tables or multiple databases.
Applies enterprise-level governance (defining which data is visible to each role), observability, and interaction logging.
Estimated development time: 400–600 expert hours.
High strategic value, easy to use and with clearly defined roles.
It’s important to note that agents do not clean chaotic data. They depend on the quality of your databases, and each query generates usage costs.
A Practical Example: Banking Agent Responding via WhatsApp
To close, we present a practical example: a conversational agent that allows bank customers to make inquiries directly through WhatsApp, accessing internal company data.
This agent can answer common questions such as:
- When is my next loan payment due?
- Where is the nearest branch?
Here’s how the flow works: the customer sends a message via WhatsApp, which reaches the system through a Webhook connected to Meta’s API. The message is received by an API Gateway, which routes it to a Lambda adapter. From there, the agent’s core (AgentCore) interprets the request, validates permissions, and decides which tool to use to execute the query.
For example, if transactional data is needed, a tool executes a pre-approved SQL query on Redshift. The result is summarized in natural language using Bedrock, and the response is sent back to the customer via WhatsApp. If the query requires additional information, the agent can include a link to a full report stored in S3. The system can also log interactions in DynamoDB and monitor agent behavior using CloudWatch.
Do you want our consultants to help you with this?
Reach out to us.
Felipe Uribe Velásquez
Director
Camilo Monsalve Maya
Data Engineering Consultant
