AI Agent Architect for FinOps

AI Agent Architect with expertise in reviewing document-driven projects, assessing and enforcing cloud best practices (across Microsoft Azure, Amazon AWS, Oracle Cloud and Google Cloud), and producing accurate, FinOps-aligned cost estimates to optimize cloud spend and governance.

At the Microsoft Innovation Hub in São Paulo, we receive many requests related to this scenario, and after discussing it with Rafa Morales (Solution Engineer at Microsoft), we agreed that Copilot Studio is a strong solution to address it. We have been conducting many engagements over the past few months related to this scenario, and we hope that this solution can help you.

The functionality for drawing diagrams and topologies may vary according to the evolution of LLMs. This AI agent will use a separate, dedicated agent for generating these diagrams, leveraging Claude Sonnet 4.5 (Experimental), with output preferably produced in Draw.io and Mermaid.

Why an AI Agent Architect?

Many customers are multi‑cloud today, and IT teams face several challenges:

• Growing complexity: Multi‑cloud architectures are becoming more complex every day. It’s difficult to design the right solutions without strong, hands‑on expertise.
• Cost estimation: Estimating costs across multiple projects is tough but essential to build a solid FinOps strategy.
• Time‑consuming reviews: Analyzing projects for best practices and cost optimization takes significant time, and you need to accelerate this work to keep up with IT demand.

Recent AI models (especially since GPT‑3.5) have greatly improved the analysis process with high accuracy. This setup references GPT-5 Chat, but you can easily switch to newer models as they become available. Another area that has advanced significantly is diagram generation — with better support for C4, flowcharts, and similar notations.

Special thanks to Rafa Morales, Solution Engineer at the Microsoft Innovation Hub Sao Paulo, who supported me with tips and best practices for developing an Agent using Microsoft Copilot Studio.

https://www.linkedin.com/in/rafamoralesms/

This solution was developed using Microsoft Copilot Studio but it's possible to use Copilot Agent, without coding. The project using Copilot Chat is available bellow:

AI Agent Architect - Lite version using Microsoft Copilot Chat

Link

Import this solution

If you already have Copilot Studio, you can download the solution, make your own adjustments, and deploy it in your environment. Follow the steps below to learn how to do this:

WARNING: THIS IS A SAMPLE PROJECT, WITHOUT ANY OFFICIAL SUPPORT

Link About How to Import the Solution on Copilot Studio

Download the AI Agent Architect Solution

The steps bellow have instruction to deploy step-by-step.

Sample projects

These are some sample projects to test how the AI Agent Architect works. Besides the fact that the documents are in Portuguese there is no problem, even if you are using prompts in English.

WARNING: THESE ARE SAMPLE PROJECTS, JUST FOR TESTS / REFERENCE, NOT OFFICIAL DOCUMENTS OR PROJECTS.

Sample - Microsoft Azure

Sample - Amazon AWS

Sample - Vmware

How the Agent works for this scenario

Normally, IT teams prepare documents that describe the architecture or planned deployments in formats like *.DOCX or *.PDF. This solution can extract information from the text inside these documents, and it can also analyze diagrams if they are included (for example, a proposed architecture or change diagram). Other file types that can be analyzed (with different output quality):

• Terraform (.tf): If you submit.tf files, the output will be more precise, because Terraform requires you to specify the exact instance sizes and service configurations.
• YAML (Kubernetes scenarios): The solution can provide best practices according to the cloud provider, but it cannot estimate costs unless you include a pricing reference file in the internal knowledge base.
• Diagrams (JPEG, PNG, etc.): If the image clearly specifies all parameters (instance sizes, regions, tiers, etc.), the results can be almost as accurate. This is useful for customers who prefer to draw the full solution visually.

If the document includes estimated volumetry (for example, users per day, requests per second, data size, retention time), the architecture and cost estimates will be more accurate. It can be difficult to provide these numbers at the start of a project, but they lead to much better results.

This solution is a good example of how to use a multi‑agent architecture to solve the problem, using Microsoft Copilot Studio.

1. User submission → AI Agent Architect (Copilot Studio)

• Supported file types: DOCX, PDF, Terraform, PNG, YAML
• Action: Parse, extract intent (providers, CAF/WAF hints, pricing clues, diagrams).

2. Analyze & route to up to four cloud agents:

• Microsoft Azure Agent (via MCP server of public docs)
• Amazon AWS Agent (via MCP server of public docs)
• Oracle Cloud OCI Agent (via public website — CAF/WAF)
• Google Cloud GCP Agent (via public website — CAF/WAF)

Microsoft and Amazon offer MCP servers for their technical libraries for free, and you can connect to them at no cost. As of the date of this publication, I have not found MCP servers from Google Cloud or Oracle Cloud (specifically for technical libraries, not for their cloud services). As a workaround, I am connecting to the publicly available technical documentation on their websites. It is not required to use MCP servers to access technical documentation.

This demo connects to the CAF (Cloud Adoption Framework) and WAF (Well‑Architected Framework), which are the recommended reference frameworks from these cloud providers.

3. Each agent performs:

• Best practice checks: CAF/WAF, security, resiliency, operations.

You can add, for each agent, a technical library or documentation containing the specific patterns your company follows. For example, in the Azure agent, you can include a document outlining cloud patterns (e.g., “Every VM must use Premium Disks”). Since each cloud provider used by your company may have its own standards and best practices, you can attach the relevant documents to the corresponding agent.

4. Cost estimation:

• If the document contains a pricing list → estimate from it. Else → pull public pricing to estimate.

Many enterprises with large cloud contracts receive special discounts for certain services, instance types, regions, and more. This repository does not include example pricing files. However, you can upload your invoice or spreadsheet with your custom pricing as Knowledge in Copilot Studio for each Agent (see details below). Doing this allows the Agents to reference your negotiated rates when estimating costs—so results better reflect your real‑world pricing.

5. Draw.IO, Mermaid or C4 diagramming:

• Every agent can emit diagrams. For multi‑cloud documents, multiple agents run in parallel and their models are aggregated.

6. Final results back to the user:

• Best‑practices aligned with each provider’s CAF/WAF
• C4 diagrams
• Estimated costs

Using Copilot Studio with Azure Credits

This solution use Microsoft Copilot Studio as the platform. I strongly recommend because it enables to use advanced features (connecting to MCP servers, publish on Microsoft Teams, etc) but it's possible to use Microsoft Copilot Chat (but with limitations). If you have a Microsoft Azure Subscription then it's possible to consume Azure credits for Copilot Studio. Please check the links bellow for more information:

• Copilot Studio - Pay-as-you-go pricing
• Manage Copilot Studio credits and capacity

Step 0 - Get Access to Copilot Studio

If you never tried Copilot Studio before then I recommend to follow these steps described here:

• Get access to Copilot Studio

Step 1 - Create your agent

1. Open your browser and access Copilot Studio using https://copilotstudio.microsoft.com

2. On the left side, click Agents, and then on the right side, click + Create blank agent . This will start the wizard process.

3. In Details, click the Edit button and provide a Name for your agent (e.g., AI Agent Architect) and a Description. You can also upload an image file for the Icon.

4. Select the agent model. GPT-5 Chat offer great results and explanation but you can try other newest model. Reasoning models could generate more data but can cause time-out.

5. Define the Instructions. This is the most important part of the agent. The text below is a sample I am currently using for this Agent, but you can adjust it according to your needs. This agent is configured to respond in English by default, but it can answer in Portuguese, Spanish, or any other language—simply instruct the Agent accordingly. Also, don’t forget to update the Cloud Region that the agent should use as the reference for cost analysis.

Bellow is a sample for Instructions:

---

Analyze all attached documents that describe cloud system architectures and generate a complete technical report according to the instructions below. All responses must be in formal technical English. 

1. Architecture Ide