For those operating in the Microsoft ecosystem, for a fairly low monthly fee, you can build an extremely powerful, flexible army of AI agents to do your bidding. Microsoft Copilot Studio can help you automate those tedious, repetitive tasks that eat up your workday. Today, we’ll be taking a look at an agent I built for one of my clients that completely revolutionized the way they handle and manage their RFP responses.

Before we get started, here’s the tech you’ll need to acquire (you may already have these):

  • Dataverse Environment
  • Copilot License (not the free one)
  • Copilot Studio
  • Pay-as-you-go billing

Once these are settled, you can begin with the prep work.

Prep Your RFP Knowledge Base

If you don’t already have a database to store your RFP responses, then chances are you have them scattered across PDFs, Word Docs, and Spreadsheets. While you can build your bot’s knowledge with these document types, it can create confusion. If you haven’t been tracking responses then you may have contradictory information within your dataset. This will result in your AI agent producing inconsistent or conflicting outputs.

Here are my recommendations for organizing data prior to feeding the bot.

  • Use Excel: I know it's a pain to extract data from Word docs and PDFs but for the best results, you’ll want to give your agent structured data.
  • Break up responses into categories: compliance, operations, security requirements, etc.
  • Put each category into its own sheet.

With your data files cleaned and prepped, it’s time to make your agent smart.

Note: If you prefer more dynamic data sources, your SharePoint library can also be incorporated into the knowledge base.

Build Your Copilot Studio Knowledge Base

Within your bot’s configuration, there is a Knowledge section that lets you configure the data sources your bot can reference when answering questions and performing tasks. This is where you’ll upload the Excel files you created during the knowledge base prep phase.

Don’t be alarmed if uploading takes a while to complete. Even small files (less than 100 KB) take minutes to index after being uploaded. Keep in mind that Copilot, while useful, does not “learn” anything the way a trained LLM does. It’s more of a sophisticated retrieval system.

Here’s what’s happening behind the scenes.

  1. Copilot chunks the documents
  2. Chunks are converted into embeddings
  3. These embeddings are stored in a vector index
  4. At runtime the model retrieves the most relevant chunks
  5. The LLM generates an answer using those retrieved chunks

This entire process is known as RAG (Retrieval Augmented Generation). The model itself never changes even if the answers vary slightly. For use cases like this, the agent being untrainable is actually an advantage.

If the information you have changes, such as a policy shift or new regulation you have to follow, you can simply replace the old data. If you were working with a trainable LLM, then you would need to retrain the model whenever new information was introduced.

Install Guardrails for Your Agent

To ensure the most accurate answers possible, you'll want to give your agent some guidelines. Think of it as a job aid for a human. Use these instructions to get started:

  • Only answer questions using the knowledge sources provided. Search elsewhere only when explicitly requested.
  • If an answer cannot be found, respond with “This information is currently unavailable in my knowledge base”.
  • Please refrain from guessing or inventing answers.
  • If multiple answers exist for the same topic, prefer the most recent entry.
  • If a user states the information is incorrect, attempt to read the knowledge base again. If the user still states it is incorrect, instruct them to contact aihelp@mattflows.com.

You’ll need to tweak and refine these instructions as you test and receive feedback from your users. The goal here isn’t to capture every possible scenario, but rather to quickly build a competent assistant that removes friction and tedious tasks from your business workflows.

Pro Tip: If you choose to allow your bot to use the Internet to search for answers, you’ll need to toggle the Web Search option to “Enabled”.

Test Your Agent

Before hitting the publish button, it’s time to put your agent to the test. Simulate common questions users will ask and some intentionally vague ones. This will help confirm that your instructions and knowledge sources are working as intended.

Start with simple questions:

  • Does your platform support SSO (aka single sign-on)?
  • Is your organization ISO 27001 certified?
  • What compliance certifications do you maintain?

Then try broader questions:

  • How does your organization protect users from cyber attacks?
  • How does your organization secure customer data?
  • What steps has your organization taken in the last 12 months to address AI-based cyber attack risks?

Finally, test questions that are outside the scope of your knowledge base to ensure the fallback responses work correctly.

  • Who is your Cyber Security Officer?
  • Does your organization use Amazon Web Services?
  • How does your organization protect against SQL injection attacks?
  • Does your organization offer an iOS or Android mobile application?

Adding Triggers to Incorporate Your Agent Into Your Workflow

This step is optional, but if you're like me and want to squeeze every ounce of value out of your tools, this is where things get exciting.

You’ll need:

  • Power Automate
  • A Microsoft email account

If you’re unfamiliar with Power Automate, it’s Microsoft’s automation platform. I use it almost daily to automate processes in both Google and Microsoft environments.

Here’s a simple workflow that can save countless hours:

  1. A new email arrives at rfp@mattflows.com
  2. Power Automate extracts the document or email content
  3. The extracted data is passed to the Copilot agent
  4. The agent parses the document and retrieves answers from the knowledge base
  5. A draft response is generated for human review

Let the Next Employee You Onboard be an AI Agent

AI agents are here to stay. To stay competitive, your business should incorporate them into workflows where they make sense. Start with the most tedious, time-consuming tasks that don’t require creativity or complex judgment. In my client’s case, that task was RFP responses.

Follow this guide to build your own AI agent and reclaim hundreds of valuable hours in your workday. If you're looking to check out Google Workspace Studio's automation capabilities, here's a simple but powerful email automation you can try. And, if you ever get stuck, want to collaborate, or just want to say hello, I’m only a DM away on the platforms listed below.

Until next time, happy automating!