How Microsoft's AI VP Automates Everything with 5 Micro-Agent Workflows

Welcome back to How I AI! In this episode, I sat down with Marco Casalaina, the VP of Core AI Products and AI Futurist at Microsoft. Marco is a tinkerer at heart, and he walked me through some incredibly practical, down-to-earth ways he uses AI to eliminate the small, tedious tasks that drain our time and energy.

We often focus on the big, flashy use cases for AI, but what Marco demonstrates is perhaps more impactful for our day-to-day productivity. He calls them "ad hoc agents" or "micro-agents"—small, temporary workflows created on the fly to solve a specific problem. Instead of using AI primarily for coding, he uses it to manage cloud resources, wrangle hardware, and automate his to-do list.

What I love about Marco's approach is its pragmatism. He's not building complex, permanent applications. He's identifying points of friction in his day and using natural language to command an AI to fix them. It's about saving a few minutes here and there, which, as we all know, adds up to hours of reclaimed time for more strategic, creative, and enjoyable work.

We explored five specific workflows that you can replicate today. We’ll cover automating Azure administration, scanning and merging documents without touching scanner software, compressing huge video files with a single command, and building agents that schedule meetings and find content for you. These are the kinds of automations that truly make life easier.

Workflow 1: Managing Azure Roles with Warp

If you've ever been victimized by the web interfaces for AWS, GCP, or Azure, you know the pain of managing user permissions. It's a maze of clicks, menus, and confusing role names. Marco showed me how he bypasses this entirely using Warp, a terminal supercharged with AI.

The Problem It Solves

The task was simple: give his colleague, Govin, several specific roles on an Azure subscription. Doing this through the Azure web portal would have involved navigating to different sections for each role, searching for the user, assigning the role, and repeating the process multiple times. Marco estimated it would’ve taken him an hour.

The AI-Powered Workflow

Instead, he just told Warp what he wanted to do in plain English.

1. Open Warp: He starts in the terminal.
2. Write the Prompt: He wrote a direct, conversational prompt.

give him Azure AI user and Azure AI project Manager on this subscription that I'm looking at. ... actually I need to give him contributor role on the whole subscription

1. Execution and Results: Warp interpreted his request and translated it into a series of commands for the Azure CLI (`az`). It executed them one by one. I even noticed that when one command failed, the agent recognized the error, corrected its approach, and successfully completed the task. The entire process of assigning multiple roles took seconds, not an hour.

Key Insight: Guiding the AI with Rules

Marco adds a crucial layer to make this process more reliable: Rules. He has pre-configured rules in Warp to guide the agent. For example, he has a rule that reminds him to activate his owner access before trying to assign roles, a common failure point. He also connects Warp to the Microsoft Documentation to ensure it can look up the correct role names if he doesn't know them.

Workflow 2: Automating Document Scanning and PDF Merging

This next workflow is for every parent, student, or office worker who has ever wrestled with a scanner. Marco needed to scan his daughter's two-sided math practice test. The old way involves scanning the front pages, saving a file, scanning the back pages, saving another file, and then fumbling with a PDF editor to merge and reorder them. It's annoying.

The AI-Powered Workflow

Marco turned this multi-step chore into a simple conversation with his terminal.

1. Prerequisite - Install a CLI tool: Marco first installed NAPS2 (Not Another PDF Scanner 2), a free, open-source tool that provides command-line control over his scanner. He then created a rule in Warp telling it to always use NAPS2 for scanning tasks.
2. Scan the Odd Pages: He placed the stack of papers in the scanner's feeder and typed a simple command into Warp.

scan the documents from the feeder and save it to this directory as this file name

Warp immediately activated the scanner and saved the file. No clunky software needed.

1. Scan the Even Pages: For the second pass, he didn't even need the AI. He just pressed the up arrow to recall the last command and manually changed odd.pdf to even.pdf.
2. Merge the PDFs: With both files ready, he gave Warp one final instruction.

Now put together the odd pages and the even pages and just make the math practice test out of it.

To accomplish this, Warp wrote and executed a small Python script on the fly using the PyPDF2 library to interleave the pages from the two files into a single, correctly ordered PDF. It then deleted the temporary script, leaving just the final result. Pure magic.

As Marco pointed out, while the AI was handling the scanning, he was able to help his daughter with a math problem. He traded time spent fighting with software for time spent with his family.

Workflow 3: Analyzing and Compressing Video Files

Anyone who works with video knows the pain of enormous file sizes. Marco recorded a 10-minute screen share that inexplicably ballooned into a 1.7 GB file. Instead of searching for a video compression tool, he again turned to Warp.

The Problem and the Prompt

He wanted to not only fix the file but also understand what went wrong. His prompt was a mix of a question and a command.

Why is this file so big use FM Peg to re-encode it? Still keeping it at 10 80 p... and make it more normal size.

The Result

The AI first analyzed the file's metadata and reported back: the file was huge because of an unnecessarily high bitrate. Then, following his instructions, it generated and executed the precise FFmpeg command to re-encode the video. The result? The 1.7 GB file was compressed down to a perfectly normal 13 MB, all while preserving the 1080p resolution. This is a fantastic example of using AI for file manipulation, a seriously underappreciated use case.

Workflow 4: Building a Triggered Meeting Scheduler

Moving beyond the terminal, Marco showed how this concept of "ad hoc agents" is appearing in mainstream business tools. He used Microsoft 365 Copilot to build a small agent to handle meeting requests from me.

The Automated Workflow

Inside M365 Copilot, he accessed the new "Workflows" agent and gave it this prompt:

When I get an email from CLA VO requesting a meeting at a certain time, check my calendar. If that time is free, send her a 30 minute meeting invite for that time.

In seconds, Copilot built a triggered agent. This agent now runs in the background, connected to his Outlook. When an email arrives from me with a meeting request, it automatically extracts the proposed time, checks his calendar availability, and if the slot is open, sends a meeting invitation. This removes him from the critical path, allowing him to be highly responsive without constantly checking his inbox.

Workflow 5: Creating Recurring Content Alerts with ChatGPT

This same capability is also available in consumer tools. Marco showed how you can use ChatGPT to create a recurring task, essentially your own personal cron job, without writing any code.

The Simple Prompt

To stay on top of new episodes of this podcast, he gave ChatGPT a simple instruction:

every day look to see if there's a new podcast by CLA VO and notify me if there's a new one.

The Result

ChatGPT immediately configured a recurring job. Every day at 9 AM, it automatically checks for new episodes and will send him a desktop notification if it finds one. This is a powerful way to automate information gathering for any topic you care about, from stock prices to new research papers to podcast episodes.

The Power of Ad Hoc Agents

What ties all these workflows together is the idea of creating small, disposable AI agents to solve immediate problems. Marco isn't building products; he's eliminating friction. Whether it's through the command line in Warp or a GUI in M365 Copilot, the line between using an agent and building an agent is blurring.

To make his interactions even faster, Marco uses a tool called AutoHotkey to create keyboard shortcuts for his most common prompts. For example, he can type MBF5 and it will expand to a pre-written prompt like, "Answer from the perspective of Microsoft in 500 characters or less with no bullets or formatting." This lets him get consistent, high-quality results from AI with minimal effort.

I encourage you to think like Marco. Look for the small, repetitive, and annoying tasks in your own day. Chances are, a simple prompt to the right AI tool can automate it away, giving you back precious time to focus on what truly matters.

Thanks to Our Sponsors

• Rovo—AI that knows your business
• Lovable—Build apps by simply chatting with AI

Episode Links

• YouTube: https://youtu.be/diy3kmUl8mY
• Spotify: N/A
• Apple Podcasts: N/A

Find more from Marco Casalaina on LinkedIn. And you can find me, Claire Vo, on X and my website.