Power Automate via GitHub Copilot

I've seen a few articles pop up recently regarding moving away from maker portals or building Power Apps and Power Automate flows using natural language. I thought to myself...I wonder if AI can really replace building some of these things, or if it's something that people will test out, say is cool and go back to building traditionally? 🤔

I am loving the parts that AI can build for me, skipping things that take are just monotonous, things I have to do a lot like setting up Solutions and Publishers, common framework or scaffolding to start a project, is a great thing which could be automated via an AI skill.

My AI harness or interface of choice is the GitHub Copilot App. I started playing around with it earlier this year and it's now my go-to when it comes to anything AI and building.

I checked out the plugin menu to see if the there was a plugin for authoring Power Automate flows and also installed the other model-apps, canvas-apps, code-apps and anything else that sounded Power Platform-y. Maybe that's for a future post.

I start prompting and it takes me through the steps of getting connected to the environment and I tell it which environment and what I want to do. I start it off with creating a solution, using an existing publisher and everything works as expected.

Most of the above and what I want it to do, I do in a couple of prompts

Once everything is logged in and solutions/publisher created, it builds the flow and asks me if I want it to be turned on. Before I say yes, I go searching, find the solution, open it, but no flow inside.

I find it under My Flows, and go back and get it to add the flow to the solution and turn it on.

I open the flow and verify it is built how I requested and it is. I go to open it and I get an error.

Returning to the classic designer does open the flow successfully and I can then toggle to the new designer fine, so maybe this is just a quirk of prompting a flow. I confirm my manual trigger and my formula in the compose step is right. I'm impressed it got the formula correct.

formatDateTime(utcNow(), 'yyyy-MM-dd')

Next I prompt something slightly more complex, a condition and a connector that requires a connection reference and a connection.

It thinks about it, updates the flow and asks me some questions about the parameters required, like location and what units I want the output in. Very smart that it understands these...but then again, it's just JSON, so why wouldn't it? I give it the details and it finishes up.

It then proactively tests the flow for me and tells me the outputs 😱

That's pretty cool!

I open up the flow and find that it has created a formula for the condition, with no direct instruction that it had to do that or anything, it just figured it out.

Looks great in the new designer too!

If I want to I can even ask GitHub Copilot to show me the schema of the flow, as it's all just JSON under the hood.

It's one thing to look at the schema and be able to understand what the flow looks like, it's another to see it in the designer...but you can also ask for a visualisation as well.

The visualisation of this is the part where I start to maybe see a divergence here. If I have a desire to create a bunch of basic automations, or even give a bunch of instructions to an agent to build the scaffolding of an automation, then this seems like a quicker way to do it. However, if it's a complex flow, where I need to populate a bunch of fields from one system/source to another, where I will have lots of switch statements, conditions, loops...will this be the best place to do it? Considering I can't visualise the flow in real time and instead have to prompt for a visualisation? I'm not so sure. But maybe I'll keep playing around with this.

Have you tried this yet? What do you think?

Ciao for now!

MCJ