We don’t hear Microsoft use the “Power” term much anymore. In the past two quarterly earnings calls, neither Satya Nadella (CEO) nor Amy Hood (CFO) have mentioned Power Platform. These days, if MS would talk about power in financial performance context, it would more likely be about where to find electricity for all the new datacenter projects.

Power is out, Copilot is in. Not big news at this point anymore — to any of the subscribers of this newsletter, at least. This week’s issue isn’t about MS investor relations, though. It’s about the newly released App Builder agent for Microsoft 365 Copilot and how that challenges the earlier “no cliffs” path Microsoft has been talking about with Power Apps.

Released as part of the same Frontier program as the Flow Builder agent that I wrote about a while ago (now rebranded as Workflows agent), the experience is remarkably similar. Instead of opening any Maker portal or Designer UI, you just chat in Microsoft 365 Copilot. You enter a prompt, then the App Builder starts building it for you. This is where it happens:

I must warn you that this AI agent has literally no brakes at all. If you happen to just mention it in a Copilot Chat, like “Hello, @app builder”:

It will not respond back to you with a text. Instead, it will immediately initiate a 12-step process of generating an app for you. Below is the Hello App the agent built in response to me testing whether I could talk with it to ask about my existing apps, for example. (Lesson: no, you can’t use the agent for this.)

We can all guess why there’s no validation in place before the app creation begins. Because of the metrics! Just like SharePoint Agents were intentionally designed to be generated in 10 seconds, this thing is a Paperclip maximizer. It only wants to create more paper clips apps to pursue its goal:

Do not stop to ask whether an app/agent should be created. Remember the Growth Mindset.

Let’s park the snark for a moment and look at the outputs from App Builder. In short, the agent will create typical React web apps with the UI elements you’d expect. Cards with rounded corners, always starting with a dashboard that counts various numbers of records. Like in this example where I said a bit more than “hello” to the App Builder:

If you have never tried vibe coding tools aimed at casual app creators, I bet the AI creation looks pretty impressive. These magic tricks that LLM-based tools can do is a key reason behind the high expectations for AI being able to replace human workers at least in some tasks fairly soon. In practice, these coding skills for basic web apps have been in your regular ChatGPT or Claude for quite some time now, thanks to all the source code data the foundational models have been trained on. Because computer code is just a very structured type of language.

When you wrap these capabilities into features explicitly shown to users in software UIs, they become more tangible than merely some dark art of prompt engineering. There’s the saying that most AI startups today are merely wrappers around OpenAI APIs. In practice, the value of software products comes from how well they can connect these raw tech capabilities into useful tools for humans, which is still requires plenty of effort and design skills.

For comparison, let’s throw the same prompt at the Swedish GenAI startup Lovable that has focused on the vibes-based app building for a bit longer than Microsoft. This was the assignment:

Above you saw the version from M365 Copilot App Builder agent. Below is what Lovable generated from the same one-shot prompt:

Okay, in terms of UI design and creativity, Lovable clearly beat Microsoft. As I expected, the outputs from a single agent within the M365 Copilot domain won’t match the level of products developed specifically for this purpose. Just like I wouldn’t give up my Canva subscription and start only using the Create tab in Copilot, the App Builder agent isn’t a replacement for best-of-breed tools in AI app generators.

Business apps aren’t just about the user interface, though. The data story is where Microsoft should logically be able to shine. If you’d enable Lovable Cloud as the back end of your Paperclip ERP, the data would get stored in a Postgres database provided via Supabase. If, on the other hand, you choose to build your Paperclip Production solution with Microsoft App Builder, the data will be in… SharePoint.

Oh…kay.🫤

For context, this meme pic about SharePoint vs. databases has earlier been shared even by Microsoft employees working with M365 technology:

But that’s just how it goes now. Whenever you prompt the App Builder agent to create an app for you, a new SharePoint site is provisioned. It will have SharePoint Microsoft Lists to represent each table that the app needs for storing data. And yes, the app will fake the relationships between lists.

Previously, citizen developers used to resort to SharePoint as the data source for their Power Apps canvas apps, to avoid having to buy a $20/user/month license for premium features. Today, when you purchase a $30/user/month Microsoft 365 Copilot license for your employees, they get to use an AI agent that builds apps with SharePoint as the data source.

“But surely this is just the starting point! By having App Builder generate the basic UI and logic, makers get a personalized template that they can open in Power Apps and continue building…” Nope. These are not Power Apps. At least not canvas apps or model-driven apps. What the App Builder produces are code apps.

If you open the special environment in which the app artifacts are stored, you’ll see the App Builder creates apps of type “Code”:

There’s no edit button enabled, because these things aren’t meant to be editable inside the Power Apps maker portal:

As of today, you can’t touch the code even through an IDE like VS Code. It is one of the limitations listed in the instructions for Build apps with Microsoft 365 Copilot: “Generated code isn’t viewable or editable. App makers cannot see or modify the code App Builder generates.”

I know it’s a preview and I believe things will change once we get closer to GA release. Nevertheless, the direction seems clear: either you prompt the AI agent to modify the app, or you take full ownership of the code that you’ve directly manipulated through your IDE of choice. The choices for app development are either A) use only chat or B) edit code + chat (with GitHub Copilot, Cursor, Claude Code or whatever pro-dev AI tool you prefer).

To me, this doesn’t sound nearly as powerful for a citizen developer as what PowerApps promised back in 2015. The original “no cliffs” story of you can start with a zero code app created by a business user and gradually evolve that to more complex scenarios. You didn’t jump straight into the deep end, instead there was a gradual path that the low-code tooling and the model-driven development with the help of Dataverse provided to app makers.

I thought about naming this newsletter issue “The Return of The Cliff”. As well as doing a funky Suno AI version of the lyrics from this 90s hit. But I had to leave that for another day…

We need to keep in mind that these new agents behind a Microsoft 365 Copilot license are not the only path that MS sells for making apps. Nor is it the only place where AI is being used to reimagine the app building experience. Let’s take a look at the tooling on the Power Platform side.