My whole professional career has been about the quest to see how much tech stuff one can do in a business context without writing code. This started way before the terms “no-code” and “low-code” were a thing. Working with customer data and marketing scenarios, I first did my tricks with graphical BI tools, then configurable business applications like Microsoft Dynamics CRM/XRM.

Five years ago, in 2020, the time felt right to go all-in with low-code and start leveraring this approach outside the Dynamics 365 world. Power Platform had become an actual thing to be developed and sold by Microsoft. Meanwhile, Office 365 customers were starting to realize that citizen devs were already building apps and flows with the capabilities included in their current licenses. So, I co-founded the first Power Platform consultancy in our market and replaced my CRM hat with my low-code hat. This thing just had to become huge, right?

Today, Power Platform is more popular than ever. Yet I wouldn’t choose the same “100% low-code” strategy if I were founding a company in 2025. Because it seems obvious to me that the pendulum has shifted from “less code” to “more code”, thanks to large language models. The next few years will likely see a massive increase in code-first solutions rather than GUI driven configuration.

I will reflect on the possible outcomes from this paradigm shift in a future newsletter issue. But first, let’s discuss how the emergence of AI-driven “vibe coding” has impacted earlier assumptions about of how citizen devs, pro devs and designers might embrace low-code tools.

Vibe coding with AI

ChatGPT and similar LLM based tools are great at coding because code is ultimately a language spoken by computers. Another inherent quality of LLMs, their hallucinations, is also much less dangerous in coding than in most other scenarios. This is because it’s easy to “fact check” the code output of AI: run the code, see if it works. Not something you could easily apply in various business processes where tech vendors like Microsoft and Salesforce are now encouraging customers to adopt AI agents. That’s why I believe the current wave of GenAI will impact solution building more than operating those solutions.

It used to be difficult for us non-programming human beings to instruct computers to do what we want them to. To make this possible, the abstraction layers that offer GUIs to translate our intents into computer language became a big business. Now, the arrival of LLMs challenges these assumptions. Suddenly, it’s possible for us humans to type things in our human language and get that translated into computer language. Everyone now has the option of doing it via GUI or through the “code” of written language. Understandably, people are drawn to this new possibility of doing things on their terms - rather than learning to operate yet another GUI.

As an example, when I was recently exploring a customer need that included having a custom UI on top of product data managed in Dataverse, I decided to not bother with a PowerPoint mock UI that I might have previously drawn. I didn’t even want to create a fake UI with a canvas app since it needed to look like something a business user would want to use. After all, canvas apps are often too much like clickable PowerPoint slides: functional but ugly.

Instead, I fired up ChatGPT with the o1 model that now supported running simple code in its Canvas. This meant I could specify my idea and requirements in the chat, then let the AI “reason” for a moment and produce an output that could be immediately previewed in the canvas. I didn’t have to learn any dedicated tools for producing mock UIs, which I’m sure there are several out there for scenarios like this. I just typed it into the same tool I use daily anyway - and it worked.

Sure, it didn’t look and work exactly the way I had envisioned it in my mind. But it was definitely close enough to keep on working with. A few days later, I wanted to try a tool specifically aimed at turning prompts into apps. So, I simply asked ChatGPT to describe the outputs it had created so far in a way that would serve as a detailed prompt for lovable.dev. After a few minutes, I had again a new version of the app, in a more refined format, ready to preview online, and to modify via natural language prompts as needed.

I get the feeling that much of what was referred to as prompt engineering in the early days of ChatGPT 3.5 has now been absorbed into the built-in capabilities of GenAI tools. Meaning, you can ask AI to get a better AI prompt. What this also covers is the possibility to transport instructions from one service to another. Even the crude copy-paste method can be sufficient for many ad-hoc needs.

Beyond the plain English instructions, the output from app generation tools like lovable.dev is nothing like a Power Platform solution package. You get the raw code, with shortcuts and integration points to services that can help you run the code and manage it. There are no guarantees or guardrails in this world, like you would have in a low-code product environment such as Microsoft’s Power Apps. If something fails or breaks, you can only ask AI to try and fix it.

Or you could fix it yourself - if you understand how the code works. For a low-coder like me, this isn’t going to be an option. Yet even experienced software developers who could technically produce better outputs themselves are starting to let AI handle things. This is the essence of the vibe coding concept. Coined by Andrej Karpathy, here’s how he describes the meaning of the term:

“It’s not really coding, but we do it anyway.” While this clearly isn’t a suggestion that all of traditional programming should be replaced with vibes, there’s little doubt in my mind that people will be using AI in this way more & more. Not because it’s better than human experts, but simply because it is possible. It makes the act of working with code accessible to a far broader audience than ever before.

This brings us to the impact that AI code generators will have on any plans created before they became a thing. Some of which is as recent as the rise of low-code platforms for business apps.

The early death of Fusion Teams

Low-code tools like Power Apps have been adopted almost exclusively by business users who don’t know how to create apps with pro-dev tools. Because there’s more people in this target audience than there are professional developers, the impact to business applications and software business has been significant. Analysts and vendors have been positioning these as the solution to address the skill gap of not enough developers in the world compared to the projected numbers of apps to be built. Which has now risen to 1 billion:

Just like with AI generated code, though, citizen developer built apps and automations haven’t exactly met all the requirements for enterprise IT systems. Not only has there been pressure to set up governance models for Power Platform. The fact is that any model alone will not be able to change the quality of casually built apps or flows. The citizens can only go so far - but luckily Power Platform was designed to offer a “No Cliffs” experienced for developers of all kind. Right?

Microsoft started promoting the concept of fusion development three years ago. The idea is based on “using low-code capabilities and combine them with code-first components to meet business needs and create Fusion applications.” In short, you would bring pro-devs in to help the citizen devs by building things on the Microsoft stack that weren’t feasible or sustainable if done via low-code tools alone.

While today there are true professional developers and architects familiar with code-first development in many Dynamics 365 teams building bigger business apps and integrations, this hasn’t been too common in the space of pure low-core Power Apps. If the projects and initiatives haven’t been big enough to have dedicated roles for pro-devs to specialize in, it’s been a struggle to get them convinced why they should bother learning anything about these GUI based Power Platform tools.

Is there a way to lure professional software developers to work with Power Apps, or even with the new “UI for AI”, meaning Copilot and its agents? Right now, it’s hard to see how a movement like this could be organized by Microsoft. Even if the tooling around Power Platform is becoming more pro-dev friendly, like with the built-in Git integration of Dataverse, those improvements are not going to be reason enough for developers who have a choice of tools. It’s gonna remain a business driven requirement to “settle” for low-code.

This means that the business side would have to keep demanding lower percentage of custom code in solutions built for them. But what happens to such demands when the business folks start to tinker with code on their own?

Could citizens really work with code?

AI remains an abstract concept, until you can show something tangible it can produce. This is why the examples where code generators are given instructions and the AI spews out a running application or a game can so easily go viral on social media. Like watching robot dogs do ever more impressive tricks, this kind of videos tend to get under our skin a lot more effectively than fluffly promises about business benefits. Maybe it’s the uncanny valley in action.

Are these tricks or real skills, though? In his blog post “AI Coding Fantasy meets Pac-Man”, Gary Marcus took aim at a recent example published by The Guardian. In it, the reporter approached seven people on X and asked them if they could get Grok to generate a Pac-Man clone.

Marcus points out that the entire game play can be described in a page. Furthermore, the original source code of Pac-Man on the Atari 2600 is available on GitHub. Despite of there surely being plenty of Pac-Man related material in Grok’s training data, the end results of trying to get AI to generate a game to meet the bar of a clone were a massive failure.

A legendary video game from 45 years ago is not within the realm of possibilities for today’s GenAI tools to generate yet. Will it be in a year, with ever bigger models? Even if that would be the case, would it prove this technology alone to be viable for creating real games to play or apps to use in business? The problem really being that once we stop looking at the cool tricks of the computer generating something, there remains the problem of how to get it to generate something very specific? Something reliable enough to warrant the Full Self Driving era of AI.

If you can’t take your hands off the wheel completely, you still need to be able to drive the vehicle when needed. There’s no skipping the part about acquiring a driver’s license. That problem is similar to the low-code scenarios of apps that could perhaps be used in limited areas for low-risk scenarios. You can let a kid drive a toy EV in the back yard. You can’t let them drive a real EV to go to the supermarket - even with all the driver assist technologies available today. When there are no guardrails set by others, someone has to take responsibility. Today, that “someone” is always a human.

You know what could be a great guardrail for AI? Having a feature where the code generator itself tells the user to not proceed unless they are skilled enough to take responsibility of the output. This was recently demonstrated in a hilarious example: “Cursor AI assistant tells vibe coder: learn to code”:

This is more than just a funny piece of AI hallucinations. Because if there’s something LLMs excel at today it is being a personal coach that helps you to learn about topics of interest. You can throw endless “explain this to me like I’m…” type of prompts at the AI assistants and they’ll frame the content from either their training data or web results in a highly personalized way. If you really wanted to learn how to understand code and also write some yourself, AI is an amazing tool to help you.

Today, when I’m building either Power Fx formulas or trying to manipulate data with Power Automate cloud flow expressions, the barrier for me to go and ask ChatGPT is super low. I generally find the current built-in Copilot tools from Microsoft to be much less useful than a generic ChatGPT where I can throw my specs, screenshots, and any question to for processing. I’m essentially vibe coding with low-code, rather than doing traditional searches for blog articles that would be a close enough match for my problem at hand.

Just like we’ve seen Stack Overflow die as software developers ask AI tools for answers instead, I believe the citizen developers will behave just the same way. The question really is whether they would end up using targeted Microsoft built experiences like the Power Apps plan designer, or will they choose to explore a world outside a specific vendor ecosystem? Which would naturally be not a good result for someone like Microsoft that has managed to capture a sizeable share of citizen developer business app activity into their proprietary cloud.

It’s a phenomenon worth paying attention to. So, I personally joined a local Meetup dedicated to vibe coding, just to understand better how people with different experiences in tools and ecosystems are seeing its potential in everyday business.

Whereas professional UX designers never were fans of the Power Apps approach that allowed quick creation of functioning yet ugly apps, the vibe coding approach is surely to appeal much more to this audience. Several AI code generators are promising to convert Figma designs to working apps. While MS did build a one-off feature like that a few years ago, they never invested in making this actually attractive for professional designers.

Now, if this designer audience learns to use AI app generator tools that keep them firmly outside the MS ecosystem, what do you think will be chosen as the technology and environment for downstream app development? Most likely not Power Platform. This might not a critical gap in the eyes of Microsoft, however, if they intend to move on from apps towards Copilot as the primary UI for information workers and agents anyway.