Over the years, I have developed several business software applications that combine my expertise with pressing client needs. Some have been for my personal use, while others have become commercial products. One of them led to software used by more than half a million businesspeople worldwide today.

I’ve always started with a concrete problem that our clients or I have encountered and haven’t found a suitable existing tool to solve it satisfactorily.

For example, as a management consultant, I involved dozens or hundreds of individuals in collective ideation or insight-gathering. There was no low-friction, independent tool available for that purpose, so I devised an ideation app, which later became a commercial product.

Until recently, creating the first working software prototypes required significant effort. The emergence of LLMs has dramatically changed both prototyping and software development.

LLMs as co-developers

With LLMs, instead of having to code every interaction or resorting to purely visual prototypes, I can now iterate on a working app early on. LLMs also help with ideation and testing, acting as a development partner.

If I want to pass the final product development to programmers, they get a clear understanding of what I’m trying to achieve. Spec-writing has also become fast and almost automated.

The use of LLMs is not limited to the early phases of software development. Every startup extensively uses AI in production, and “Big Tech” companies currently generate between 25% and 30% of their code with LLMs. Some companies project that this share will reach 50% or more by the end of 2026.

Enter vibe coding

In a recent interview with McKinsey, Sonar CEO Tariq Shaukat highlighted a seismic shift in the software world: AI is not just making coders faster; it’s changing who can code.

People with little or no programming experience can create individualized apps with AI using a technique called “vibe coding.” The term was coined in early 2025 by Andrej Karpathy, a prominent AI researcher, co-founder of OpenAI, and former Director of AI at Tesla. No-code development tools have existed for a long time, but LLMs offer even amateurs a more intuitive way to build bespoke apps.

You describe, in natural language, what you want the software to do, and the chatbot generates the code using the framework you choose. Fixing errors and tweaking the code is an iterative process in which you provide feedback and instructions until the app works the way you expect it to.

Democratizing innovation

The McKinsey interview notes that AI lowers the barrier to entry. In AEC, this democratizes innovation.

Small design firms can now build proprietary internal tools that were previously affordable only to “Big Architecture” firms like Foster + Partners or Gensler. Professionals can eventually replace the notorious Excel contraptions with working apps that anyone in the company can use.

When the person who understands the problem is empowered to build the solution, we eliminate the “lost in translation” phase between designers and software developers. This approach leads to tools that are more intuitive and hyper-specialized for the project needs.

What is possible now?

In the pre-AI era, if a structural engineer wanted a custom tool to automate a specific carbon-loading calculation for a unique timber frame, they had two choices: spend weeks learning Python or C#, or wait for a busy IT department to prioritize it.

Today, using LLMs like GPT-4o or Claude 3.5, an engineer can generate Revit and Rhino plugins by describing a workflow in natural language and receive working code in seconds (e.g., “Write a Python script for Grasshopper that optimizes window placement for solar gain”).

They can also quickly build “glue” code that connects an Excel sheet to a BIM model or a site sensor feed to a dashboard. Tools like Streamlit and Vercel v0 allow AEC professionals to create functional web interfaces for their internal tools without knowing a lick of CSS.

Impact on the AEC software industry

The traditional software giants of AEC are no longer the only game in town. This signals the end of “one size fits all.” Customers are becoming more critical, demanding, and impatient. Instead of waiting for a major software update to include a specific feature, firms are building their own microservices.

Tech companies are aware of the trend and are transitioning from providing static tools to offering intelligent assistants and automated workflows. Still, there’s plenty of room for innovation beyond these tools.

The value is shifting from the code itself to the proprietary data and logic behind it. As Shaukat pointed out, AI helps write the code, but the human must still ensure the logic is “clean” and the intent is correct.

LLMs accelerate development cycles. Firms can prototype ten digital workflows in a week, kill the nine that fail, and move the winner into production, all before a traditional dev shop would have finished the discovery phase.

Building responsibly

No great power comes without its caveats.

The McKinsey interview raises a vital warning: Speed can be a trap. We have all experienced the hallucination factor. In AEC, a bug isn’t just a crashed app; it could be a miscalculated structural load. AI-generated code can look perfect but still contain subtle logic errors that an amateur might miss.

Building apps can lead to spaghetti code that’s hard to maintain. If an architect uses AI to build a complex tool but doesn’t understand the underlying code, what happens when it breaks six months later? Without proper “code health”, firms may end up with a library of “black-box” tools no one knows how to fix.

Security is a constant concern with LLM services. Pasting project data into free public AI models to help write scripts can inadvertently leak sensitive client information or proprietary design logic. Another security risk is buggy third-party modules that LLMs might add to your app.

From creator to curator

As we lean into AI for fast prototyping and app development, AEC professionals become curators of logic. They must ensure that AI-generated logic is “clean” and that the intent is correct.

The new role raises an interesting question: Why bother learning programming when AI can master it?

Structural engineers and MEP designers use sophisticated calculation and simulation tools, but they must verify that the results make sense. Likewise, you must learn enough about programming to ensure that your AI tools remain your servants, not black boxes.

Using AI mindfully and involving professional coders in the loop are ways to manage the risks of this new, powerful technology.

AI has democratized development, but it hasn’t outsourced accountability.