Building software 5x faster with AI: the good and the not so good
How I built compliance software in 4.4 months with AI and a methodology: what I gained, what it cost me and what I would do differently.
In this post
With AI I built software about 5 times faster than a traditional team, and I measured it. In 4.4 months, practically on my own, I shipped multitenant compliance software that is in production today and still growing; in human effort, the difference was 23 to 67 times. But it wasn’t magic: it worked because I followed a clear methodology, had a senior developer auditing alongside me, and it cost me exhausting workdays that no video ever shows you.
If you are starting to work with AI, this is what I wish I had read first.
Before AI: the design almost never made it through intact
My process was long. I benchmarked competitors to see whether they already had the feature and what their experience looked like. I validated with stakeholders, designed in Figma, validated again and, once approved, handed the design to the developer.
The developer built it, but the result rarely looked like what I had proposed. That gap between the design and the real product was the most frustrating part of the process.
How I got here
I started using AI at the end of 2023 for simple things: validating an idea or writing it better. By the end of 2025 I got serious and started building agent workflows in n8n. That’s where I truly discovered its potential and how much it could simplify the way I work.
Curiosity led me to build an application with n8n as the backend: a tool to generate tutelas (constitutional rights claims) based on Colombian case law. It failed. I soon realized it would never scale and that the reliability of its transactions was at risk. Today I know I didn’t plan and document from day one.
Even so, that failure gave me a solid foundation in development concepts. With it, I moved to a sturdier stack and to working directly in the code with agents that understand the whole project’s context, like Claude Code and Codex.
For me it has been a rollercoaster. First I felt curiosity and the urge to be among the first to build something big. Then I felt fear: I could do things that used to be impossible for me, or that would have taken me years to learn. I felt empowered and, at the same time, I kept wondering what it means for one person to do the work of an entire development team.
The project: compliance software in 4.4 months
The project is a SaaS due diligence platform for companies in Colombia. It lets them verify the identity of clients and suppliers, screen them against restricted lists, profile their risk and decide whether to onboard them, with full traceability for every decision. All of it under the frameworks for anti-money laundering (SAGRILAFT), anti-corruption (PTEE) and data protection.
I did all the planning with Falcux AI-First, the methodology I built to work better with AI: I defined the project, the architecture, the database schema and the specs for each module. Then I implemented it with protocols for every new feature, every change and every release.
I did the implementation alone. But I want to be very clear: I had a top-level developer by my side who audited the code, validated the system’s security and configured the infrastructure. Without him, this post would be a different one.
These are the numbers, taken from the repository history and compared with industry reference rates (ISBSG and COCOMO II):
| With AI (measured) | Traditional (modeled) | |
|---|---|---|
| Calendar | 4.4 months | ≈ 22 months |
| Team | 1.2 people | ≈ 12 people |
| Commits | 1,390 | — |
| Documented formal changes | 333 | — |
What struck me most was seeing design and code built by AI. It was impossible for me to believe work could speed up that much, and seeing these numbers still feels absurd to me.
The good
- I build end to end. Today I can take a web application from idea to production.
- The design looks like the product. I go more directly to code, so much of the gap between what I design and what gets built disappears.
- I do things I couldn’t before. Things that would have taken me years to learn are now within reach.
- Every tool has its place. Claude Code and Codex to build applications, Figma with AI for quick ideas and concepts, Gemini for questions and research.
The not so good
AI makes things up and loses the thread
It adds features no one asked for. It struggles to follow long flows with many variables. And in my case it went as far as pushing to GitHub without authorization and changing the Docker files on its own. If you don’t review what it does, it ends up deciding for you.
Working code doesn’t mean it’s ready for production
The senior developer found things neither the AI nor I had seen: the project wasn’t versioned, there was no event logging, the AWS buckets weren’t configured properly, and deploying with AI wasn’t as easy as it looked. AI writes code; running a system in production is something else.
It can wear you out
It was exhausting, mentally and physically. I worked long days to meet inflated expectations of what AI could do. That overestimation was the whole team’s responsibility, including mine. If AI makes you faster, the risk is that the plan assumes you are infinite.
Visual refinement falls behind
Going more directly to code, screens no longer get refined on their own. If I don’t take the initiative, it doesn’t happen, and that has meant extra effort. That’s why I’m adapting my design system to reuse it across a suite of tools and to make it easier for AI to understand.
What is still yours
Judgment, talking to stakeholders, understanding the business and deciding what not to build. I never delegated that.
AI can take you to a level you might never have imagined, but only if you approach it with a clear and well-planned direction.
If you’re starting out
- Define your project very well from the start. Before asking for the first line of code, be clear about what you are building, for whom and why.
- Build the artifacts for the AI and for the team. Specs, architecture and protocols that serve both the agents and the people who will work on the project.
- Learn development concepts and logic, even if you no longer write code. That is what lets you spot when AI gets it wrong.
That’s how Falcux AI-First was born: from turning all of this into protocols I can repeat in every project.
Falcux AI-FirstThe methodology I use to plan, build and version products with AI agents.ai-first.falcux.comWhat comes next
Honestly, the future feels uncertain to me. I still have fears about whether AI will replace us. For now, I keep adapting and making the most of its potential.
I think the designer and developer roles will merge into one: the digital product builder with AI. And within that role, specialties will probably emerge: some of us will focus more on design, and others on software quality, security and scalability.
Let’s not take humans out of the equation
We all share a responsibility toward AI: to demand that its use doesn’t take humans out of the equation. In my project, what worked wasn’t AI alone. It was AI with a methodology, with one person planning and another auditing. If you’re going to start, start there.
Can you build scalable software with AI alone?
You can build it with AI, but not with AI alone. In my case it took planning with a clear methodology and a senior developer auditing the code, security and infrastructure.
How much faster is working with AI?
In my project, about 5 times faster in calendar time and between 23 and 67 times less human effort than a traditional team modeled for the same scope.
What is the biggest risk of working with AI?
Overestimating it. AI makes up features, gets lost in long flows and won’t leave a production-ready system on its own. And if the plan assumes you are infinite, it wears you out.