The Iceberg Below the Code

A Medium article by Andreas Leicher made the rounds recently: “AI-Assisted Coding: The Tip of the Iceberg in Software Development.” His thesis is solid — AI can build you an MVP fast, but the 70% of work below the waterline (architecture, security, scalability, integration) is where projects actually succeed or fail.

He’s right. But he’s also being generous.

Because even that 70% — the engineering iceberg — is still just talking about code. And code is the easy part.

The Real Iceberg Has Seven Layers

Here’s what’s actually below the waterline when you build software for a real business:

Layer 1: The Functional App (Visible — 10%)

This is the part LinkedIn celebrates. “I built a SaaS in a weekend with Claude!” Cool. It works on your laptop. Ship it to production and watch what happens.

Layer 2: Architecture & Technical Debt (15%)

AI doesn’t architect systems — it solves prompts. Your weekend SaaS has seventeen files that all import from each other in a circle, three different ways to format a date, and an auth system that stores passwords in localStorage. This is what software entropy looks like on day one — and it only gets worse. This is where Leicher’s article lives, and it’s real.

Layer 3: Security, Compliance, Infrastructure (15%)

Your AI-generated app has the API key hardcoded on line 47. There’s no rate limiting. The database has no row-level security. You’re one GDPR complaint from a very bad day. This isn’t theoretical — secrets management is an attack surface that most AI-generated code ignores entirely. Enterprise buyers won’t touch it.

Layer 4: Integration With Reality (15%)

Your clean, beautiful app needs to talk to a 20-year-old ERP system that communicates via SOAP XML and crashes if you send a field longer than 50 characters. AI has never seen your company’s legacy stack. You’re writing glue code by hand, debugging OAuth signatures at 2 AM, and discovering that the data model everyone agreed on doesn’t match what’s actually in production — because it never does.

Layer 5: Active Listening (20%)

This is where the iceberg gets deep. Before a single line of code, someone has to figure out what the business actually needs — which is almost never what it says it needs.

Active listening isn’t a soft skill. It’s the hardest part of building the right thing. An AI coding agent will build exactly what you tell it to. An operator will tell you what you actually need — and then build that instead.

Layer 6: Reverse Prompting (15%)

This is the operator’s secret weapon. Instead of accepting the brief and building to spec, you push back. You ask the uncomfortable questions:

• “What happens when this data is wrong?”
• “Who owns this process today, and will they sabotage the new one?”
• “You said 10 users — but your org chart shows 60 people who touch this workflow.”
• “The last vendor built this exact thing. Why did it fail?”

Reverse prompting isn’t being difficult. It’s saving six months of building the wrong thing. Every experienced operator knows: the fastest way to waste money on software is to build what was asked for instead of what was needed.

Layer 7: Operator Intuition (10%)

This is the deepest layer. It can’t be prompted. It can’t be automated. It comes from years of watching projects succeed and fail across dozens of companies.

It’s knowing that the real problem with the reporting system isn’t the reports — it’s that marketing and sales use different definitions of “customer.” It’s recognizing that the CRM migration will fail not because of the technology, but because of the politics around who controls the data. It’s seeing that the $200K ERP project is actually a $40K data cleanup problem dressed in enterprise clothing.

No AI has this. No vibe coder has this. It comes from being in the room, in the system, in the business — day after day, decision after decision. It’s why the ops person problem is the most important hire most companies haven’t made.

The Market Is Flooded With Layer 1

Every freelancer platform is full of people who can “build apps with AI.” Upwork, Toptal, Fiverr — they’re all selling Layer 1. And Layer 1 is, genuinely, more accessible than ever. That’s a good thing.

But if you’re a business leader making a hiring decision, here’s what you need to understand: the person who builds the app is not the person who knows what to build.

AI-assisted coding has commoditized the top of the iceberg. Which means the value has shifted down — to the layers that require judgment, experience, and the willingness to tell you things you don’t want to hear.

What This Means For You

If you’re evaluating whether to hire an AI coding freelancer vs. an operator who builds:

The Source Material

Leicher’s article is worth reading — it’s a solid breakdown of the engineering iceberg. His point that “AI delivers 30% of the work” and the remaining 70% is where projects succeed or fail is well-argued and well-sourced.

But I’d push it further: even that 70% understates the problem. The engineering work is knowable. It’s hard, but it’s procedural. The truly hard part — the part no AI can do — is understanding the human system the software has to survive in.

That’s not a coding problem. That’s an operator problem.