The Fractional CTO Was the Warning. Cognitive Debt Is the Invoice.

Last week, I wrote about the Fractional CTO's Dilemma. The two extinction paths: Token Monkey (knows every tool, understands nothing) and Vibe Coder (ships fast, production falls apart). I described what it's like to live in both traps simultaneously — fixing an agent's timezone bug on Tuesday, evaluating LLM extraction quality on Wednesday.

Here's what I didn't tell you.

There's a third trap. It's the most dangerous one, because you don't even know you're in it.

The Third Trap: Cognitive Surrender

Sebastian Christoph published a piece recently that crystallised something I'd been feeling but couldn't name. He calls it Cognitive Debt.

You know technical debt. It lives in your codebase. You track it, you schedule sprints to pay it down, you feel guilty about it at standup.

Cognitive debt lives somewhere worse: in your people.

It's the accumulated cost of not understanding the code that's running your business. Not because you're lazy. Because the AI wrote it so fast, and it works, and the tests pass, and you moved on.

Margaret-Anne Storey's research lays it out in three layers:

Technical debt — the code is messy. You know this one.

Intent debt — the code does something, but nobody remembers why it does it that way. The original requirement is buried in a Slack thread from six months ago.

Cognitive debt — nobody on the team can actually explain how the system works. Not the junior. Not the senior. Not the CTO.

These three layers compound. Technical debt makes intent harder to recover. Lost intent makes cognition impossible. And cognitive debt makes every future decision slower, riskier, and more expensive.

The Data Is Already Here

This isn't theory. The evidence is stacking up.

Anthropic's own internal study found that engineers using AI scored 17% lower on knowledge quizzes about their own software. The company that builds the AI is telling you: using it makes you understand your code less.

A METR study showed experienced developers were 19% slower when using AI coding agents. Not faster. Slower. Because the cognitive overhead of verifying AI output exceeds the time saved by generating it.

DX's longitudinal research across 40 companies found AI productivity gains of only around 10%. Not the 10x we keep hearing about.

Martin Fowler put it perfectly:

"A rather dodgy collaborator who's very productive in the lines-of-code sense of productivity, but you can't trust a thing they're doing."

He's right. And if that collaborator were human, you'd be having words with HR.

What This Looks Like in Malta Right Now

This isn't just a Silicon Valley problem. It's especially acute here.

The MGA just announced a record number of license applications. Malta's FinTech infrastructure is going cross-border. The ecosystem is accelerating. More code is being written. More systems are being built. More AI is being used to do it.

But here's the question nobody's asking: who's building the verification layer?

The speed is real. The infrastructure is real. But speed without understanding is just accelerating toward the wrong destination.

How We Handle This at Opteia: The 90/90/90

I'm not here to tell you not to use AI for coding. That would be hypocritical — 90% of our code is AI-generated. We use it every day. We build our products with it. We sell AI services to clients.

But here's the other 90: 90% of our verification, judgment, architecture, and intent is human-driven.

That's not a contradiction. It's the point.

When the AI generates a module, a human reviews it. Not just "does it pass tests?" but "do I understand why it works?" If the answer is no, we don't ship it. We break it apart, we name the pieces, we make sure the team can explain it to someone who's never seen the codebase.

Ajey Gore, former CTO of Gojek, described the future org chart: "The team that used to have ten engineers building features now has three engineers and seven people defining acceptance criteria, designing test harnesses, and monitoring outcomes."

We're already there. Our bottleneck isn't writing code. It's verifying it. And we think that's exactly right.

The Practical Playbook

Here's what we actually do, not just what we preach:

1. Explanation reviews on every PR. Before merging AI-generated code, the reviewer has to explain why it works. Not what it does. Why. If you can't, you don't understand it well enough to own it.

2. Specs before generation. We write the requirement, the acceptance criteria, and the edge cases before we let any agent touch a keyboard. Spec-driven development is cognitive debt prevention.

3. Rotate ownership. No piece of AI-generated code becomes one person's mystery. If you built it with AI last sprint, someone else reviews it this sprint. Fresh eyes catch what familiarity normalises.

4. Preserve the hard parts. Domain boundaries, error handling, security-critical paths — humans write these manually. Not because AI can't, but because the act of writing them builds the mental models you need when things break at 2am.

5. Name the debt. In our architecture reviews, we ask: "Who can explain why this module works?" If the answer is "the AI," that's a red flag. Code without a human owner is orphaned code.

The Hard Truth

Jeremy Howard said it best:

"If you're not actively using your design and engineering muscles, you don't grow. You might even wither."

AI is not going to slow down. The tools are going to get better, the code generation is going to get faster, and the temptation to surrender cognition is going to get stronger.

But here's what I've learned running an AI company: the scarce resource in software engineering is no longer the ability to write code. It's the ability to determine if code is correct.

The Fractional CTO's dilemma was the warning. Cognitive debt is the invoice. And the only way to pay it is to stay human in the loop — not as a bottleneck, but as the verification layer that makes speed sustainable.

That's our 90/90/90. It's not perfect. But it's honest. And in an industry full of 10x claims, honest might be the most competitive advantage of all.

We build AI agents for businesses in Malta and beyond. If you want to adopt AI without the cognitive surrender, let's talk.