James here, CEO of Mercury Technology Solutions. Hong Kong - January 5, 2026
I ran an experiment recently. I took a codebase I’ve maintained for nearly a decade—a living, breathing, revenue-generating fossil—and asked an AI Agent: "What are the top three problems with this project?"
The AI’s report was impressive. It was professional, logical, and technically accurate. It pointed out:
- Divergent file structures.
- Inconsistent naming conventions.
- High coupling between modules.
From a "Computer Science 101" perspective, the AI was 100% correct. But as the person who has kept this ship afloat for ten years, my reaction was: "These are real observations, but they are not real problems."
1. The "Fresh Grad" Syndrome
The AI analyzed my code like a brilliant university graduate who just read Clean Code for the first time. It applied a universal, theoretical standard to a specific, historical reality.
- The AI sees: "This code is ugly and violates the Single Responsibility Principle."
- The Human sees: "This code is ugly, but it has handled millions of transactions without crashing for five years. Refactoring it introduces risk with zero business ROI."
The AI identified Technical Debt. It failed to identify Technical Asset.
2. The Difference Between "Broken" and "Benign"
This is the blind spot of Large Language Models.
- Agents are Risk Scanners: They are great at listing every theoretical vulnerability.
- Humans are Risk Underwriters: We know why we tolerate certain risks.
This project has "bad architecture" because different teams added features over a decade to meet urgent client deadlines. That "bad architecture" is the scar tissue of survival. The AI doesn't know that the "high coupling" in Module B exists because decoupling it would break a legacy API that our biggest client still uses.
The lesson: Not all "problems" are equal. Some are cancers that will kill the product; others are just benign cysts that look ugly but hurt no one. The AI cannot tell the difference.
3. The New Human Core Competency: Value Judgment
This experiment clarified the true role of the human in an AI workflow. We are not here to generate lists of issues. We are here to make the Value Judgment.
An Agent can tell you: "Here are 50 ways to improve this code." Only a Human can tell you: "Ignore 48 of them. Fix #49 because it's a security hole. Fix #50 because it blocks the new feature."
The Agent expands your peripheral vision. It ensures you don't miss anything due to ignorance. But the decision to Act or Ignore is a function of:
- Historical Context ("We tried that in 2019, it failed.")
- Business Context ("We are pivoting away from this module anyway.")
- Accountability ("If this refactor breaks production, I get fired, not the AI.")
4. The Agent is a Lever, Not a Leader
If you blindly follow the AI's advice, you will spend the next six months refactoring a working system into a "perfect" system that generates the exact same revenue. You will be "busy," but you will not be "productive."
How to use Agents correctly:
- Use them for Discovery: Let them find the cracks you missed.
- Use them for Options: Let them suggest "theoretical improvements."
- Keep the Gavel: You are the judge. You decide what goes into the backlog and what goes into the trash.
Conclusion: The Power of "No"
AI does not make professional experience obsolete. It makes it more critical. In a world where an AI can generate infinite "To-Do" lists, the most successful leaders will be the ones with the strongest "To-Don't" lists.
If you find yourself arguing with an AI about why a "problem" isn't actually a problem, congratulations. You aren't resisting progress. You are doing the one thing the machine cannot do: Contextual Prioritization.
Mercury Technology Solutions: Accelerate Digitality.
