A great code bloat is arising as AI turns managers into software programmers

In boardrooms, realization has dawned that natural language can now program computers. Anthropic’s Claude has emerged as the darling of the non-technical professional, praised for its ability to transform a vague business requirement into a functional application.

The executive who once waited months for a dashboard can now conjure one in an afternoon. This shift expands human agency. It also introduces a new category of risk.

I recently watched a chief operating officer with a terrifyingly efficient grasp of logistics and a newfound affection for AI models build a global inventory tracker in an afternoon. It was a masterpiece of ‘vibe coding.’ It looked and felt right. To her, it was a triumph over the company’s stodgy information technology department.

Monkey see, monkey do; and so, her colleagues are following. The sales boss has started assembling a client portal. Soon, the organization will have multiple new systems that solve immediate problems while quietly refusing to acknowledge each other’s existence.

Such tools, while functional, are digital black boxes. It is the software equivalent of a modern mansion with the wiring done by an enthusiastic poltergeist. The lights come on when you enter, but don’t you dare touch the walls.

When humans painstakingly author code, there is usually an underlying mental map that allows for future troubleshooting. I have written before of an Anglo-Indian musician named Desmond, long since emigrated to Australia, whom I worked with decades ago and who approached Cobol with the same rhythmic precision he gave to his verses; every line had a reason.

Even imperfect human systems carry traces of intent. Not so with AI-generated code.

A great code bloat is taking birth in the minds of a million managers. As every employee becomes a casual builder of software, corporate networks are being populated by thousands of small, fragile scripts that lack written requirements, documentation, security protocols and any semblance of a unified architecture. It may or may not scale. But it will surely decay. Fast.

The fundamental agency cost here, as my old professors Jensen and Meckling would have noticed, is not that the machine is unreliable but that it has no sense of scope. It will solve the prompt you give it today while quietly and unwittingly setting up a catastrophic failure three months from now.

The immediate consequence is not failure, but fragility. But when the systems do break, diagnosis becomes slow and uncertain. Responsibility is diffused across undisciplined ideation, prompts, models and partial fixes. No one holds a complete map of the system. Such programs are unexploded digital ordinance that could erase balance sheets.

This is where the Indian IT services industry re-enters the story, though not in the role it would prefer. For decades, these firms built their empires on labour arbitrage, providing the heavy lifting of manual coding. The future may involve the digital equivalent of forensic plumbing, going in to discover why a system built by marketing has suddenly stopped recognizing the existence of the euro. The consultant becomes a plumber with a computer science degree.

This is a vastly different business than Indian IT is used to. It is about assurance, not effort. Clients will not pay for more code; they will pay for confidence that their systems behave predictably. Trust might become the only billable commodity.

It’s possible that the insurance industry becomes the biggest buyer since enterprises will no doubt look to cover the risk of AI failures, and insurers will certainly find a way to cover them while reducing premiums for those that have efficient human plumbers working against AI blockages.

In the pre-AI era, a senior developer would scrutinize every line of code to ensure it was elegant and secure. That process is under pressure. If the volume of code grows tenfold because everyone is a coder, the human capacity for review cannot keep pace. Sampling will replace scrutiny, and assumptions, certainty.

Some say we need AI to check AI, but this assumes that the second model possesses a higher degree of truth than the first. Not so. What it often leads to instead is polite agreement between two black boxes to ignore the same structural flaws. Without humans who understand both the why and the how, the enterprise loses its ability to recover from failure.

When one looks inside IT service firms, one finds the same problem. The engineers providing oversight are themselves increasingly reliant on the same AI tools. The industry risks creating a feedback loop where machines generate systems that humans supervise with the help of other machines.

Separately, the risk for IT service firms is that they may inadvertently participate in their own obsolescence by prioritizing volume over verification. If they merely use AI to churn out code faster, they contribute to the very bloat that will eventually bankrupt the technical debt of their clients.

The winners will be those that recognize why the product is no longer the code itself, but the guarantee of its utility. This requires a shift from code as craft to code as consequence.

We are all coders now. God help us.

The author is a technology consultant and venture capitalist.