If I were the CEO of a software company today, I could probably lay off 70–90% of my junior developers. AI would write the code. AI would open the pull requests. AI would also handle much of the testing.
But the real question is something else entirely: who will still understand the software five years from now?
In 2023, the question we were asking was: can we replace humans with AI?
In 2026, the question is becoming: can we afford to replace humans with AI?
These are two very different questions. The first concerns technical capability. The second concerns the final bill. And that bill includes items that almost nobody has accounted for so far: token costs, knowledge-loss costs, technological dependency costs, maintenance costs for AI-generated code, and cognitive costs associated with the erosion of skills.
Five categories. None of which appear in an AI-first pitch deck.
This article examines them one by one.
Vibe Coding and the Paradox of Lost Documentation
Vibe coding works. In fact, it works surprisingly well. You describe what you want, AI generates the code, and the code runs. The term was coined by Andrej Karpathy in early 2025 and, within a few months, became an everyday practice for thousands of teams.
The problem is not generating code. The problem is understanding that code six months later.
More and more companies are accumulating millions of lines of software produced by Claude Code, Cursor, GPT, and Gemini. But an increasing proportion of that software has been generated without anyone truly following the reasoning that led to the solution.
The code exists. It works. But the knowledge behind its creation often does not.
In the past, a developer remembered why a function had been written in a particular way, which trade-offs had been consciously accepted, and which bugs had already been encountered and avoided elsewhere in the system. That memory lived in people's minds even before it appeared in commits.
Today, increasingly, only the final result remains. The prompt that generated it has disappeared into a chat. The reasoning that ruled out three alternative solutions was never documented anywhere.
It is the difference between knowing a route and following Google Maps. You reach your destination either way. But if your phone runs out of battery in a neighborhood you never learned to navigate, you are lost.
Outsourcing the Brain
For decades, London taxi drivers have had to pass an examination called "The Knowledge": a mental map of approximately 25,000 streets and 20,000 points of interest across the city, memorized without any digital assistance.
Neuroscientist Eleanor Maguire spent years studying the brains of people undertaking this challenge. The findings, replicated and refined in subsequent studies, reveal a robust result: those who pass "The Knowledge" develop a measurable increase in the size of the posterior hippocampus, the brain region involved in spatial memory. Those who abandon the training do not.
Then GPS arrived.
The question is not whether GPS is useful. Of course it is. Nobody is seriously suggesting we should go back to navigating entirely from memory.
The question is different: what happens to a cognitive ability when we stop exercising it?
The same mechanism may now emerge with AI, but on a much broader scale than spatial navigation alone. It does not affect just one profession. It affects writing, reasoning, programming, summarizing, and the ability to hold a problem in our minds long enough to understand its structure.
We are not necessarily losing the ability to produce. We may be losing the ability to understand what we produce.
The MIT Study and Cognitive Debt
In 2025, the MIT Media Lab published a study entitled "Your Brain on ChatGPT", which used EEG to measure the brain activity of 55 participants aged between 18 and 25.
Participants were divided into three groups: those writing an essay using ChatGPT, those using a traditional search engine, and those writing without any tools.
The results showed that participants using ChatGPT exhibited the weakest brain connectivity of the three groups, particularly in areas associated with memory and semantic reasoning. Those using a search engine demonstrated intermediate engagement. Participants writing without tools activated the broadest and strongest neural networks, with significant involvement of creative, memory-related, and executive regions.
However, the most widely cited finding from the study was another: during the initial sessions, 83% of participants who had written their essays using ChatGPT could not quote even a single line immediately after submitting them. Participants in the brain-only group remembered their texts accurately.
MIT researchers gave this phenomenon a name: cognitive debt.
Like financial debt, it accumulates almost imperceptibly, loan after loan, delegation after delegation. And like financial debt, sooner or later the bill comes due: less critical thinking, less creativity, weaker memory, and a more distant relationship with one's own work.
This is not entirely new. It has already happened with calculators. It has already happened with GPS. The difference is that this time we are not delegating a single calculation or route, but writing, coding, research, synthesis, and reasoning: almost everything we associate with thinking at work.
The risk is not becoming less productive. In the short term, productivity actually increases dramatically. The risk is becoming less autonomous.
The Madness of Tokenmaxxing
There is a social phenomenon we know well: looksmaxxing, the obsession with optimizing every detail of one's physical appearance until it becomes a status symbol.
Companies are now developing an operational equivalent: tokenmaxxing. The contemporary obsession with maximizing token consumption as though it were, in itself, a measure of progress.
Companies internally celebrating millions of prompts submitted. Millions of lines of code generated in a quarter. Billions of tokens consumed over a year.
Numbers that end up in presentations, internal leaderboards, and adoption targets, as if the marginal cost of every token were irrelevant and sheer volume were already an achievement.
But volume is not the outcome. Volume is the cost of achieving the outcome.
The Uber Case
In 2026, Uber exhausted its entire annual budget for AI development tools, particularly Claude Code and Cursor, in approximately four months.
Among the factors accelerating consumption was an internal leaderboard ranking teams according to how extensively they used AI tools: a mechanism intended to encourage adoption that ultimately encouraged consumption instead.
The most interesting statement in the entire episode is not the amount of money spent. It is what Andrew Macdonald, Uber's President and COO, said when asked to connect that expenditure to features actually delivered to users: that connection did not yet exist.
He added that drawing a direct line between increased spending and features actually released was extremely difficult, and that without such a connection, the trade-off became harder to justify.
CEO Dara Khosrowshahi nevertheless confirmed that approximately 10% of the code committed at Uber is now produced by autonomous agents. Uber subsequently introduced individual spending limits for AI tools.
For years, we measured technical work in person-hours, sprints, velocity, and revenue generated. Now many companies have begun measuring it in tokens consumed.
But tokens are not an output. They are an input.
Measuring success in tokens is like measuring a factory by how much electricity it consumes rather than by how many valid products come off the production line.
The Return of Human Beings
This is perhaps the most counterintuitive part of the entire story: some of the most aggressively AI-native agencies and software companies in the market have started hiring again.
Not because AI does not work. It does, and that is precisely why these companies exist.
But as adoption grows, costs emerge that were not considered during the initial wave of enthusiasm: verification of generated output, systematic quality control, maintenance of code that nobody in the company actually wrote, governance of access to and use of tools, onboarding new employees onto systems that are partly opaque even to those who "built" them, and continuous supervision of processes that tend to drift when left unattended.
Every automation creates new points of fragility. Every autonomous system requires someone who understands its behavior deeply enough to recognize when it stops behaving correctly.
In this scenario, human beings do not disappear. Their role changes. From producing to supervising. From writing to understanding.
The Greatest Risk: Losing Organizational Memory
Ultimately, companies are not made of software. They are made of knowledge.
When someone leaves an organization, that organization does not simply lose a productive resource.
It loses context accumulated over years, experience of what has already failed and why, insights never formalized in any document, knowledge of exceptions that no operating manual describes, and the historical memory of decisions made and the reasons behind them.
If we replace people with generative systems too quickly, we risk discovering the consequences too late, when the raw material that truly mattered was not the code, the documents, or the processes.
It was understanding.
And understanding, unlike code, cannot simply be copied and pasted from a repository.
Conclusion: Why Vibe Coding at Hoxell Does Not Replace Human Supervision
The question is not whether artificial intelligence can write most of the code required to develop a software platform. Today, it can already do so, often at a speed unimaginable for a team composed entirely of human developers.
The real question is whether it is wise to entrust AI with the entire process without someone fully understanding what is being generated, the decisions embedded in the code, and the consequences those decisions may have over time.
This awareness is precisely what shapes the approach we have adopted at Hoxell.
We use vibe coding to accelerate development, experiment more quickly, and reduce the time needed to turn an idea into a concrete feature. But AI-generated code is never treated as a finished product or left without supervision.
Every output is reviewed by people capable of understanding its logic, assessing its quality, and intervening when errors, vulnerabilities, or unforeseen consequences emerge.
AI can write the code, but responsibility for architecture, security, maintenance, and final decisions remains human.
At Hoxell, therefore, vibe coding does not replace technical expertise: it amplifies it. It enables people to work faster without sacrificing the knowledge required to understand and govern what they are building.
Because generating code has become relatively simple. Understanding that code, maintaining it, and taking responsibility for it remains a profoundly human task.
Davide Bernasconi