Over the last two years we have heard the same speech:
“AI will replace millions of jobs.”
“Companies will become much more efficient.”
“You will be able to do the work of two or three people.”
And, being honest, part of that is already happening. I myself would not have been able to build NeuraPRO in the time I did without Claude, Cursor, GitHub Copilot, and ChatGPT. Today I use AI every day in a multinational company to automate processes, develop internal tools, and manage projects. In addition, I have spent months studying how it works internally: agents, MCP, Skills, and Subagents.
The more I learn, the more convinced I am that we are living through a technological revolution comparable to the arrival of the Internet. But I am also starting to believe that we are looking at the problem from the wrong angle.
It is not capability. It is not security. It is not model quality.
It is something much simpler:
Who pays the bill?
When AI stops being a demo
A few months ago we had practically free access to Claude at my work. And exactly what one would expect happened: we started building. A lot.
In my case I developed a Planner fully integrated with Jira, Confluence, and our internal portal. Today I can create complete projects, maintain dependencies between tasks, reorganize calendars without breaking the plan, assign engineers, create tasks automatically, update Jira with one click, calculate effort, measure overall progress, and export the plan to share it with other teams.
It still has a lot to improve, but it already solves real problems. And it will probably end up being used far beyond my team.
Then everything changed
As the weeks went by, limits started to appear: internal talks, best practices, emails recommending prompt optimization, using cheaper models whenever possible, and thinking before launching complex queries.
And that is when I understood something.
It was not a technical problem. AI kept working exactly the same. It was not a security problem.
It was a cost problem.
And I think many companies did exactly the same thing: first they allowed teams to experiment freely and then they analyzed the data. How much was used? How much time was really saved? How much did it cost?
If I had had to make that decision as a technology leader, I probably would have done exactly the same thing. Because there is no better way to understand a technology than seeing it working in production.
It is not just a perception
My experience does not seem to be an isolated case.
Over the last few months, several reports have appeared showing that many companies are starting to control AI consumption much more closely. Some studies indicate that more and more organizations are imposing budgets, usage quotas, or promoting cheaper models because operating costs are growing much faster than expected.
At the same time, companies like Microsoft, Google, Meta, OpenAI, and Anthropic continue to announce investments of tens of billions of dollars in infrastructure to support the growing demand for AI models.
Technology is advancing, but making it run remains extraordinarily expensive.
Let us do a simple exercise
Imagine a company with 5,000 employees. Not all of them develop software; many simply use AI for their daily work: reading documentation, preparing presentations, summarizing meetings, analyzing contracts, consulting internal procedures, writing code, or replying to emails.
Now imagine that each person makes only 20 relevant queries per day. Not simple questions, but queries where AI must read documentation, understand Jira tickets, review Confluence pages, analyze files, use tools through MCP, or reason before answering.
Each one consumes context, processing, and tokens.
Multiply that by five thousand people, then by twenty working days, and then by twelve months.
Suddenly AI stops being simply a productivity tool. It becomes a new expense line that can represent millions of dollars per year.
And that bill keeps growing as we use increasingly sophisticated agents.
Agents are impressive…
…but they also consume a lot.
Over the last few months, I have been studying in more depth how modern agents work, not only from a practical point of view, but also from what happens behind the scenes.
An agent does not make a single query: it reads documentation, consults Jira, searches for information in Confluence, invokes tools through MCP, consults other specialized agents, reasons, queries again, generates tasks, and validates results.
Each of those steps consumes tokens again. Many more than most people imagine.
As an engineer, I find it incredible.
As a Project Manager, I inevitably think something else:
Is it sustainable when an entire organization starts working like this every day?
Is it really cheaper than a person?
This is probably the question I have been asking myself the most lately.
For a long time I heard an idea that seemed obvious:
“It will be much cheaper to replace people with AI.”
Today I am no longer so sure.
Not because AI is not capable —it is—, but because the doubt is somewhere else.
If an experienced engineer solves a task in one hour, the cost is known. If an agent needs to analyze hundreds of pages of documentation, go through tickets, consult several tools, reason for several minutes, and generate an output, the cost also exists. And we are still learning which of the two models is more profitable depending on the type of work.
I think that is where the real debate is: not only asking what AI can do, but how much it costs to do it millions of times every day.
The biggest lesson all this left me
Curiously, the best result I got was not asking more questions, but no longer needing them.
Building my Planner consumed a lot of tokens, but now that it exists, much of that cost disappeared. The logic was implemented. I no longer need to constantly ask a model how to reorganize tasks or calculate dependencies: the system simply does it.
And that completely changed the way I see AI.
Maybe the greatest value is not in talking endlessly with a model.
Maybe it is in using that intelligence to build tools that later work almost by themselves.
My reflection
I do not think we are living through a technology bubble. AI works. I see it every day. I used it to build NeuraPRO, I use it in my work, and I keep studying it because I am convinced it will change the way we work.
But I do believe we are living through a stage where expectations are moving faster than reality. Today we talk a lot about replacing people and very little about costs, return on investment, or sustainability.
And I increasingly wonder how human work will evolve in this new scenario. How we will find the balance between productivity, technology, costs, and people.
Because one thing is proving that AI can do a job. Another very different thing is proving that it can do it every day, for thousands of people, for years, at a cost a company is willing to assume.
Maybe that is why we are seeing more and more organizations promoting smaller models, optimizing prompts, establishing budgets, and teaching people to spend fewer tokens.
Not because AI has failed.
Quite the opposite.
Because for the first time it stopped being a demonstration and started being used for real.
The next AI revolution will no longer depend only on creating smarter models.
It will depend on making that intelligence economically sustainable.
And that will probably be the most important innovation of all.
Related reading:
- NeuraPRO: My MVP Worked. My Customers Didn’t.
- When Jira Is Not Enough
- Are We Entering a Silent Labor Crisis?
✍️ Claudio from ViaMind
“Dare to imagine, create and transform.”
FEATURED POSTS
ABOUT ME
ViaMind Ecosystem
- ViaMind — The brand hub for ideas, technology and strategy.
- NeuraPRO — Practical software for managing real operations.