Why is everyone suddenly complaining about the cost of AI?
There is a story people tell about British-ruled India, where poisonous cobras had become a serious problem.
The authorities came up with what seemed like a simple solution: pay people a bounty for every dead cobra they brought in.
At first, the plan appeared to work. People started killing cobras, turning them in, and collecting the reward. The metric was clean, simple, and comforting: more dead cobras meant fewer live cobras.
Then someone figured out the game.
If the government was willing to pay for dead cobras, breeding cobras could become a pretty good business.
Once the authorities realized what was happening, they canceled the bounty. At that point, the breeders released the snakes, which were now worthless.
In the end, according to the story, there were more cobras than before.
It is a perfect story because it captures something organizations do all the time: they measure something, turn that measurement into a target, and then act surprised when people optimize for the target instead of the outcome.
Every company has its cobras.
Bugs created so they can later be fixed. Tickets closed without actually solving the problem. Meetings multiplied to prove alignment. Miles driven mistaken for sales performance. Documents produced mistaken for thinking.
And now, in the age of artificial intelligence, we have found a new cobra to count: tokens.
Before going any further, let me make one thing clear.
I know not everyone will like what I am about to say. I believe that truly understanding artificial intelligence requires a level of expertise that very few people in the world have. Maybe a few thousand. I am not one of them.
I belong to a different group: people who observe, ask questions, try to understand where things are going, and attempt to offer answers. Partial answers. Imperfect answers. Debatable answers. But, hopefully, useful ones.
Over the last few months, something subtle but important has been changing: the way major AI providers count tokens and subsidize their consumption.
For a while, the goal of the big AI providers was obvious: get us to try AI, get us to use it, and weave it into our daily work. In a sense, make us dependent on it.
I do not mean “dependent” in a conspiratorial way. The dynamic is much simpler.
AI works.
In many cases, it works extremely well.
Once you start using it seriously, it is hard to go back. But first, you need to try it. And to get you to try it, someone has to absorb part of the cost.
So far, fair enough.
The problem starts when companies adopt AI without really knowing how to measure its impact. So they choose the easiest metric available: the number of tokens consumed.
It is an understandable temptation.
Tokens can be counted.
Tokens fit nicely into reports.
Tokens make charts.
Tokens make dashboards happy.
But measuring AI adoption by token consumption is like measuring a manager’s effectiveness by how many meetings they attend. Or judging a salesperson by how much gas they burn. Or evaluating a student’s intelligence by how many pages they highlight.
The metric is real. It just does not mean what we think it means.
Then, when the bills start showing up, everyone starts scratching their head.
Is AI really making us more efficient?
Are the costs justified?
Is it worth it?
Are we spending too much?
Should we cut usage?
Those are fair questions, but they are often framed the wrong way. Because they start from the wrong metric.
The real question is not how many tokens we consume.
The real question is what those tokens produce.
A useful answer?
A better decision?
An error avoided?
A process accelerated?
An employee enabled to do better work?
A workflow simplified?
A customer served faster?
Those are the hard questions.
Which is exactly why we tend to avoid them.
Measuring what matters is hard. Measuring what is easy is comforting. The problem is that organizations, like people, often confuse what can be measured with what is actually important.
This brings us back to one of the oldest laws in computing, one everybody knows and everybody forgets: garbage in, garbage out.
Put garbage in, get garbage out.
We can dress it up. We can make it sound polished. We can ask it to use a professional, empathetic, visionary, or strategic tone. We can run it through the most powerful model available.
But if the input is confusion, AI does not produce clarity.
It produces well-written confusion.
That may be the most uncomfortable part of the whole conversation.
AI does not pay off our debts. It amplifies them.
If a company has serious technical debt, AI will not magically erase it. If a process is messy, it will stay messy, just faster. If decision-making is fragile, AI will not automatically make it sound. If an idea is bad, cooking it with AI will not make it good.
In fact, it may become more dangerous, because it will look better.
AI has an extraordinary talent: it can make almost anything sound plausible.
Try this experiment. Take a terrible idea and discuss it with an AI assistant. If you push hard enough, phrase your questions carefully enough, and look for confirmation instead of criticism, after a while you may start believing that your terrible idea is actually a breakthrough.
Not because AI is evil.
Because many AI systems are designed to help us, support us, and go along with us. And helping someone does not always mean challenging them.
Sometimes it means giving them a more elegant version of their mistake.
That is why I do not believe AI turns lead into gold.
More often, it turns lead into fool’s gold: something that shines, something that looks valuable, but is not.
Does that mean we should not adopt AI?
Absolutely not.
That would be a huge mistake.
Every major technology has had an immature phase. Think about machine tools in factories. The early ones were powerful and transformative, but also dangerous. They lacked the safety guards we now take for granted. Inexperienced workers got hurt, but so did skilled ones.
The answer was not to abandon machine tools.
The answer was to learn how to use them better. Add safeguards. Define procedures. Train people. Understand where they made sense and where they did not.
AI will be no different.
A smart business leader does not look only at today’s ROI. They look at future ROI. They ask: what does it cost to adopt this technology? But also: what will it cost not to adopt it?
For AI, the question is the same.
The issue is not whether to use it.
The issue is how to use it without getting stupid faster.
So where do we start?
First, with experimentation. But serious experimentation, not theater. Handing out licenses, counting tokens, and declaring a digital transformation underway is not a strategy.
Companies need to choose concrete use cases. Measure concrete results. Understand where AI reduces friction and where it adds noise. Where it speeds up work and where it simply creates more stuff for someone else to review.
Above all, we need to stop expecting AI to pay off our technical and organizational debt.
It will not.
It will make that debt more visible.
Then there is the question of efficiency.
Using AI well does not mean using the most powerful model, with the largest possible context, for every single problem.
That is the corporate equivalent of calling the country’s top surgeon to put on a Band-Aid.
We need to become more thoughtful and more disciplined.
Keep the context tightly focused on the actual problem.
Break complex problems into simpler ones.
Use smaller models for simpler tasks.
Reserve the most powerful models for work that truly requires reasoning, synthesis, evaluation, and orchestration.
In every company, people would love to have the smartest person in the room involved in every meeting, every decision, and every project.
But you quickly learn that this is not possible.
And more importantly, it is not even a good idea.
The same applies to AI.
Finally, we need a new role: the devil’s advocate.
I have to smile when people ask AI: “Do not make mistakes” or “Do not hallucinate.”
That is a bit like telling a person, “Please be smart.”
Good intention. Weak method.
It is much better to build a process for criticism.
Once AI has produced a proposal, another AI should try to tear it apart. Look for weak assumptions. Highlight contradictions. Identify risks. Ask uncomfortable questions.
And no, it is usually not a good idea to do everything in a single pass.
Asking the same model to generate an idea and immediately critique it often leads to a weak, polite, accommodating critique.
Separate the roles.
Better yet, let the devil’s advocate be a different model, with different instructions, and maybe even a different philosophy.
Because the future of AI in business will not be decided by who uses it the most.
It will be decided by who learns to use it best.
And using it better also means measuring it better.
Not tokens.
Not self-reported hours saved.
Not the number of prompts written.
Not the number of documents generated.
But the quality of decisions.
The reduction of errors.
The speed of real processes.
The elimination of useless work.
The ability of people to focus on better problems.
AI is an incredibly powerful technology.
But it is not a religion.
It is not a magic wand.
It is not an organizational debt forgiveness program.
It is a mirror.
If we put confused processes, vague goals, and bad metrics in front of that mirror, the image coming back will be sharper, faster, and more convincing.
But it will still be the same image.
Just with more tokens.
A blog of AI, tech, and mixed stuffs 
Leave a comment