I've been observing how companies approach artificial intelligence for a few months now, and there is a pattern repeating so often that I can no longer ignore it: many people are running in a direction that won't give them the results they expect.
I'm not saying this from a place of superiority. I'm saying it from the trenches, looking at job offers, conversations with founders, and strategic plans that mix ambition with a rather fuzzy idea of what AI can actually do for a real business today. And I think it's worth sorting out what I'm seeing, because the decisions made now will shape the coming years.
The Mirage of Having Your Own Model
I find the first symptom in how hiring is done. I see companies looking for profiles to "adopt AI," and upon reading the offer, many ask for experience training LLMs or hint at a specific fantasy: that of magically launching their own artificial intelligence model.
Here we need to be honest. Training your own model from scratch is not a one-quarter project with a couple of motivated engineers. We are talking about investing, at least, a hundred million dollars just for the training, not counting specialized talent, infrastructure, and the cost of making mistakes along the way.
And the reality is that almost no company needs that to leverage their business growth with AI. Confusing "using AI" with "making AI" is like thinking that to have your own website you must first build your own browser. The layer that actually moves the needle for the business is much closer, and much cheaper.
The Dream of Autonomous Agents
The second symptom is the opposite, but equally problematic. Other companies hear about AI agents and imagine completely autonomous robots doing a person's job from start to finish.
Let me be clear: I'm not saying the future won't lead us there. It probably will. But there is still a long way to go, and building today's strategy on a capability that isn't reliable yet is the fastest way to burn budget and internal trust.
Agents do work, but they work within limits: tasks with clear boundaries, human supervision, and processes where an error is not catastrophic. The difference between "an agent that saves me hours" and "an agent that replaces my team" is not one of degree; it is one of nature. And mixing them in a strategic plan generates expectations that no one can deliver on.
So, What Does Make Sense?
If both extremes—the custom model and the autonomous robot—are traps, where do you start? This is what I truly believe today:
You Probably Don't Need Your Own LLM
I repeat it because it is the most expensive mistake. Before thinking about training anything, squeeze what already exists. Commercial models are a very powerful commodity today, and the value for your business is not in the model, it is in how you connect it to your processes and your data.
Leverage Automation Where AI Excels
Here lies the true short-term return. There are dozens of processes in your company—support, documentation, data analysis, draft generation, classification, summaries—that AI does faster and cheaper today. It is not glamorous, but it pays the bills and frees up your people's time.
Provide Access and Training to Your Employees
It is useless to have the best tools if your team does not know how to use them day-to-day. The real advantage is not determined by the tool, but by the fluency with which your people integrate it into their work. Access, training, and permission to experiment. That is the investment with the best cost-benefit ratio you will find.
Plan Costs, and Assume They Will Rise
You have to plan costs from the start, and be very clear that they tend to grow for at least a few years. As you integrate AI into more processes, consumption scales. This is not a problem; it is a feature of the model. But if you don't budget for it, you'll be in for a surprise.
Evaluate Headcount vs. Token Cost
This is the uncomfortable decision that almost no one wants to look at head-on. In many contexts, it is still more advantageous to pay juniors than to bet 100% on AI. The key is to diversify, as in any investment: don't put all your budget into tokens or all into heads. The balance depends on your business, your margin, and your phase, but betting on the absolute—in either direction—usually turns out to be expensive.
Blitzscaling, as We Knew It, No Longer Fits
And this point is what makes me think the most. Blitzscaling in scaleups as we understood it about three years ago no longer fits in this AI world. The logic of growing at any cost, burning cash to win the market before anyone else, clashes with a scenario where efficiency is king again and where the tools you have today allow you to do more with less. The rules of the scaling game have changed, and it is wise to accept this before designing the plan for the next two years.
What If You Really Do Need Your Own LLM?
To be clear, it's not a "never." There are specific businesses where a custom model or serious fine-tuning does make sense: highly specific data, particular regulations, or a competitive advantage that lives precisely in that layer.
If that is your case, you are not alone: both Nvidia and AWS offer services designed to help you train your models without having to build all the infrastructure yourself from scratch. But arrive at that conclusion out of actual business need, not the fantasy of having "your own AI."
Truths Are Changing, and That Is Fine
If there's one thing I'm clear about, it's that we are in a moment where truths change constantly, and best practices do too. What I write today might be outdated in six months, and that's okay.
Therefore, my only recommendation is this: do not take anything as an immutable maxim. Stay open to changes and willing to correct course when the data tells you that you were wrong. In technology, ending up clinging to an old certainty is much more expensive than rectifying in time.
And you, what are you seeing in your companies? Do you recognize any of these patterns, or are you finding different paths to integrate AI without falling into extremes?
