I have spent thirty years building digital infrastructure, and there is one mistake now repeating itself in almost every company I visit. It is not a mistake about tools. It is a mistake about order.

Every week someone proudly shows me a screen. Three artificial intelligences running at once: one answers customer messages, another prepares the Instagram posts, another helps write the Friday reports.

Then I ask a single question: do these three talk to each other?

And that is where the demo ends.

The crossroads: licences or architecture

If your AI strategy consists of handing out licences across departments, you are not innovating. You are automating chaos.

There is an enormous difference between having artificial intelligence in your company and being an AI-Native Company. The first is a recurring bill for disconnected apps. The second is a different topology, where intelligence sits at the centre and the tools hang from it.

Almost everyone is doing the first and believing they are doing the second.

Mistake one: installing apps without touching the core

One app generates your images. Another translates your emails. Another answers messages.

It is exactly the same mistake you made fifteen years ago when you outsourced communications to an external agency without ever sitting down to explain how your business actually works. The material came back correct: no typos, well laid out, delivered on time. And completely alien to you.

An AI with no access to your context is not an extension of your brand. It is a very fast external supplier.

Before you switch on a single tool, you need a central brain that everything else hangs from. In technical terms: a vectorised knowledge base built from your real history — quotes, emails, complaints, manuals, the specific way you say things — reachable through RAG by any agent acting on your behalf.

Without that step, everything you build on top is decoration.

Mistake two: nobody is watching the dashboard

The second mistake is quieter. Agents get installed to do specific jobs, and there is nobody above them governing what they do.

This has a name: Shadow AI. Tools running inside your company, making decisions and talking to your customers, outside any architecture and outside any control. When an agent starts quoting old prices or promising impossible delivery dates, no alarm goes off, because no alarm was ever designed.

Automation without supervision is not autonomy. It is abandonment.

In a properly built architecture this is solved with orchestration: an n8n layer on top, with logs, alerts and a human on the loop reviewing output before it reaches the customer. It is not sophisticated. It is simply doing the job properly.

Mistake three: starting from the bottom

And this is the expensive one.

When you start by automating the tasks at the base, you have to train every tool separately. To each one you explain the tone, the customer type, the product range and the prices. One by one.

The day you change a price, you have to go and change it everywhere. The day you reposition the brand, everywhere again.

But the real problem is not maintenance. It is that those tools do not talk to each other. Each one learned your company at a different moment and has been drifting on its own ever since. So the tone of your social media slowly pulls away from the tone of your customer support, and both pull away from the reports you sign.

Nothing breaks. It is simply that every day your company looks a little less like itself.

A Formula 1 engine on a wooden cart

This is the image I always use, because it is exactly what happens.

You are mounting state-of-the-art models on inherited processes nobody has touched in fifteen years. The result is not efficiency. It is a high-speed crash, only in slow motion and spread across several quarters of your accounts.

AI does not fix a bad process. It runs it faster.

How I would do it: Step Zero is the boardroom

If I had to start from scratch in a company tomorrow, this would be the order. And it does not start with the messaging app.

One. Infrastructure before applications. Your own VPS: secure, scalable, properly structured. Your data in your house, not scattered across fourteen SaaS products.

Two. Load the memory. The whole company vectorised and well classified: contracts, emails, minutes, manuals, commercial history. This is the brain. Everything else is hands.

Three. Agents for the boardroom, not for the base. A real advisory board for the general manager. Imagine being able to sit Steve Jobs, Peter Drucker and Henry Ford down to analyse your company. In real life that board does not exist and no small business could afford it. With AI it exists and it is affordable.

Four. Put them to debate. Not answering isolated questions: arguing with each other about your product, your market and your structure, until a conclusion emerges that you would never have reached alone.

Five. With that strategy on the table, transform the company from the top down. From executors to directors.

Six. And here comes the part nobody expects: half the automations you would have installed at the start are no longer needed. Not because they work better. Because they have stopped making sense.

At PerfilesIA I have it built exactly this way: thirteen specialised agents, each with its own function, hanging from a single memory and a single governance layer. It is not a demo, it is a company running. When I say this can be done, I am not speaking from theory.

So what did you want the messaging bot for?

Here is the argument that closes the door.

Why do you want an AI answering your customers on WhatsApp, if before long the customer will not be writing to you at all? Their AI will. And it will want to talk to yours: request data, compare terms, negotiate deadlines, close.

I do not have the exact date, but my bet is that system-to-system communication will arrive considerably sooner than most people calculate. And the day it does, the entire layer you are automating today to speak with humans loses its purpose.

That expiry date is not yours to set.

What I want you to take away

Do not automate a company that is going to become obsolete.

Blow up the inherited processes and build the new topology. Start at the top, because the top is where you decide what is worth automating and what should simply be removed.

When you get it right, it is not the processes that become automatic. It is the company.

Your first AI application: did you install it at the top or at the bottom?