World 1-1: where it all begins

Décio Dalke Jr. — Arquiteto de negócios. Sócio-gerente no ecossistema MitUP (mitup.pt).

World 1 of AI series — level 1 of 5: World 1-1 · World 1-2 · World 1-3 · World 1-4 · World 1-5

Anyone who played Super Mario remembers the first level. You press the button, Mario appears on the left edge of the screen and, a few steps later, a brown mushroom with an unfriendly face comes walking towards you. Nobody explained anything. But the level was designed so you’d discover, right there, that you can jump. And that jumping on top of it solves the problem.

Nobody starts the game in Bowser’s castle.

With AI, the castle is exactly where most companies start. They open a chat, type two lines and expect the solution to come out the other end. When it doesn’t, the conclusion is usually one of two: “AI isn’t ready yet” or “AI is amazing” — depending on how impressive the answer was at that moment. Neither is right. And both come from the same mistake: skipping World 1-1.

This piece is World 1-1 of the AI world on this site (the logic of worlds and levels is in The Super Mario logic (in Portuguese)). There’s no tough enemy here. There are the rules that will make everything else make sense.

In a company, AI is an assistant: it can work as a junior or as a senior, and the person running it decides which. Before you pick any tool, two questions settle almost everything: what do you want to use AI for, and how will you check what it delivers?

So what is AI, really?

An assistant. That’s all (and that’s a lot).

It can be a junior assistant or a senior one. The model doesn’t decide that, the vendor doesn’t, the subscription price doesn’t. The user does.

Imagine having Beethoven sitting next to you. Talent isn’t the issue. But he’ll compose the music the way you ask for it. Ask badly, and you get a masterpiece on the wrong subject — or something mediocre on the right one. The quality of the output depends on the quality of the request. That holds for the genius, and it holds for AI.

And there’s the other side. You don’t hand the running of the company to the intern, however brilliant, however many books they’ve read. They’re still an assistant. They can do a lot, and do it well, but the decision and the responsibility stay with whoever signs. AI is the same: it knows more than any of us. A mandate, it doesn’t have.

What do you want to use it for?

Here’s where the first confusion lives. “Using AI” isn’t one thing. I see at least four different uses, and all of them can be valid:

• Auxiliary brain. AI helps you organise topics, track open items, cross-reference information and prepare decisions. You still decide; it carries the weight of memory and housekeeping.
• Outsourcing decisions. You let AI decide: which email is urgent, which proposal passes screening, which text goes out.
• Operational executor. AI does the work: fills in, classifies, writes, moves files, builds reports.
• Guru. You ask, it answers. Research, explanation, second opinion.

Each of these uses calls for a different operating method. The guru requires you to know how to judge the answer. The executor requires rules, limits and someone checking the result. If you outsource decisions, you need written criteria — otherwise you’re outsourcing the criteria too, without noticing. And the auxiliary brain needs a structure where information lives and doesn’t get lost from one conversation to the next.

Opening a chat, starting to talk and expecting it to work is wishful thinking. It might work once. It won’t hold up as an operation.

In my own operation, this became clear when the same AI started serving two roles. As a manager, I use AI to draft, analyse and review contracts, build plans and handle complex matters spread across several tools. On the development side, the work is different: running code, analysing logs, implementing architecture decisions. Each role has different criteria and requirements.

I had a single set of instructions for everything. It kept growing, became bloated and generic — and was never going to cover the quirks of each role. I started noticing the AI was getting lost with so many instructions piled together. Not to mention the cost: every instruction that was useless for the task at hand got loaded, and paid for, every single time (World 1-2 explains why). I split the method by role. Each one carries only what it needs.

How do you know it’s right?

This is the question that separates people who use AI from people who get used by it.

AI having every tool doesn’t mean it knows how to use them. It’s like handing a complete toolbox to someone who has spent their whole life in an office and asking them to fix a machine or put up a wall. You need to know the job, the tool, what the result should look like — and how to tell whether it came out right. AI has plenty of theory. The application, and the checking, stay with whoever runs it.

There’s an image I use a lot. AI lets you move across surfaces you haven’t mastered. You don’t need to know how to swim in deep water: it’s the boat that takes you out there and does the heavy lifting. But if you’ve never left the beach, can’t swim and don’t know how open-sea fishing works, you risk doing something stupid without even knowing what, or why. You can’t correct what you can’t evaluate. People who can swim in deep water stay calm in the boat. People who have never taken their feet off dry land shouldn’t start on the high seas.

So before asking which tool to use, the question is a different one: am I able to measure, gauge and verify what it gives me? If the answer is no, the problem isn’t the AI.

An example of my own, a very homely one. I asked an assistant for a summary of my running workouts, using data from my watch. Back came a confident sentence: the longest run on record was 10 km. It was wrong. Across the 92 days of data, the longest was 15 km. The assistant didn’t “lie” — it calculated from a recent, partial table and wrote “the longest” as if it had looked at everything. The summary’s main recommendation changed with the right number. Since then, any sentence with “the biggest”, “the smallest”, “never” or “always” only goes out after the calculation has been run over the full series.

Checking also means knowing how to ask for the check. I was reviewing a contract proposed by a client and asked the AI for a review, without saying whose side it was on. It picked one on its own: it analysed the contract from the counterparty’s point of view and suggested changes that weakened our side. My manual review is what caught the imbalance. I asked, and the agent replied that it had done the analysis as the other side’s lawyer. That was it.

The AI hadn’t got the law wrong. It had picked a side without saying so. Since then, legal analysis has fixed roles: a lawyer on our side, a lawyer for the opposing party and an impartial judge. All three read the same document, each with their interest declared.

Notice that, in both cases, nothing was a flaw in the tool. The assistant answered with what was in front of it — a slice of the data, a request with no side defined. And that’s exactly what World 1-2 will open up: what’s in front of it, and why that changes everything.

So where do you start?

You don’t need to wait until you reach some very high level before you start. You need structure, organisation and awareness — but it’s by using it, spending little and paying close attention to the result, that you gain experience and maturity.

World 1-1 in Super Mario isn’t easy because the game underestimates you. It’s easy because it teaches you the rules that will save you later, in the castle.

With AI, the World 1-1 rules are two: know what you’re using it for, and know how you’ll check what you get. The rest is gameplay.

AI isn’t born junior or senior. It’s only as good as whoever runs it.

On to World 1-2?

Quick questions

What is AI, for a company? An assistant. It has access to more general knowledge than any person, but it doesn’t know your company and has no mandate: the decision and the responsibility stay with whoever signs.

What are the possible uses of AI in a company? Four: auxiliary brain, outsourcing decisions, operational executor and guru (consultation). Each one calls for a different operating method.

How do you know whether the AI’s answer is right? By being able to check before you use it: complete data, written criteria and, when there are sides at stake, defined roles. If you can’t evaluate the result, the problem isn’t the AI.