World 1-3: up high

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

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

The third level of Super Mario’s first world pulls the ground out from under you. The stage turns into a string of loose platforms high in the air, with nothing but empty space between them. The higher you climb, the farther you can see. And the longer the fall.

With AI, this is the level where almost everyone asks the wrong question: “which tool should I use?” Chat, agent, automation, the app that came out last week. Tools change every six months, and an article that just lists them ages right along with them. What doesn’t change is something else. Every level you climb, you trade oversight for scale.

There are five levels of AI use in a company: chat; chat with context and tools; AI inside your work tools; an agent on your computer; and automation with no interface. At each level, the AI does more without you watching. The choice isn’t about the tool. It’s about who checks the result, what can break and what a mistake costs.

What changes when AI can reach your files?

I started a piece of work on my phone. I opened a chat, explained what I wanted and asked for help organising it. The answer was good, and it ended with a question: some information was missing, and it lived in documents on my computer. Without it, there was no way to get it right.

Fair question. The phone chat couldn’t reach that far.

I carried on in the desktop app, where the AI has access to my folders. First detail: I had to paste the whole conversation in by hand. The context didn’t travel from one side to the other on its own. Second detail: in the first answer, the question was gone. It went and read the documents, found what was missing, and kept going.

Same AI, same request. What changed was the reach. (And notice that the phone chat did the right thing: data was missing, so it asked instead of guessing. In World 1-2, that had a name: missing data, stop and ask.)

That’s what this level is about. Not which tool is best, but what each level can reach — and what it starts doing without you.

What are the levels?

I see five.

1. Chat. You ask, it answers, you copy and paste. Nothing leaves without going through your hands. If there’s a mistake, you’re standing right in front of it (and that’s where the World 1-1 question comes in: can you check it?).

2. Chat with context and tools. Projects with fixed documents, memory across conversations, web search, connectors to your email, calendar and drive. The AI starts to know your situation without you repeating everything each time. This is where most people who’ve moved past the basics are. The gain is big; the new risk is subtle: it now answers based on things you didn’t read at that moment.

3. AI inside your work tools. Copilot in Word and Excel, Gemini in Gmail, Claude in Excel. The AI goes to the work, instead of the work going to it. It looks like level 1 in different clothes, and it isn’t. The output lands straight in the document, the spreadsheet, the email draft. The copy-and-paste moment disappears — and that was, without anyone noticing, the moment of review.

4. An agent on your computer or in your browser. You hand over a task and it carries out several steps: opens files, edits them, runs commands, browses websites. The desktop app from that story lives here, and so do coding agents. You don’t approve each step; you approve the request and check the result.

The hardest lesson at this level, in my own operation, came from a misread permission. I had given the agent autonomy to work on my files. Hours later, it took that as licence to take over my computer’s mouse and keyboard: it opened windows on top of mine and ran a script, without asking. The rule was written the same day. Working on files is one permission. Controlling the machine is another, and it needs confirmation every single time, no exceptions.

5. Automation with no interface. n8n, Make, direct API calls, agents that fire on their own on a schedule or on an event. Nobody is watching when it runs. Oddly enough, this often isn’t the most “intelligent” level: it’s a fixed workflow, designed by someone, with an AI step in the middle. The difference isn’t intelligence. It’s that nobody sees it. And a mistake here doesn’t happen once: it repeats on every run, until someone notices.

In my operation, the proof came from a WhatsApp number I have connected to Claude. I use it to capture ideas, social media links and the voice notes I send while I’m out running. An engine processes each item, with nobody watching. One day I sent a link, and the processing agent decided on its own to use a paid agent to extract the page content. Nobody had asked it to. I found out because the processing summary mentioned, quite casually, that the paid agent had been used.

It wasn’t luck that it told me. I built the summary to account for what the engine did, precisely because nobody is watching when it runs. That’s what let me see it. From then on, the paid agent was taken out of its reach, and the process was moved behind an autonomy gate: a layer that decides what the agent can do on its own and what needs authorisation. The agent didn’t do anything absurd. It did what seemed reasonable to get the job done. The problem is that nobody had told it where reasonable stopped.

Two things were left out on purpose. Talking to AI by voice, or sending it a photo, isn’t a level: it’s a way of giving input, and it cuts across all of them. And running a model on your own computer is another conversation, one for World 1-5.

So which level should you use?

The temptation is to treat the ladder like a career: start with chat and keep “moving up” to automation, as if the top were the goal. It isn’t. Each level suits a type of work, and climbing without need means paying for risk and getting nothing back.

The choice comes down to three questions:

• Who checks the result, and when? At level 1, you, on the spot. At level 5, nobody on the spot — so the checking has to be designed in advance.
• What can this break? A wrong draft gets deleted. A sent email, an overwritten file, a posted payment — those don’t.
• What does a mistake cost, and how many times does it repeat before anyone sees it? In a chat, it’s one mistake. In automation, it’s one per run.

Drafting a difficult reply to a client calls for level 1 or 3: the AI proposes, you decide every word. Organising a folder of two hundred documents fits level 4: you check the result at the end, and it can be undone. Classifying what comes in every day, with nobody available to look, is level 5 — and it only works if someone decided, beforehand, how to know the classification is still right.

Notice that the World 1-1 question hasn’t gone away. It’s just moved. Down low, you check afterwards. Up high, you have to decide how you’ll check before you jump.

Up high, Mario doesn’t fall because he jumped badly. He falls because he jumped before seeing where he’d land.

Quick questions

What are the levels of AI use in a company? Five: chat; chat with context and tools; AI inside your work tools (such as Word, Excel and Gmail); an agent on your computer or in your browser; and automation with no interface.

What’s the difference between a chat and an agent? In a chat, every answer goes through your hands before it becomes work. An agent carries out several steps on its own (opens files, edits, browses) and you check the result at the end.

Is AI automation smarter than a chat? Not necessarily. It’s often a fixed workflow with an AI step in the middle. The difference is that nobody is watching when it runs, which is why the checking has to be designed in advance.