The levels of AI: three ladders we often confuse

Not long ago, “artificial intelligence” meant HAL 9000, Skynet or a robot with a metallic voice. Today, AI sorts your inbox, suggests a route and helps you write code. It went from science fiction to infrastructure before we had time to name things properly.

That is where much of the noise begins. We use the same term for a spam filter and for the idea of a machine smarter than humanity. Some people predict the apocalypse; others dismiss AI as glorified autocomplete. They may both have a point because they are not talking about the same thing.

And there is something else, which is what interests me most here: plenty of people already use AI every day. They type a question into a chatbot, copy the answer and stop there. That is not nothing, but it is the first step of a much longer staircase.

A better question than “what is AI?” is “which level are we talking about?” There are three different ladders: how you interact with it, how deeply it is embedded in your work and what the technology can actually do. I will start with the two you can climb yourself.

1. Your role in the interaction

This ladder does not measure how advanced the technology is, but how much control you take when using it. It is the easiest one to climb and the one that changes the result the most.

1. Passive consumer. AI recommends, filters and ranks things around you, usually without being visible. You did not choose it.
2. Occasional user. You ask for something and judge the tool by its first answer. You can spot this in one detail: every conversation starts from scratch. The next step is inexpensive: provide context before asking (who you are, who it is for, what works for you and what does not) and ask for two or three versions instead of one.
3. Co-pilot. You provide context, iterate, challenge and verify. That last verb is the one people skip. A model produces text that sounds right, which is not the same as text that is right: it will cheerfully call a method that does not exist in the version of the library you are using, wrapped in otherwise idiomatic code. Whoever signs off on the result is still you. And the better you understand your subject, the more the model gives you, which is the opposite of what most people assume.
4. Solution designer. You decide where the model belongs, define its limits and checks, and plan for failure. Here the question stops being what to ask and becomes how much to delegate: a model that can call tools, read files or run code does far more per turn and fails in more expensive ways.

This is not a moral hierarchy. Moving up costs time and is worthwhile only when the result justifies it. But if you see yourself at step 2, the payoff from becoming a co-pilot can be noticeable within a single afternoon. I wrote about that idea, AI amplifying judgment rather than replacing it, in When the machine does the typing.

2. Maturity of use

The first ladder is about what you do inside a single conversation. This one is about what survives between conversations: how deeply AI is embedded in a job or a business. It applies to one person as well as to a team, although the higher levels need responsibilities shared across an organization.

• Level 0 · Exploration. A few experiments, with no recurring use.
• Level 1 · Ad hoc use. People copy, paste and prompt, starting again each time. It is an easy place to get stuck.
• Level 2 · Repeatable use. Saved instructions and templates, reusable context (your documents, your data, your examples) and clear criteria for when AI is appropriate and when it is not.
• Level 3 · Integration. The model works inside products or processes: it is called from other software, uses company data, executes tasks and someone evaluates whether the outputs are any good.
• Level 4 · AI-native. Processes are designed around AI, including the unglamorous essentials: cost, privacy, traceability, monitoring and fallbacks.

There are two jumps here, not one. The first, from 1 to 2, is the one many people still need to make and the cheapest to make: it mostly consists of no longer discarding the context and lessons from every conversation. Keep what worked, reuse the context, learn which tasks the model genuinely helps with.

The harder jump, and the one I often see teams struggle with, goes from 2 to 3. AI stops being a personal trick and becomes engineering: the team must test, measure and own the outcome. A demo that succeeds eight times out of ten may be impressive. In production, that percentage says little unless we know what fails, what failure costs and who reviews the result.

Not every organization needs to reach level 4. A professional using AI consistently and responsibly can gain more than a company whose entire strategy is a presentation.

3. So what about AGI?

The two ladders above are climbed with technology that already exists. This third one is what makes the headlines, and it runs from what exists today to what remains hypothetical. It helps you understand what the argument is about, not what to do on Monday. There is also no agreed boundary between its terms, so take what follows as a rough map rather than a settled scale.

Narrow AI (ANI): what we use today

Narrow AI is built to perform particular kinds of tasks: translation, recommendations, fraud detection, speech recognition, coding assistance.

Large language models break this classification, and it is more honest to say so than to force them into a box. The three-tier scale was drawn up long before one system could write code, summarise a contract and translate a letter in the same session. By breadth of task, an LLM is not narrow at all. What it lacks sits elsewhere: it does not learn from yesterday’s work unless you feed that work back in, it loses track of a goal over long chains of steps, and its accuracy drops in ways that are hard to predict once a situation drifts away from what it saw during training. Broad in what it can attempt, narrow in what it can be trusted to sustain. Most arguments about whether these models “really” understand anything are arguments about which of those two halves each side is looking at.

The mechanism underneath is worth two sentences, because it explains both halves. The model is trained on an enormous amount of text to predict what comes next, and then tuned with human feedback so that it follows instructions instead of merely continuing your paragraph. The first part gives it its range; the second is why it feels like a conversation.

None of this makes these systems small. Most of the AI producing real value in production today sits at this level.

AGI: a disputed horizon

Artificial general intelligence would approach human flexibility across a broad range of intellectual tasks: picking up unfamiliar skills, carrying knowledge from one domain into another, holding a goal over time.

The debate is hard to follow because the people having it are not measuring the same thing. There is no shared definition, so two forecasts about when AGI arrives can point at different destinations. The bar also moves: chess, then translation, then competitive programming each counted as proof of intelligence until software did them well, at which point they became merely tasks. And there is no agreement about what is missing. Some expect more data, more compute and better versions of today’s architectures to be enough. Others think it takes real progress in memory, causal reasoning or contact with the physical world.

“When will AGI arrive?” is a fascinating question that almost never helps with today’s decisions. What can current AI already do well in my work? does.

ASI: speculation with serious consequences

Artificial superintelligence would outperform people across virtually every domain, possibly including the design of better AI. It stays hypothetical, and the safety and alignment questions it raises deserve serious work: how do you make sure a far more capable system pursues the goal you actually gave it? What those questions should not do is take over every conversation about a tool that today drafts reports and classifies documents.

The next useful step

The three ladders move independently, which is the practical reason to keep them apart. A company can build narrow AI deep into its operations and get a great deal out of it; a person with the best model available can still be using it as a search engine. My own position, for what it is worth: today’s systems sit at the first level of the technical ladder, most of us sit between levels 1 and 2 of the second, and the move that changes the most is from occasional user to co-pilot. The third ladder does not depend on you. The first two do.

If you recognise yourself in the occasional user, start by not asking blind. Two ways of asking for the same thing:

Summarise this report.

Here is my team's monthly incident report, six backend engineers.
The reader is the head of operations: not technical, and they'll give it two minutes.
I want half a page covering what broke, who it affected and what we'll do
differently next month. No internal service names. Give me two versions, one blunt
and one more diplomatic, and tell me how they differ.

Length is not what makes the second one work. It carries the reader, the format, the length, what to leave out and a request for alternatives, which is the part almost everyone skips. Asking for two versions costs exactly what asking for one costs, and it forces you to choose instead of settling for the first answer because it reads well.

The jump from ad hoc use to repeatable use has a very concrete shape: a file. I keep the instructions that work in the repository, next to the code, not in the chat history, which is where they go to die. Each one has three parts: the instruction, an example output I accepted, and why I accepted it. That third part is the one hardly anyone writes down, and the only one still useful six months later, when you switch models and need to check whether the new one behaves the same.

If your use is already integrated, stop evaluating by feel. Twenty real cases pulled from production, each with the output you would consider correct next to it, and you can answer the only question that matters when you touch a prompt or move to a new model version: did it improve, or did it just change? That set of cases is what the industry calls evals, and it is far less sophisticated than the name suggests. Decide in advance which failures carry weight, too. In a document classifier, mistaking a receipt for an invoice costs one click; extracting the wrong amount costs a bad accounting entry and a call from the client. Writing that difference down before you ship is a good part of what separates level 2 from level 3.

Losing your fear of AI does not mean assuming it is harmless or capable of everything. It means knowing which ladder you are on, which step you occupy and whether climbing further is worth it.