Thinking Is Method

Let me start with the conclusion: there is no single right way to use AI in research. The idea that conversing with AI is “novice” and building agents is “advanced” does not describe how research works. It is a sales narrative. In disciplines where judgment must be produced — not merely executed — conversation is not the bottom rung. It is where the value is made.

The pedagogy of falling behind

Researchers who use AI to think out loud — asking, correcting, conversing — are increasingly told they are basic users. The “right” way, we hear, is to automate and become agentic. Anything else is standing at the technology’s front door.

We saw this narrative with blockchain and the metaverse. Every technology cycle needs us to feel a deficit, because courses, tools and consulting are sold on top of that deficit. Today’s version is a ladder of “user levels”: the conversational user at the bottom, the executor and the builder above, the occasional user barely counted at all.

The problem is not the tools. It is the ladder. It does not describe how researchers work — it prescribes it. And the prescription is commercial, not technical.

Functions, not rungs

The four ways of relating to AI are functions, not stages. One researcher moves through all of them, often within a single project. Some build first and return to conversation when their judgment gets disordered. Others converse for weeks and build in an afternoon. The transit is multiple and reversible.

The question that matters is not how the tool is used. It is this: is the judgment already given, or does it have to be produced?

When judgment comes ready-made, starting agentic makes sense. Four cases: executing a standardized method someone else designed; working in fields where an external truth corrects quickly (code compiles or it does not); exploring unknown territory by generating variants; and learning a new domain by building and disassembling. In all four, the judgment already exists somewhere — the world provides it. None of this is falling behind or getting ahead. It is fitting the approach to the task.

Where judgment is the work

Research that works with meaning — semiotics, user experience, audience reception — cannot compile itself. There is no external signal that flags the error. The researcher establishes the criteria, or no one does.

In this craft, conversation is where judgment is forged. And the agentic mode, far from being a superior stage, is often the destination: execution and building arrive when the method is so clear and so rehearsed that it can finally run on its own.

There is something the progress narrative never names, and for researchers it is worth gold: the most powerful move is not automating tasks but encapsulating your own method. Anyone can use a generic agent. Turning your judgment into an instrument — one that analyzes the way you do, not the way any model does — is not being an advanced user of someone else’s tool. It is owning an instrument of your own.

A case from practice

In my own work I built a narrative ecosystem around sustainability: a cartography of published discourse, analyzed through the social semiosis framework of Eliseo Verón, the Argentine semiotician who studied how meaning transforms as it circulates.

It began as conversation and manual analysis — slow, by hand, pure judgment. From repeating that analytic gesture, a methodology precipitated: discursive territories, reading rules, a stable way of mapping meaning. Only then, with judgment mature, did it become agentic design: an instrument that reproduces the method and scales it across a corpus no single person could read.

The phase the ladder would call “novice” was not the opposite of the agentic function. It was its raw material. Without that slow, reflective conversation, there would have been nothing worth automating.

The takeaway for our industry

For research leaders and insight teams, three recommendations follow.

First, audit the task before choosing the mode: if judgment is given, automate early; if judgment must be produced, protect the thinking phase — it is the deliverable.

Second, measure AI maturity by ownership of method, not by volume of automation. A team that has encapsulated its own analytic criteria into instruments is more advanced than one running generic agents at scale.

Third, resist the deficit narrative when buying training or tools. Ask vendors one question: does this help us execute someone else’s judgment faster, or build our own?

To get from one point to another there are many routes and many vehicles. No one claims the airplane is more evolved than walking: it depends on the distance, the load, the terrain. Researchers who begin by defining judgment did not choose the slow road. They chose the only transport that crosses unmarked terrain. That is never falling behind. It is method.