Why ChatGPT is not a strategy process

Strategy is a sequence of dependent decisions

There is no single task called “do strategy”. A brief is interrogated. The problem is diagnosed. The audience is understood. Evidence is interpreted. A positioning is chosen. Campaign and creative decisions follow. Each step changes the next one. If the audience work changes the problem, later work needs to know that.

If the positioning closes one route, campaign strategy should not casually reopen it in a new conversation. The difficult bit is not producing good text at each stage. It is keeping the work connected.

This is why allocating strategic tasks to AI also requires a decision about the sequence.

The user becomes the integration layer

An experienced strategist can make ChatGPT behave like a process. They know which task should happen next. They know what context to paste in. They remember which previous decision is authoritative. They notice when two outputs contradict each other. They carry the work between conversations and clean up the joins.

At that point the software is providing intelligence on demand. The process still lives in the strategist. That can work extremely well for one senior person. It is harder to turn into a repeatable agency capability.

Chat does not create commitment

Strategy needs moments where the team decides something. This audience definition is accepted for the assignment. This positioning route is chosen. This contradiction remains unresolved and blocks the next stage. A chat history can contain all of those statements without making their status clear. Then the conversation continues.

Three days later somebody can generate a new answer that quietly assumes something different. The text exists. The process state does not.

Why better strategists get more from ChatGPT

Experience gives a strategist a standard to push against. I once worked on a brief with Johanna and said early on that the strategy would probably need to focus on one particular thing. We went through the evidence and arrived at the same place. When she asked how I had known, the honest answer was that I had seen this movie before. That is pattern recognition, and it still needs checking.

A model gives that judgement more material to work with. It can also produce an agreeable, well-structured argument for a weak conclusion. Ask what evidence would overturn the recommendation, and check whether the answer changes because the evidence changed or because you expressed a preference.

The evidence does not justify saying AI always widens the gap between strong and weak practitioners. In the HBS and BCG experiment, gains depended on the task; lower-performing consultants gained more on tasks inside the model’s capability frontier. My concern here is uncritical approval of persuasive output, not a claim that less experienced people cannot benefit.

“Human in the loop” is too vague

People often solve the AI quality problem by saying a human will check the output. Strategists do not usually check strategy. They judge it. Does the answer ring true? Does the evidence support it? Is the choice useful? Will it help the creative team? Should it survive into the next stage? The distinction matters because standard LLMs are extremely good at presenting plausible work with confidence.

General LLMs are still part of the answer

General LLMs remain part of my work. They offer flexible exploration, while a shared process provides sequence, accepted decisions, review criteria and stopping points.

Before the next handover, name the decision that has been accepted, the evidence behind it and the person allowed to reopen it. Make that record available to the next task in the strategic workflow, rather than relying on someone to remember which conversation contained it.