Your question to the AI is the bottleneck, not the model

When it comes to choosing an AI model, I hear deliberate decisions in very few conversations. Instead, it is either habit or “use the best one” (out of fear that otherwise it won’t be “good enough”). One costs money and time unnecessarily, the other—just as unnecessarily—costs quality.

Yet handling model and compute effort is simpler than it looks:

• Model size scales judgment: nuance, weighing things up, and above all ambiguity.
• Thinking effort scales thoroughness: multiple steps for things that can be checked.

My experience looks like this:

▪ Low thinking effort: small model. Structuring, reformatting, summarizing, translating. Or simply for everything I can check myself within seconds. Three quick attempts beat one expensive one here.

▪ Day-to-day business: standard model. Drafting texts, putting research results into context, developing ideas, routine code. Standard tasks that need more depth than legwork.

▪ Choose high thinking effort: when the problem is multi-step and checkable—costing, debugging, planning with many dependencies. Very important: the model can remain a small or older one as long as the task is clear (= solvable through diligence).

▪ Large model: when judgment is called for. Strategy, decision-making, positioning, contradictory sources, delicate communication. A job posting would say: “tolerance for ambiguity.” This is where the difference becomes glaring.

▪ Both high (big model, high effort): only when three things come together: 1. no clear path to a solution, 2. real costs of error or irreversible decisions, 3. enough context for depth to emerge at all. But that is MUCH rarer than I thought.

▪ And when a small model with low effort? With a thin prompt, with no cost of error, at high volume. A top model with a thin prompt delivers an overpriced result. I don’t send the SWAT team to do a traffic stop. About three quarters of my requests to AI probably fall into this range.

That brings me to the core: the bottleneck is almost never the model! The bottleneck is your question. If you are not clear about what you want, the largest model will only give you a better wording of the same unclear request.

Many people know SISO: Shit in → shit out. The formula I find more strategically relevant is:
Perception in → Decision out.

The quality of a decision is limited by the quality of the perception, the framing, and the information that precede it. (This formula, by the way, comes from my sparring with the AI—I couldn’t have put it more succinctly myself just then.)

In guiding AI transformation, effectiveness and efficiency are left on the table because money is invested in tools without it being clear, in day-to-day work, what exactly people want to ask or solve. Clarifying the request (for the clerk as much as for the CEO: “What is this really about right now?”) is not a preliminary stage of using the tool. It is half the work.