Ask a model a question with five moving parts and it will answer immediately. Fluently, confidently, and quite often wrongly, in a specific way that is hard to catch: the answer is the most plausible-sounding answer rather than the one the numbers support.
That is not laziness and there is no effort dial you forgot to turn up. A model generates one token at a time, and each token gets roughly a fixed slice of computation. Demand a verdict in the first token of the response and there is nowhere for a five-step calculation to happen. So it pattern-matches: it produces the kind of answer that questions like yours usually have.
Chain-of-thought works because the reasoning becomes part of the generated output, and every later token is conditioned on the steps already written.
This is one of the most replicated findings in AI research. Wei et al. at Google published it in 2022, and asking a model to work through a problem before answering improved performance on complex reasoning tasks by roughly 20 to 50%, depending on the task. It has held up across models, across labs and across years, which is rare.
Two mechanisms are doing the work, and they are worth separating.
Commitment. Writing "the base rate is 18%, so 85% of customers scoring above 0.75 is impossible" puts that sentence in the context window. The next token cannot quietly ignore it. The model has committed to an intermediate step, and every subsequent step has to live with it.
Conditioning. Each written step narrows what the following tokens can plausibly be. By the time the conclusion arrives, the relevant facts are sitting right there in the context, written by the model, for the model.
The second benefit, which people undersell
Errors become inspectable. Without a chain you get a wrong answer and no idea where it came from. With one you can read down the steps and find the exact line where it went sideways, which usually turns out to be an assumption you never stated. That is the difference between "this is wrong" and "this is wrong because it assumed the deposit was not returnable", and only one of those is something you can fix.
Modern reasoning models (Claude's extended thinking, OpenAI's o-series) are essentially this technique built into the architecture rather than requested in the prompt. The idea did not go away when the models improved; it moved inwards.