A large language model is trained, above all, to be helpful. And helpfulness, in conversation, is almost indistinguishable from agreement.
Ask whether you should make the hire and it helps you make the hire. Ask whether to pivot and it helps you pivot. It answers the question you asked, in the direction you were already leaning, with more fluency than you had alone. This is a feature for most uses and a defect for one: the high-stakes decision, where the thing you most need is the question you didn’t ask.
Why agreement is the wrong default
For a reversible, low-stakes choice, an agreement engine is fine — speed matters more than scrutiny. For the decision that spends your runway or defines your next two years, the opposite is true. You don’t need a faster version of your own reasoning. You need an adversarial one.
A second opinion that always agrees with the first is not a second opinion.
This is the entire reason a decision audit is structured as an audit and not a chat. An audit applies the same discipline to every decision and is built to find what you missed — not to help you do faster what you already intended.