I use an AI as a thinking partner for reviewing my own decisions and routines, but only once I stopped treating its confidence as evidence of anything.
I use an AI chat regularly the way some people use a journal with a reply function: I describe a decision I am about to make, or a routine I have fallen into, and ask it to push back. This is genuinely useful, and it is useful for a narrower reason than it first appears — not because the model knows anything about me, but because articulating a decision out loud, to something that will actually respond, surfaces the gaps in my own reasoning faster than thinking about it silently ever does. Writing a decision down does something similar, but writing to something that answers back forces a level of specificity that a private note rarely gets, because a note does not ask a follow-up question.
The failure mode I watch for constantly is treating the model’s confidence as if it were evidence. It will tell me, in the same even tone, something it has genuinely inferred from what I told it and something it has effectively guessed, and nothing about the delivery marks the difference. A human coach who does not know an answer usually signals that somehow — a pause, a hedge, an admission. A model under-signals this by default, which means the burden of noticing when it is speculating sits entirely with me.
What actually works is being precise about what I am asking it to do. "Is this a good decision" is a question with no ceiling — it will always produce an answer, because there is always something plausible to say. "Here is the reasoning I used, find the step that does not follow" is a bounded task, and a bounded task is one I can actually check the output of, because I know exactly what a correct answer would look like and can hold the response to it.
The place this earns its keep is not the big decisions, where I already consult people who know me, but the small recurring ones — the daily and weekly routines that never feel important enough to bring to another person, but that compound the most because they repeat. Having something available at the moment I am actually deciding, rather than saving it up for a conversation that may not happen for weeks, changes how often the review actually happens, and a review that happens is worth more than a better one that does not. None of this replaces the deeper conversations with people who actually know the context, but those conversations do not scale to a daily cadence, and pretending they should is how the small routines quietly go unreviewed for months at a time.
None of this makes it a coach in the sense of someone who knows my history, notices the pattern I cannot see because I am inside it, or tells me something I do not want to hear because they are genuinely invested in the outcome. It is closer to a very fast, very patient sounding board that has read a great deal and remembers nothing between sessions unless I tell it to.
That is a real and specific use, and pretending it is more than that is exactly the mistake that makes people trust an answer they should have checked. The value is in the speed and the availability, not in any claim to actually understand the person on the other end of the conversation, and keeping that distinction clear is most of what makes the habit worth keeping.
I have also noticed the habit changes depending on how I open the conversation. Starting with the decision already made and asking for a reaction gets a much softer response than starting with the reasoning laid bare and asking it to find the weak link — the second framing gives the model an actual task with a checkable answer, while the first invites agreement dressed up as analysis, which is the opposite of what I am there for.