AI as a Thinking Partner, Not a Decision-Maker

"Should I hire someone or outsource this?"

Ask a model that and you will get an answer. It will be well structured, balanced, and completely untethered from your actual situation — your cash position, your tolerance for management, whether you enjoy the work, what happened last time you hired.

The answer is confident because the model is always confident. It is not confident because it knows.

The reframe

A model has no stake in the outcome, no accountability, and no knowledge of your context beyond what you typed. Those three absences make it a bad decision-maker and a genuinely good thinking partner — the same qualities that make it useless for choosing make it tireless for examining.

It will argue any side without ego. It will not get bored on the fourth pass. It has no relationship to protect.

Five things worth asking it

1. Steelman the option I am not choosing. You already know why your favourite is good. Make it build the best possible case for the alternative — properly, not as a straw man.

2. What am I assuming that I have not stated? This surfaces the load-bearing beliefs you have stopped noticing. Then: how could I test each one cheaply?

3. Run a pre-mortem. "It is a year from now and this failed. Write the account of why." Anticipated failure is far easier to see than hypothetical risk, and this framing reliably produces the specific, boring reasons rather than dramatic ones.

4. What would change my mind? If you cannot answer this, you have not made a decision — you have made a commitment. Naming the evidence in advance keeps you honest later.

5. Who does this affect that I have not considered? Second-order consequences are exactly what a tired person misses at 6pm.

What to give it

The quality of this depends entirely on what you feed it. Vague inputs produce the generic business-school answer.

Give it the real numbers, the real constraint, and the thing you are secretly hoping is true. That last one matters — stating your hope explicitly lets you ask it to attack that hope specifically, which is the most useful move available.

The failure mode to watch

Confirmation. It is very easy to keep rephrasing until the model agrees with you, and then feel validated. Models are agreeable by default; they will find merit in most things.

Two habits help. Ask for the counter-case before you say what you prefer. And when it does agree, ask what would have to be true for it to be wrong.

Where the line sits

Widening the options, testing the reasoning, finding the blind spot, doing the sums — all fair game.

Weighing what matters to you, carrying the consequence, living with the choice — yours. That is not a limitation of the technology. It is what a decision is.