Ask anyone in San Francisco what AI is doing to their company and youʼll get an answer about productivity: code written faster; decks drafted overnight; tickets handled without a human touching them. All true, and all boring by now.
Spend a few months raising money here and you notice something less discussed.
You’ll be surprised, but it isnʼt the work being automated, but the judgement around the work.
You see it in small ways first: a follow-up email that reads a little too neatly. The same objections in the same order from three people who have never met. A memo with that particular consistency to it — every paragraph equally confident, nothing at stake.
None of it is bad, and thatʼs the odd part. Itʼs all perfectly reasonable, which is what makes it hard to notice.
A lot of people have quietly moved the first draft of their thinking into a model, and the model is very good at producing something defensible.
However, defensible means a low bar when everyone has the same tool.
When your own analysis, and that of your competitorʼs, come from the same place, you havenʼt gained an edge. Youʼve both bought the same average opinion at the same price.
This cuts deepest in investment because investment is fundamentally a bet on something that hasnʼt happened yet.
A language model works by compressing what has already been written. Thatʼs not a flaw, but the design.
It makes these systems excellent at anything with precedent and unreliable at anything without one.
Ask it to summarise a market and it will beat most analysts. Ask it whether a category that doesnʼt exist yet will matter in five years and it hands you the consensus of people who couldnʼt have known, in the same steady voice.
Every generational company looked unreasonable at the start.
Recently, I sat in on a lecture at Stanford by William Barnett, who studies why organisations miss the things that end up mattering, and he put it better than I could: “We are so bad at predicting, but so good at retrospectively rationalizing.”
His example was Google, which struggled to raise money because nearly every search engine before it had failed.
The evidence was real. It was also pointing the wrong way.
His other observation has stayed with me. “Your people fear being a fool more than they hope to be a genius.”
That was true long before AI. What changes now is that the safe, defensible answer is instant and essentially free.
The non-consensus position still has to be said out loud by someone prepared to be wrong in front of the room.
The same pattern shows up inside large companies, where I spend most of my time now.
Executives have never had more analysis, and it hasnʼt made the hard calls easier because those calls were never blocked on analysis.
Instead, they were blocked on committing to one future over another and being accountable for it.
AI produces more material for the decision and does nothing to make it.
Thatʼs what we kept running into when building Principle — and almost every conversation with a leadership team lands in the same place: they donʼt lack information, they lack a view.
None of this is an argument against the tools. I use them constantly and my company is built on them.
Itʼs about which part of the job you hand over. Delegating the research is sensible.
Delegating the position is how a market full of well-informed people makes the same mistake at the same time.
Agency is a scarce thing now. Execution has never been cheaper, and deciding something and owning it has never been rarer.
So the question for any leader this year isnʼt whether AI can do part of your job. Itʼs which part youʼve already stopped doing yourself.