2026-09-21 · AI
AI Doesn't Make Your Judgement — It Amplifies the One You Already Have
Over the past stretch I built three analysis skills with a coding agent: trading comps, a one-page DCF, and a weekly macro panel. Before starting I assumed the payoff would be time saved. It wasn't. The real change was that my workflow got forced into written form, where nothing vague can hide.
It can't do a division reliably — and that's not really the problem
At first I had the agent produce the comps table directly. It fetched plausible data, and the median was always slightly off. The reason isn't mysterious: it is generating a plausible answer, not computing one. Ask it for the median of four multiples and you get a number that looks right and isn't.
So I split the work: the agent finds the data, aligns definitions and writes the conclusion; a Python script does every calculation, and the agent reads the result afterwards. The script carries the red lines — WACC at or below terminal growth throws an error, missing data is marked N/A, no filling gaps with an invented number. That division of labour sounds like common sense, but what it actually changed was the structure of trust: I can now say exactly how each column was produced.
When your process is fuzzy, it scales the fuzz into something that looks professional
This is the strongest thing I took away from the three builds. If your process is really "pick a few peers by feel, roughly align the multiples," the agent will execute that fuzzy process with enormous diligence and hand you something beautifully formatted, correctly worded and quietly wrong. It won't correct your method. It will run your method faster.
The flip side is that when the process is clear, it's remarkably good. Screening peers used to be a matter of judgement for me; now it's three hard criteria — overlap of business model, size band, geography and regulatory environment. Once it was written down I noticed something uncomfortable: I used to quietly relax the second criterion, because genuinely comparable companies are scarce and one extra name makes the numbers look better. Once the shortcut is written down, you become too embarrassed to keep taking it.
Writing the SKILL.md is itself a retro
To hand a task to an agent you have to state what each step takes in, what it produces, and when it must stop. Everywhere you can't say it is somewhere you haven't thought it through.
Take "downside floor." In my head it had always been a comfortable blur: roughly what this company is worth when conditions are bad. Trying to specify it exposed that the blur wouldn't stand up — does bad mean growth gets revised down, or the discount rate goes up? What if both? How much? It ended up as three parameters: growth haircut, WACC uplift, margin compression, each requiring a stated reason. Only after that did the vague version become something I can run repeatedly.
The judgement is still mine
What the tool can take over is fetching, arithmetic, formatting and repetition. What it can't take over is the flicker of "that number isn't right" when I look at a figure. That instinct has an unglamorous source: I've lost real money on numbers of exactly this kind.
So my position on AI is neither embrace nor suspicion, it's a division of labour. Whatever can be written down goes out; whatever can't stays with me. And then you have to be honest that the second part is the part that's actually worth something.