VertikalsBook a prototype callBook
All Field Notes

Hands, Not the Brain

AI is the best pair of hands you'll ever work with. It makes a dangerous brain to hand the keys to.

Download PDF

Everyone's racing to put AI in charge. The pitch is always some version of "point it at the problem and let it run." It's the most expensive mistake I watch companies make right now — and it's expensive precisely because it almost works.

Fluency is not judgment

AI is so articulate that it reads like thinking. It isn't. It is the most capable pair of hands ever built: it drafts, generates, summarizes, translates, and refactors faster than any team you could hire, and it never gets tired or annoyed at the tenth revision.

But hands don't know where to go. They don't carry the context of why this customer, this market, this bet. They don't feel the cost of being wrong. Point a fluent model at a badly framed question and it will hand you a beautifully written answer to the wrong question, with total confidence. That confidence is the trap. We're wired to read fluency as competence, and the model has infinite fluency and zero skin in the game.

Draw the line before you delegate

Here's the discipline that keeps teams out of trouble. Before you hand anything to AI, split the work into two lists.

  • Brain — defining the problem, deciding what "good" looks like, choosing between options, judging when to stop, owning the result.
  • Hands — producing drafts, generating options, doing the volume, handling the mechanical transformation.

Delegate the second list freely and without guilt. Keep the first. Then look at your two lists: if the decision, the definition of done, and the accountability have all landed on the AI's side, stop. You haven't delegated the work — you've abdicated the job.

The skill you give up is the one you'll need most

There's a failure mode pilots have a name for: automation complacency. Hand enough of the flying to the autopilot and the human slowly loses the instinct to tell when something's off — right up until the moment the autopilot can't cope and the human has to take over cold.

The same thing happens when you outsource judgment to a model. You erode the very ability to catch it when it's wrong. And it will be wrong — usually at the exact moment the stakes are highest and the answer is most plausible-sounding. The judgment you stopped practicing is the judgment you'll be reaching for when it matters.

How to actually run it

A few rules that have held up for us:

  1. You frame the problem and set the bar. The model doesn't get to decide what "done" means.
  2. The model generates options; you make the call. More options is a gift. More decisions made for you is a liability.
  3. Own the output like you wrote it — because your name is on it, not the model's.
  4. The harder and more irreversible the decision, the more human it should be. Reversible and low-stakes? Let the hands run full speed. One-way doors get a person.

Keep your hands on the wheel and let AI do the rowing. You'll move faster than the teams who gave the wheel away — and you'll still be pointed where you meant to go.

Point AI at the work. Never at the decision.