What Not to Delegate to AI When You Run It All

Hand AI everything it can do well, and you still keep four things: taste, strategy, key relationships, and the final call on anything irreversible. Those are not bottlenecks you failed to remove. They are the reason the portfolio holds together. AI does the volume. You do the judgment that decides where the volume points.

I run around twenty companies alone. AI writes drafts, triages inboxes, ships code, reconciles books. If I delegated the four things below, I would not be faster. I would be running twenty companies badly instead of well.

What should you never delegate to AI?

Taste comes first. AI can generate a hundred landing pages. It cannot tell you which one sounds like a person who means it. Taste is the thing buyers feel and cannot name, and it is the only part of the output that a competitor with the same model cannot copy. I edit the last ten percent of anything customer-facing myself, because that ten percent is the brand.

Strategy is second. Which venture gets my attention this quarter, which one I starve, which market I walk away from: those are bets, not tasks. A model optimizes inside the frame you give it. It does not choose the frame. I cover how I make that call in how I decide which venture gets attention today, and it is not a decision I would ever automate.

Third, the relationships that actually move money. A model can draft the follow-up. It should not be the one a partner thinks they are talking to. When a deal matters, the other side needs to know a human with authority is on the line. Faking that is a short-term win and a long-term hole.

Fourth, anything irreversible. Sending money, signing contracts, deleting data, publishing a public claim. I let agents propose these and I approve them. The approval is cheap. The mistake is not.

Why keeping this work is leverage, not a bottleneck

People assume the goal is to delegate one hundred percent and go to the beach. That is the wrong target. The goal is to delegate everything except the decisions only you can make, so that all your attention lands there.

Think about where your edge comes from. It is not that you can write faster than a model. You cannot. It is that you know which twenty things are worth writing at all, out of a thousand possible things. That knowledge lives in the four categories above. Delegate them and you delegate your edge.

This is the same reason I do not hire a team to make these calls either. The judgment does not transfer cleanly to a person any more than it transfers to a model. I wrote about that in why I do not hire a team, even though I could. The point is not that people are bad. It is that the coordination cost of explaining your taste to someone else, over and over, can exceed the cost of just doing the taste-level work yourself.

How to draw the line in practice

Use a simple test on any task: if this goes wrong, is it cheap to fix and invisible to customers? If yes, delegate it fully. A bad draft is cheap. A miscategorized transaction is cheap and reversible. Let the agent run.

If a mistake is expensive, public, or permanent, the agent proposes and you dispose. This is the same human-in-the-loop pattern I use across every venture, and it is what makes broad delegation safe rather than reckless. Tools like Girard AI exist to run the volume while keeping approval gates on the parts that bite.

Then watch for scope creep in the wrong direction. The failure mode is not that AI takes over your strategy on its own. It is that you get busy, stop reviewing the ten percent that matters, and quietly let the machine set your taste by default. Every generic-sounding company you have ever seen got there this way.

The discipline is boring. Keep a short list of what stays with you. Guard it. Automate aggressively around it. That is how one operator covers twenty companies without any of them turning to mush. The leverage is in where a solo operator actually finds leverage: not in delegating judgment, but in delegating everything else so your judgment has room to work.