How do you use AI agents without losing your own judgment?

How to use AI agents without losing your judgment as a founder

Hand the agent the legwork, not the verdict. As Ethan Mollick frames it in Co-Intelligence, the centaur split works: AI does the search, drafting, and pattern-finding; you keep the framing, the trade-offs, and the final call. The week you stop being able to explain a decision in your own words is the week you have outsourced too much.

The mistake I see founders making right now is not under-using AI agents. It is the opposite. They have built such smooth pipelines — Claude reading their docs, an agent drafting their emails, another summarizing every call — that within a quarter they cannot tell you, in their own words, why their company is doing what it is doing. The output looks great. The thinking underneath has quietly thinned out. I have done this to myself more than once, and the recovery is harder than the prevention.

The most useful frame I have found for this is Ethan Mollick's centaur-versus-cyborg distinction in Co-Intelligence. Centaurs split work cleanly: the human handles the parts that require human judgment, the AI handles the rest, and the seam between them is conscious. Cyborgs blend the work — every paragraph, every decision, is co-generated, and over time you cannot tell where your thinking stopped and the model's started. Centaur usage compounds your skill. Cyborg usage, especially on the things you care about most, slowly erodes it. The good news is that the split is a choice you make per task, not a personality trait.

Here is the split I have landed on after a year of running my own work this way. I hand the agent the legwork: web research across forty sources, draft outlines, first passes on routine emails, transcript summaries, harvesting questions from Reddit, cleaning up messy notes, code review of small changes, due-diligence checklists. I do not hand the agent the verdict: the framing of a strategic problem, the trade-off that defines a positioning choice, the hard email to someone I care about, the hiring decision, the "what does this mean for us" reading of a market signal. The legwork is where AI's leverage is enormous and the downside of being wrong is small. The verdict is where my judgment is the actual product, and where being subtly wrong compounds in ways that are hard to undo.

The mechanism this protects is what Kahneman called System 2 — the slow, effortful, deliberate kind of thinking. System 2 is metabolically expensive. Your brain will offload to System 1 whenever it can, and AI is the most attractive offloading target ever invented. Mollick is explicit about this risk in Co-Intelligence: in his MIT and Wharton studies, people who used AI on tasks slightly outside their skill ceiling got better results but worse learning. They got the answer; they did not get sharper. Repeat that pattern on your strategic decisions for six months and you have not become a better founder — you have become more dependent on a model that will not be in the room when you actually have to lead.

The check I run on myself once a week is borrowed from coaching. It is closer to Sir John Whitmore's GROW model than anything technical: pick the most important decision you made this week, set a timer for ten minutes, close the laptop, and write — by hand — the reasoning behind it. Not the result. The reasoning. If you can articulate the trade-off, the alternatives you considered, and what you would do differently with the same information again, you are still doing the thinking. If you find yourself stuck after three minutes, paraphrasing something the agent told you without really understanding why, that is the early warning sign. That is the week to dial back.

There is also a small workflow rule I have stolen from Cal Newport's Slow Productivity: protect at least one block of deep work per day where no AI assistance is allowed. Not because AI is bad — it obviously is not — but because the muscle of sitting with a hard problem, getting bored, getting stuck, and pushing through to your own answer is the muscle that atrophies fastest when the model is one keystroke away. Forty-five minutes a day is enough. The block does not have to produce anything. It has to produce you doing the thinking.

The honest summary, after watching this play out across dozens of founders I coach: AI agents will make you faster at almost everything. Whether they make you better depends entirely on what you hand them. Give them the legwork, keep the verdicts, and run the weekly check. If you can still explain your decisions in your own words at the end of every week, the leverage is working for you. The day you cannot, you are not delegating anymore. You are being replaced — by yourself, in advance.