You can outsource your thinking, not your understanding

The insights report is flawless: sound methodology, clear findings, clean charts. An AI agent produced it in a fraction of the time it used to take an analyst. Then the client asks, “Why?” And then: “Where did that come from?” Two whys in, and the presenter has nothing left to hold on to.

The uncomfortable part is that the report wasn’t wrong. It was good, and the person defending it didn’t know why. AI can perform that analysis and do it well; what it doesn’t do is leave the context in the presenter’s head. Before, when an analyst read and synthesized the material, they could drill down to the example when asked: the exact mention, the nuance, the case that didn’t fit. Nobody studied that context deliberately; they absorbed it by doing the analysis. It was a free byproduct of the work. When we automate synthesis, we save analysis time and, without anyone budgeting for it, lose that byproduct.

A line Andrej Karpathy quoted at Sequoia AI Ascent 2026 captures it perfectly: “You can outsource your thinking, but you can’t outsource your understanding” ¹.

A series of seven experiments involving more than 10,000 participants found that, compared with web search, people who learned from LLM summaries reported shallower learning and later wrote shorter advice with fewer factual references ². In software development, this already has a name: comprehension debt, the growing gap between the code that exists in a system and what someone genuinely understands about it ³. Reports work the same way. And like any debt, it gets paid later, with interest, at the worst possible moment: a client’s objection, a crisis, a negotiation.

If a junior employee presents a report they don’t understand, we say they’re unprepared. If an executive does it with an agent, we call it efficiency.

The answer isn’t to go back to doing everything by hand. It’s to demand the same level of knowledge as if you had done it yourself. Anyone presenting a report needs to know the context cold: what the data says, where it came from, which cases support each conclusion and which ones contradict it. How they come to know it is up to them. What matters is the standard, not the method.

That comes at a price, and it needs to be said plainly: not all the time AI saves on production can go toward producing more. Some of it has to go back into studying and understanding. Because the client isn’t paying for a document. They’re paying for someone who can answer the second why.