AI can write the memo, but can it prove the thinking?

Forty-two percent of workers who received low-effort AI-generated work in a Stanford and BetterUp Labs survey said they saw the sender as less trustworthy. Former Amazon executive Kristi Coulter reports that Amazon’s internal AI assistant, Cedric, promises to deliver six-page narratives in seconds.

In 2004, Jeff Bezos banned PowerPoint from Amazon’s senior leadership meetings and replaced slide decks with six-page narrative memos, read in silence at the start of each meeting. The format required a tightly argued account before debate began, shifting discussion from bullet points to the substance of the argument.

Cedric’s speed threatens Bezos’s method. Bezos said: “There is no way to write a six-page, narratively structured memo and not have clear thinking.” The process, not the polished output, supplied evidence of judgment.

An Amazon veteran told Coulter that drafting forced the writer to confront the idea, clarify what they wanted to say, and sometimes change their mind several times. Silent reading disciplined readers too: everyone worked through the same argument before anyone spoke, so discussion began with substance.

A polished document without that work is merely text, and time spent studying it is wasted.

AI can sever the link between thinking and text production. In a review described by Steve Huynh, a former Amazon principal engineer, a product manager discussed the six-page memo he was supposed to write. The exchange showed that he was reading it for the first time and had not thought through the proposal at all.

Readers are adapting. Coulter reports that one Amazon manager now requests longer documents because AI will summarize them; Huynh says some employees tag suspect memos “AI;DR,” marking them as AI-generated and unread.

Once a written product stops serving as evidence that its sender thought about the content, readers stop giving it attention.

What three jobs does a memo perform?

The target is the decision, not the memo. A good memo bundles three functions: it forces the author to think, provides evidence of that thinking, and gives readers a shared understanding. AI can handle communication, but speed cannot restore the first two.

Michael Lewis’s Moneyball offers a useful model. After the Oakland A’s lost star Jason Giambi to free agency, they stopped searching for one equivalent replacement and assembled cheaper players whose combined statistical impact matched Giambi’s contribution. The lesson for AI is to preserve a workflow’s aggregate outcome rather than substitute AI for each human step.

Where should leaders require human thinking?\n

Leaders need two changes: restore the forcing function and its evidence together, then address shared understanding separately. Not every document needs deep thought: a weekly progress report can assemble and transmit facts, so automation may cause little harm. A future vision or contested recommendation needs more.

Polished prose no longer guarantees deep thought, so the process must restore that evidence. Leaders should examine the author, not just the document. At Amazon, the writer had to answer for every line once silent reading ended.

In an AI-assisted meeting, the defense can become the entry gate: a memo earns an agenda slot only if its author can walk the room through the argument and take questions without the document in front of them.

Cedric cannot do that. Announcing the defense recreates the blank page’s function by making the writer know the argument before polished prose can stand in for it.

Terence Tao, the Fields medalist, proposes the same rule for mathematics. If authors cannot deliver a clear, expert-level explanation of an AI-assisted result, they should not publish it. Even a formally verified proof remains incomplete when no human can explain it; the same test applies to a memo.

Who has to understand the output?\n

Restoring thought does not automatically solve comprehension. AI can generate genuinely valuable writing faster than people can absorb it, and mathematics is now overrun by journal articles and preprints carrying AI-generated proofs that no one can properly digest.

In September 2026, OpenAI said its models had solved one of the Millennium Prize Problems; in October 2026, it released hundreds of AI-generated results in one batch.

Software has verified many proofs by checking their logic, but correctness does not explain why a result holds. The Association for Human Mathematics put it bluntly in a statement reposted by Tao: “We reject OpenAI’s assertion that this release advances our subject.”

Pavel Etingof, an MIT mathematician, argues that an AI result enters mathematics only when a person reads it, understands it, and extracts something from it; a thousand unread proofs add nothing. Companies face the same condition: AI can shoulder generation, but people must still do the understanding.

Human understanding is scarce. Leaders will need to filter and direct the flow of outputs. arXiv, the preprint server for scientific papers, now limits authors to two submissions per month. A company could apply the same logic by limiting how many memos receive a silent-reading slot each month.

What should replace the document?

For two decades, a six-page memo was reasonable evidence that someone had thought hard. AI has broken that link, and no policy can fully restore it. Leaders should stop treating the artifact as proof of the work it was meant to produce, or they will get more six-pagers and fewer people who understand what sits inside them.

The better future is not faster memos; it is meetings in which each memo earns attention because its author can defend the thinking behind it.