In 2004, Jeff Bezos banned PowerPoint from Amazon’s senior leadership meetings in favor of narrative memos, read in silence at the start of each meeting. The idea was that a well-structured and tightly argued memo communicated its ideas in a way that a deck built around bullet points could not. As Amazon moves into the AI era, it is developing internal tools that can help produce polished documents much more efficiently. Its internal AI assistant, Cedric, promises to deliver “six-page narratives in seconds,” according to Kristi Coulter, a former Amazon executive.
But no matter how good Cedric’s outputs are, using an AI tool to trim the traditional writing process down to mere seconds undercuts what made well-written memos valuable in the first place. The memo itself was never the point. What mattered was what writing the memo forced the human to do. As Bezos put it, there is “no way to write a six-page, narratively structured memo and not have clear thinking.” The process, not the output, was the real source of value.
How AI breaks the memo
Bezos wasn’t promoting writing for its own sake. Writing mattered to him because it was a forcing function for rigorous thinking about important decisions. As one Amazon veteran told Coulter, writing the memo forced them to confront their own ideas and clarify what they wanted to say. And this often meant that writers would change their mind several times while writing the narrative.
It is true that building meetings around memos disciplined the readers as well. Because meetings began in silence, everyone worked through the same argument before anyone spoke, and the discussion started from the substance. But if the content of a memo carries less of the writer’s thought, the value of reading it diminishes in proportion to that fall. Time spent studiously reading well-written but empty texts is time wasted.
The underlying problem is that AI can break the link between thinking and the production of text. Steve Huynh, a former Amazon principal engineer, recently described a review meeting in which a product manager began discussing a six-page memo he was supposed to have written. Almost immediately, it became clear that he was reading it for the first time. He was presenting something that he had not thought about at all.
Readers are adapting to the change in the writing process. Coulter reports that one Amazon manager now asks for longer documents because he plans to have AI summarize them, and Huynh says some employees now tag suspect memos “AI;DR”: AI, didn’t read. In a Stanford and BetterUp Labs survey on workslop, 42% of workers who had received low-effort AI-generated work said they saw the sender as less trustworthy. Once a written product stops being evidence that someone thought about the content, readers simply stop giving it their attention.
The memo’s 3 jobs
The trap for leaders is to treat this as a problem about memos. Ultimately, businesses don’t need memos. They need good decisions. The memo mattered only because, for a long time, it reliably communicated the results of the thinking process that lay behind the writing.
Baseball provides a useful analogy here. In his book Moneyball, Michael Lewis tells the story of how the Oakland A’s responded to the loss of a star player, Jason Giambi, to free agency. Instead of trying to replace him with another player of the same caliber, the A’s realized they didn’t need another Giambi, per se. What they needed was the measurable impact he had on their results. So rather than search for one player to fill the gap, they tried to recreate him “in the aggregate” by finding several cheaper players who together delivered the same statistical impact.
There is a lesson here for how we think about Amazon’s memos—and, indeed, for how we think of much else in the new AI-powered workplace. We shouldn’t be looking for simple substitutions of AI work for human work. Instead, we should look at how work produces value in aggregate and then redesign workflows to reproduce the end result, not the individual steps.
A well-written memo bundles three distinct jobs into one artifact. It forces authors to think, it provides evidence that the author has done the thinking, and it gives its readers a shared understanding of the issue at hand. What the Oakland A’s can teach us is that, rather than trying to restore all of those functions in one stroke, the better path is to make two distinct moves: one that restores the first two jobs together, and one for the third.
Force the thinking, then prove it
The rebuilding process starts with honesty about where thinking is actually needed. A great deal of human writing does not provide the basis for making important decisions, so in those cases there is little reason to build a replacement process that forces the thinking and evidential functions alongside the communicative role of the writing. A report summing up project progress over the past week, for instance, can simply be an assemblage of information, and the process of collating and communicating it can happily be automated. But when it comes to visions of the future or arguments about which decisions are best, leaders have to adapt to the fact that polished prose no longer guarantees deep thinking.
The first move is to both rebuild the forcing function and provide evidence of the underlying thinking. This can be done by making the author, rather than the document, the thing that gets examined. Amazon’s meetings already included the seed of this requirement. Once the silent reading ended, the writer had to answer for every line. The change for the AI era is that the defense becomes the initial gate. A memo earns its place on the agenda only if its author can walk the room through the argument and take questions without the document in front of them. That is not something the author can hand to Cedric, because Cedric will not be in the room, and knowing the defense is coming does what the blank page used to do.
Terence Tao, the Fields medalist, has proposed the same rule for mathematicians: If the authors of an AI-assisted result cannot give a clear, expert-level talk on it, the result should not be published. A proof that no human can properly explain, he writes, “should be viewed as incomplete, even if it has been formally verified.” The same holds for a memo.
Who does the understanding?
The memo’s third job, helping create shared understanding, seems the least problematic, but it raises a counterintuitive challenge. We tend to think that AI-written texts are problematic because they waste our time by spreading poorly thought-out ideas. Restoring the underlying thinking should, then, automatically solve the communication problem as well. But things are not quite so clear cut. Even when the content is genuinely good, AI can help humans produce more valuable outputs than other humans can absorb.
We can see this in the challenges that the discipline of mathematics is facing. Mathematicians are being overrun by journal articles and preprints containing AI-generated proofs that nobody can properly digest. In September, OpenAI said its models had solved one of math’s Millennium Prize problems. A month later, it released hundreds of AI-generated results in one batch. Many of the proofs have been verified by software that checks their logic, but knowing that a result is correct is not the same as understanding why it is. 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, a mathematician at MIT, argues that AI’s results truly become part of mathematics only when a human reads them, understands them, and gets something out of them. A thousand new proofs sitting unread on a shelf add nothing to the discipline. The same logic holds for companies grappling with a surfeit of good writing outputs; no matter how intelligent a memo is, it becomes a part of a company’s decision-making only if the relevant humans absorb it. That points to a new division of labor: AI can increasingly shoulder the generative burden, but the understanding still has to be done by people.
That makes human understanding a scarce and valuable resource, and leaders may have to start rationing it. Working out how to filter and direct the flow of outputs becomes an entirely new type of task. The preprint server arXiv, for scientific papers, now limits authors to two submissions a month. A company can do something similar, such as limiting the number of memos that get a silent-reading slot in a given month.
The artifact is not the thinking
For two decades, a six-page memo was reasonable evidence that someone had thought hard. AI has broken that link, and no policy will fully restore it. What leaders can do is stop mistaking the document for what it was meant to produce. The alternative is more six-pagers than ever, and fewer people who understand what’s in them.
