Clay Banned AI Docs That Cost More to Read Than to Write

Clay published an internal AI writing policy this week, and the notable thing is what’s missing from it. No section on legal exposure. Nothing about client data going into a model. Nothing about hallucinated facts in customer-facing copy.

All four principles are about one resource: the time your colleagues spend reading what you produced.

Varun Anand, Clay’s co-founder and COO, announced it on LinkedIn on August 10. The document was written by Sophie Alpert, an engineer there, and the adoption path is the part worth noticing:

“Originally this was just for eng, but other teams found it so helpful we expanded it company-wide.”

A policy that spreads because other departments asked to be covered by it is a different object from one handed down by legal. It means the problem it solves was already being felt everywhere.

The four rules

Alpert published the underlying argument on her own site on June 25, six weeks before the company-wide rollout, under a title that is the whole thesis: there are no lossless transformations of natural-language text. Her framing is that every rewrite changes meaning, and when the rewriting is done by something that doesn’t hold your intent, information is lost.

The four principles, as Anand posted them:

1. You must stand behind every idea and sentence. “It is your responsibility to make sure that the entire document is representative of your own thoughts before you share it.”
2. Writing is thinking. “Spending time on the writing process teaches you more about your topic. If you circumvent this process, you will walk away with a poorer understanding of the subject matter.”
3. More time should be spent writing a document than consuming it. “If you generate a document from a short prompt then ask your readers to go through the longer output, you are disrespecting their time. They can talk to ChatGPT themselves if they want to.”
4. Longer is not better. “AI makes it easy to generate long docs, and it loves padding them with sentences that say nothing. If you’re producing docs from a short prompt, consider just sharing the prompt.”

It isn’t a ban. Brainstorming with a model is fine, drafting is fine, proofreading is explicitly fine, and even pasting raw output into a document is allowed as long as it’s labeled as raw output. The line is drawn at passing off unattributed generation as your own thinking.

Rule three is the one that generalizes, because it contains a test anyone can run: does this document cost the room more time than it cost me? Alpert’s version of the math is that reader time carries “a multiplicative cost,” which is the obvious point everybody skips. One person saving 20 minutes by generating a doc can spend an hour of other people’s attention, and nobody’s calendar shows the trade.

It’s also the counterweight to a problem we’ve covered from the output side, where AI made content cheap to produce and differentiation expensive. Clay is applying the same logic inward: cheap to generate has never meant cheap to absorb.

Why a marketing team should care about an engineering memo

Marketing is where this asymmetry is worst, because marketing runs on documents that exist to get other people pointed the same way: the brief, the positioning doc, the campaign readout, the quarterly narrative. Those artifacts have no value except the shared understanding they create. A brief that took 4 minutes to generate and takes 6 people 20 minutes each to decode has destroyed value while looking like output.

There’s a labor angle sitting underneath it too. The Content Marketing Institute surveyed 644 marketers in February 2026 and found 76% doing the work of more than one job, with 50% absorbing new responsibilities without a raise or promotion. That’s the same squeeze showing up in which marketing roles companies still post for. A team in that state has every incentive to generate the deck and move on. Rule three is a constraint aimed squarely at the coping mechanism of an overloaded team, which is why it needs to be written down rather than assumed.

Enforcement is worth being clear about, since it’s the thinnest part of the reporting: it happens through peer review, someone asking about a specific sentence. That works at Clay’s size and with Clay’s norms. It’s a social mechanism, not a control, and nobody has published whether it changed behavior.

The pushback

The Hacker News thread on Alpert’s post is small, 31 points and 11 comments, so treat it as a temperature reading rather than a survey. The dissent in it is the useful part. Commenter owentbrown argued that “handwriting docs adds less value than writing high quality instructions to your agent,” which is a real position: if the skill is now specifying well, then the prompt is the thinking and the prose is just rendering.

Alpert’s rule four half-concedes it. If the doc came from a short prompt, share the prompt. What she won’t concede is that a reader should be handed the expanded version and made to reverse it.

The portable version

You don’t need Clay’s document. You need its question, and it fits in one line for your next team meeting: for each recurring document we produce, is more time going into writing it than into reading it?

Anything where the answer is no is a candidate to be cut to the prompt, cut to bullets, or cut entirely. That’s a real operating change, and unlike most AI governance it costs nothing to adopt and produces a shorter meeting the same week.