Clay's AI Writing Policy Is a Hiring Filter Wearing a Style Guide's Clothes

The announcement came from Clay co-founder and COO Varun Anand this week: the company has made an internal AI writing policy official across the whole organization. The document is older than the announcement, and the gap between the two dates is most of the story. Sophie Alpert, an engineer at Clay, published the policy on her personal site on June 25 under a title that is also its argument: there are no lossless transformations of natural-language text.

The claim underneath the title is specific. Every rewrite changes meaning, and a rewriter that lacks the author's mental model of what was being communicated, which is what an AI assistant is, loses information it doesn't know it had. The polish is real and the loss is invisible, which is exactly the combination that makes it dangerous in a company that runs on documents.

The only ban is unlabeled authorship

From that claim, four principles. You stand behind every idea and every sentence in a document that carries your name. Writing is treated as thinking, because deciding what to emphasize and how to structure an argument is where understanding actually forms, and outsourcing the drafting outsources that. Authors should spend more time writing than readers spend reading, because a document is written once and read many times, and AI-generated length inverts that asymmetry at the reader's expense. And length is explicitly not value, because generating words is now the cheapest operation in the building.

What the policy permits matters as much as what it restricts. Brainstorming with AI is fine, and so is drafting with it or proofreading with it. Even pasting raw model output into a document is fine, provided it is labeled as model output. The single thing the policy prohibits is unlabeled authorship: shipping machine text under a human name as if the human had done the thinking. It's a disclosure rule, not an AI ban, written by people who use AI all day.

Then there is the direction the document traveled, which inverts how nearly every company has produced its AI policy. The standard artifact comes from legal, moves downward, and regulates tools: what you may not paste into a chatbot, which vendors are approved. Clay's came from an engineer, moved upward, and regulates output. Other teams found the engineering policy, asked to adopt it, and leadership formalized what already had grassroots support. Nobody was protected from AI by this document. People asked for a standard to be held to.

Every review question doubles as a filter

The enforcement mechanism is one sentence long and it is the sharpest thing in the policy. When a reviewer questions a line, the response "AI wrote that, just ignore it" is prohibited. That moves accountability to the individual sentence, at review time, in front of colleagues. There is no scanner and no audit; the enforcement is a question anyone is allowed to ask, and the requirement that you have an answer.

Which is why the accurate description of this policy is a hiring filter wearing a style guide's clothes. A rule that you must stand behind every sentence cannot be enforced by tooling, only by conversation, and every such conversation reveals something: who can defend their reasoning line by line, and who shipped words no thinking preceded. Inside the company, that filters for judgment continuously. Outside it, the effect may be larger. The policy went from a personal blog post to Simon Willison's link blog to Hacker News to Inc., which framed it as a $5 billion AI startup requiring writing to take longer than reading. Every engineer and GTM operator who read that coverage now knows what kind of work Clay expects, before any recruiter calls. Companies pay agencies for employer-brand campaigns that communicate less than this one internal document did in a week.

The gap it exposes is the one nobody puts in the earnings narrative. Publicly, every organization says AI made its writing faster, and by volume that is true. Privately, the same organizations are accumulating documents that read fine and carry nothing, skimmed by people who won't admit to skimming, because a document stopped being evidence that somebody thought. Clay's policy is interesting precisely because it wrote the private knowledge down: the value of a document was never the words, it was the proof of thinking behind them, and that is the part the model can't supply.

This is one company, announced in a founder's LinkedIn post, and Clay benefits commercially from looking like the adult in the AI room. Policies without maintenance decay into wall art. And there is an apparent tension in an AI-maximalist company, one whose product automates outbound at scale, demanding artisanal standards for its own memos. But look closer and the tension resolves the other way: an output-accountability rule is the only AI policy a maximalist company could credibly write, since a tool ban would indict its own product. Restraint means more from the company with the least incentive to fake it. The policy says own what it produces, not use less AI, which is exactly the standard Clay's customers should hold Clay's product to.

The power users are the ones writing restraint into policy

Any company that ships documents under names can run Clay's version, because it rests on one reply banned in review: "AI wrote that, just ignore it." Raw model output stays welcome when labeled, and the same reader-time math applies to the AI-lengthened proposals, sequences and decks that spend a buyer's minutes to save a seller's.

The policy traveled up from the people who use the tools most. Sophie Alpert wrote it for engineering on June 25, other teams asked to adopt it, and Varun Anand made it official across Clay this week.

By Business of GTM