Clay Has Made an Internal AI Writing Policy Official Across the Whole Company

Co-founder and COO Varun Anand announced the policy this week. It was written by Sophie Alpert, an engineer at Clay, and it was originally meant only for the engineering team. Other groups inside the company read it, found it useful, and asked to be included. It now applies company-wide.

Alpert published the full text on her personal site in June under the title "There are no lossless transformations of natural-language text," which is a fair summary of the argument. Every time a piece of writing gets rewritten or rephrased, the meaning shifts a little. When the rewriting is done by something that does not know precisely what the author was trying to say, some of what they meant gets lost on the way through.

What the policy allows

The policy is not an actual ban. Brainstorming with AI is permitted. Drafting with it is permitted. Proofreading is explicitly permitted. Even pasting raw model output into a document is allowed, provided it is labelled as raw model output. The example Alpert gives is flagging an idea a chatbot raised and asking colleagues whether it is worth chasing.

What it regulates is unlabelled authorship, and it does that through four principles.

The first is that you have to stand behind every idea and every sentence you send. Before a document goes out, it needs to represent what you actually think, in full. The second is that writing is thinking. Deciding what to emphasise and how to order an argument is where understanding gets built, and skipping that work leaves you knowing less about your own subject than you would otherwise. The third is that more time should go into writing a document than into reading it. The fourth is that length is not a virtue. Alpert cites Pascal on the letter that ran long because there was no time to shorten it, then points out that generating length has become the cheapest operation available to anyone with a keyboard.

One sentence does most of the enforcement

The policy also sets out a scenario. A reviewer points at a line in your draft and asks what you meant by it. According to the policy, "Oh sorry, AI wrote that, just ignore it" is not an answer you are allowed to give.

There is no classifier in this policy, no disclosure field, no percentage threshold, no detection tooling. There is a colleague, a specific sentence, and a question you have to be able to answer in the moment. Accountability sits at the level of the individual line, and it gets tested in review by someone who will notice.

Most corporate AI policies are compliance paperwork. Do not paste customer data into a public model, do not upload the roadmap, sign here. Legal owns them, almost nobody reads past the second heading, and compliance is a checkbox. Clay's policy contains no legal exposure at all. It is a standard for the quality of the work, which goes some way to explaining why other departments went looking for a copy of it. Compliance documents do not usually generate demand.

Documents as evidence

The second principle carries a further claim. For a large category of internal writing, the document is not really the deliverable. A technical spec, a project status update, an incident retrospective: these exist to show that somebody thought hard about a problem. The prose is what the thinking leaves behind.

Generate one from a prompt and you have produced the residue without doing the work it was supposed to evidence.

A good deal of marketing work sits in the same category. Positioning documents, campaign briefs, messaging frameworks, the quarterly strategy memo. Nobody reads any of these for the writing. They read them to find out whether a decision was made and what it was. A generated positioning document looks identical to a real one and contains no decision anywhere inside it. It will clear review, sit in the drive, and fail the first time somebody has a hard call to make and goes looking for the reasoning behind it.

The reader maths

One person writes, many people read, and the costs run in opposite directions. Spend four minutes prompting a document that takes nine minutes to get through, send it to thirty colleagues, and you have turned four minutes of your own time into something close to four and a half hours of everyone else's. Time spent tightening a draft before it goes out is paid once and recovered by every reader. Time skipped gets billed to each of them separately, and keeps getting billed for as long as the document stays in circulation.

For anyone running owned media, this stops being a question about internal documents. Publishing at volume works only for as long as each individual piece justifies the read, and the current economics of generation make it very easy to cross that line without noticing you have. Alpert suggests a test for this. If you are producing a long document from a short prompt, think about sending the prompt.

The obvious objection

Clay sells a go-to-market platform. It helps thousands of companies research accounts and personalise outreach at scale, and AI does a lot of that work. Its engineering team has now written a policy against letting AI write the things Clay employees send each other. That reads as a contradiction until you work out where the line falls.

The product does research and assembly, finding signal, enriching records, drafting against a human's stated view of what matters. The policy governs the output of that view. LinkedIn drew a similar distinction in July when it began cutting distribution for generated posts, and said at the time that it was aiming at content with no perspective behind it rather than at the tools themselves.

Whether a line like that survives commercial pressure is a reasonable thing to keep asking. A company whose growth depends on volume has an interest in volume. But the distinction holds together, and Clay has committed it to writing, which is more than most of the industry has bothered to do.

Why it spread

The four principles are not especially novel. Plenty of companies could have written something similar, and a handful have. The route this one took through the company is less usual. Engineering wrote a standard for its own use. Other teams read it and asked for it. Leadership then formalised something that already had support behind it. AI policy does not usually arrive that way. It normally comes down from legal, unrequested, or from the executive team, instructing everyone to use more AI faster, with nothing attached about what the resulting work should look like.

Most companies have spent two years pushing their people to use these tools. Guidance on the output has been considerably thinner. Clay has published its version, and told everyone else to help themselves.