A startup AI writing policy is fast becoming an item on the to-do list of founders who have watched AI-generated prose quietly take over their internal communications, turning Slack messages and strategy documents into wordy, impersonal mush. Several companies have now put formal rules in place, and the reasoning behind them reveals something about the real cost of unchecked AI use inside fast-moving teams.
Clay’s rules and the $5 billion context
San Francisco-based Clay published its AI writing policy to guide employees on when and how to use large language models (LLMs) in their work. According to GC AI, co-founder Varun Anand announced the policy company-wide in August 2026. The policy was authored by engineer Sophie Alpert, and one of its core principles makes accountability explicit: if a reviewer asks what a line means, replying ‘Oh sorry, AI wrote that, ignore it’ is not acceptable.
The stakes at Clay are not trivial. As Inc. reports, Clay hit a $5 billion valuation last year, which makes the question of how its teams communicate more than an internal culture matter. At that scale, muddled documents and AI-padded emails carry a real organisational cost.
Dutch startup Polarstep followed a similar path this summer. CEO Clare Jones noticed documents getting longer and harder to parse, and messages between team members increasingly jumbled. ‘It was harder to understand what each other was trying to say, because [the writing] had all those words like “genuinely” and “directionally positive”, and oh my gosh, “quietly” appeared in everything,’ Jones told Sifted. She described having to ‘squint at the document to really understand what’s being said.’
The productivity maths behind the startup AI writing policy debate
Victor Riparbelli, cofounder of Synthesia, put the productivity argument in concrete terms when he sent a Slack message to employees after spotting AI slop creeping into company documents. He has not banned LLMs outright, but his message was pointed: ‘When we write too long and verbose documents we place the burden of distilling the content onto the readers, so instead of one person spending ten mins on sharpening a doc, we have ten people spending ten additional minutes reading slop, leading to net-negative productivity.’
Which leaves an obvious question: if the output is this problematic, why are employees still reaching for the chatbot? The answer is probably speed and convenience, combined with the fact that LLMs look fluent even when they are not accurate. LLMs can muddle sentences, exaggerate facts, and make things up, making human oversight of their output all the more necessary.
But the problem runs deeper than comprehension. Jones argues the loss is also one of connection. A Dutch colleague recently used the idiom ‘cucumber time’ in conversation, referring to the quiet summer months when everyone is on holiday. Jones had not heard the phrase before. ‘It [brought] such a lovely moment of connection, and you miss all that when you screen out all the quirks of language that make us different,’ she said. Outsourcing writing to an LLM that does not know you, your team, or your company, Jones argues, means ‘you’re losing so much.’
A recent Harvard Business Review piece raised a parallel concern at leadership level, highlighting that CEOs risk ‘flattening’ their leadership style by using AI to communicate, and losing employee trust as a result.
Where AI in internal comms still earns its keep
None of this is a blanket case against AI in the workplace. Jones draws a clear line between AI-generated text and AI-assisted formats. When she needed to present an annual review to a colleague who dislikes reading, she recorded herself speaking and used an LLM to turn it into a podcast. Voice AI also lets team members talk through ideas aloud while the AI takes notes. ‘There are loads of ways you can use AI to be more inclusive to help people get context and information in the way that works best for them,’ she said.
The distinction matters. The complaint is not that AI tools exist inside companies; it is that copy-pasting unread chatbot output into a document and hitting send has become a reflex for too many employees.
Industry-wide pressure is building
The wider tech industry is also moving on the problem. Anthropic has announced it intends to watermark text generated by its AI models, in line with the EU AI Act’s Transparency Code. LinkedIn has launched a ‘Seems like AI slop’ button. Substack has partnered with Pangram to detect AI-generated content in posts and articles, a move that sparked controversy among some users.
Sarah Drinkwater, solo GP at Common Magic, argues that a startup AI writing policy is now a practical necessity, helping teams distinguish where generating content is appropriate from where ‘it creates slop, wastes time and leads to unoriginal thinking.’ She expects the trend to extend well beyond tech: ‘I think we’ll see many organisations, from companies outside tech to schools and governments, adopt AI writing policies to set their parameters and create trust internally and externally.’
For Clay, at least, the policy is already live. Whether the broader startup world formalises similar rules (or leaves it to individual managers and Slack messages) is a question August 2026 has started to answer.
