A lot of the AI-and-work conversation is framed as replacement: which tasks disappear, which roles shrink, which jobs survive. That framing isn't wrong, but it skips past a more immediate effect: AI is compressing the mechanical portion of many jobs faster than organisations are building the judgement portion to fill the space it leaves behind.
The part that's disappearing
Drafting a first-pass proposal, summarising a call, searching for a comparable deal, producing a report: these are increasingly a prompt away. For years, doing these things reasonably well was itself a signal of competence. It no longer is, because doing them is no longer hard.
The part that's left
What's left is the part that was always harder to teach: knowing which of ten AI-generated options is actually right for this customer, this deal, this moment. Knowing when the model's confident answer is confidently wrong. Knowing when to override a recommendation because of context the model doesn't have.
That's judgement. And most organisations have spent decades building systems to evaluate judgement (performance reviews, deal reviews, win/loss analysis) without building many systems that actually develop it.
Why this matters now, not later
As long as the mechanical work absorbed most of someone's time, weak judgement could hide behind busyness. Compress the mechanical work and judgement becomes the visible, differentiating layer much sooner in someone's career, often before they've had the repetitions to build it.
This creates a real risk: junior people producing senior-looking output with junior-level judgement behind it, because the tools closed the production gap without closing the experience gap.
What seems to help
- Deliberate exposure to ambiguous decisions, not just clean case studies. Judgement is trained on situations without an obviously correct answer.
- Fast, honest feedback loops where someone can see the actual consequence of a judgement call, not just whether the output looked polished.
- Treating AI output as a draft to interrogate, not a decision to accept. The habit of asking "why would this be wrong" is itself a trainable behaviour.
What this looks like in practice: writing
The clearest example I've seen of an organisation acting on this is Clay's AI writing policy, published by co-founder Varun Anand. It's short, and notably it doesn't try to detect or restrict AI use. It sets out four principles instead:
- You must stand behind every idea and every sentence. If a reviewer asks what you meant by a line, "sorry, AI wrote that, ignore it" is not an acceptable answer.
- Writing is thinking. Deciding what to emphasise and how to structure it is where you learn the subject. Skip the process and you walk away understanding less.
- More time should be spent authoring a document than consuming it. Generating a long document from a short prompt pushes the cost onto every reader. One person's saved hour becomes twenty people's wasted ten minutes.
- Longer is not better. If you're expanding a short prompt into a long document, consider just sharing the prompt.
What makes this a judgement policy rather than a writing policy is where it puts the burden. It doesn't ask "was AI involved?" It asks "do you own this?" Ownership is exactly the muscle that atrophies when the production gap closes: you can now ship a document you don't fully understand, and nobody will notice until someone asks a follow-up question.
The third principle also names something rarely made explicit. Documents are written once and read many times, so effort you decline to spend is not saved, it's redistributed, multiplied, to your readers. AI makes it very easy to look productive while quietly moving cost onto other people.
There's a nice observation buried in the fourth principle too: there are no lossless transformations of natural language. Every rewrite shifts meaning, and if the rewriting is done by something that doesn't hold a detailed model of what you were trying to say, information is lost. Readers would rather have your thinking than someone else's polish.
I liked it enough that I adapted it into a writing standard for my own team.
Why the "ownership, not detection" framing is the right one
Most organisational responses to AI writing so far have been detection-shaped: disclosure rules, AI-detector tools, bans in certain document types. Detection is a losing game technically, and more importantly it aims at the wrong target. Nobody is harmed by AI having touched a document. People are harmed by documents nobody actually thought about.
Ownership rules are enforceable in a way detection rules aren't, because they're tested by the ordinary mechanics of work. Ask two follow-up questions in a review and it becomes obvious whether the thinking happened. That's a judgement feedback loop, which is precisely the thing the previous section says organisations don't have enough of. A writing policy of this shape is one of the cheapest ones available.
None of this is really about AI. AI has just made it obvious, faster, that judgement was always the scarce resource. We were just able to get away with treating it as optional for longer than we should have.
---
Sources and further reading:
- Varun Anand (co-founder, Clay), Clay's official AI writing policy, LinkedIn: the four principles, the "no lossless transformations of natural language" point, and the ownership-over-detection framing.