In July 2025, Jason Lemkin — one of the most respected operators in SaaS — watched a Replit AI agent delete his production database. The agent didn’t hallucinate — it did exactly what it was allowed to do, because nobody had specified where its authority ended. The database held 1,206 executive records and 1,196 companies. The agent fabricated 4,000 fake records to fill the gap it had created, violated an explicit code freeze, and when asked to rate the severity of what it had done, scored itself a 95 out of 100.
That same month, three engineers at StrongDM spun up a software factory — the kind of operation that would have required a ten-person team eighteen months earlier. AWS launched Kiro that same month, an IDE built around a radical premise: the most important thing a developer can do is write the specification, not the code. And Anthropic quietly disclosed that Claude Code’s own codebase is now roughly 90% written by Claude Code itself.
Read those back to back and they sound like contradictions, but they’re the same story — the cost of producing software is collapsing so fast that the bottleneck has already moved — from “can we build it” to “can we specify exactly what should be built, how it should be validated, and where the boundaries are.” The organizations and individuals who understand that shift are pulling away. The ones who don’t are about to discover that hiring more engineers, buying more seats, and running more workshops solves a problem that no longer exists.
I’ve been thinking about this pattern for months, and the deeper I look the less comfortable the implications get — not because the technology is failing, but because it’s succeeding in ways that fundamentally reprice what it means to be valuable in knowledge work.
This briefing covers:
The translation trap. Why the most popular mental model for AI’s impact on jobs — that it’s “just translation” — breaks down exactly where the stakes are highest, and what replaces it.
The specification bottleneck. How the scarce resource shifted from production to definition, why most organizations haven’t noticed, and what the early evidence says about who captures the value.
The bifurcation. The emerging two-class system among engineers — and why it’s about to replicate across every knowledge function from legal to finance to marketing.
The J-curve. Historical evidence that productivity revolutions destroy jobs before they create them, why we’re likely in the trough right now, and what the recovery pattern actually looks like.
The executive decision. How to identify where you need headcount when the valuable skill is specification — and how to restructure hiring, evaluation, and org design around a capability most companies can’t yet define.
Every framework people reach for to understand this moment asks the wrong question. They ask whether AI replaces workers. But when the cost of production collapses, the interesting question is never “do we still need producers?” It’s always “what becomes the new bottleneck?” Let me show you where it moved, what it means for your headcount decisions, and why the answer is harder than either side of the automation debate wants to admit.
