Deployable task coverage × Governance coordination

The moat gets built before there's anything worth guarding: narrow systems carry manageable stakes, but heavy compliance costs still land in full, so certification excludes small entrants for risks that don't yet justify the barrier — governance arrives ahead of the danger it's supposedly pricing. The vertical structure then locks against itself: a system certified for medical imaging can't legally share what it learned with a pathology system next door, so the regulatory boundary blocks capability from compounding even within a single hospital. And the narrow, heavily-audited sectors end up doing unpaid PR for everyone else: the public infers AI is being handled from the visible fence around a few domains, while consumer systems in unregulated categories stay exactly as fragile as before.

Capability arrives fully formed and then waits. Broad, dependable systems can already do the work, but coordinated governance decides when the work is allowed to migrate, so productive capacity sits fully built behind a queue whose length is set by committee, not by engineering. The compliance apparatus this creates — auditors, certifiers, monitoring firms — quickly starts lobbying to protect its own revenue, so the industry meant to police the bottleneck becomes invested in keeping it wide. Nations without the capacity to run their own audit boards import an approved registry wholesale from whoever built one first, trading regulatory sovereignty for market access — and the corner that reads as the policy dream quietly resolves into a certified oligopoly where safety and market power are the same lever.

The damage stays survivable mostly by accident: weak coordination lets narrow tools proliferate into spam, biased hiring screens, inflated grading, and grey-market support desks, and because no single deployment is catastrophic, none of it forces a reckoning — harms accumulate as background noise rather than crisis. Small jurisdictions turn that gap into an export product, standing up zero-regulation data havens that host the scrapers and spam operations targeting economies next door. The redeeming accident is that limited capability caps the tuition: organisations get to learn — cheaply, loudly, repeatedly — the habits of weak testing and opaque procurement they will need to unlearn before the tools get dangerous enough for those habits to matter.

This is less an apocalypse than a slow flooding: broadly capable agents outrun jurisdiction, so firms route work to whichever regime lets it move fastest, and each jurisdiction pockets the productivity gain while betting someone else absorbs the displaced. The web fills with synthetic media faster than any chain of custody can track it, so courts, journalism, and peer review lose the evidentiary footing they need to establish plain facts — not because anyone lied especially well, but because verifying anything at all becomes structurally expensive. Credentials stop gatekeeping work that un-audited agents already do end-to-end, and dependable systems get embedded into real operations well before any coordinated incident-response regime exists — so the eventual failure isn't a product recall, it's a cross-border economic event with no agency positioned to own the cleanup.