One sentence in Anthropic’s own employment report quietly repriced the AI economy: at the most agent-saturated workplace studied, machine hours outnumber human hours six to one, and the scarce input is not compute or talent but human attention. Set that beside this morning’s other coverage, and the pattern holds: the UK’s public bodies are drowning in machine-written correspondence, and the US office that reads frontier models’ test results has lost its independence and most of its staff.
🔍 THE BOTTOM LINE
The productivity story everyone priced assumed the expensive half was producing output. Three datasets this month say it has quietly become checking it. At Anthropic, agents write faster than staff can read their output; in Britain, tribunals absorb claims generated in seconds, answered in months — single claims up 28 percent year-on-year; at the Commerce Department, the office verifying frontier models has no permanent director and a White House line. Verification — reading, checking, adjudicating, testing — is becoming the priced service of the AI economy. Almost none of it is priced yet.
The maker-checker split is now visible, and the makers outnumber the checkers
Anthropic’s report is self-reported, one company — and still the clearest published topology of AI-native work. Agent hours passed human hours in March; by August the ratio was six to one, with output per person rising. But the binding constraint moved somewhere no hiring plan budgeted for: even non-technical employees would spend half their week just reading what their Claude agents produce. One engineer described facing “an army of Claudes, with every one asking for my attention.” The company’s answer: hire for “autonomy and judgment” over technical expertise — junior hires fell to 9 percent of technical intake as hiring kept growing.
New Zealand’s job market shows the same split from outside. The country now holds an annual AI awards night, with roughly 2,300 mapped roles and salary bands clustered between $74,000 and $220,000, as our awards coverage tracked. But the fastest-growing asks in that market, per SEEK’s September gauge, are governance and oversight roles: reading what the machine produced. The demand the labs created on the maker side has arrived first on the checker side.
The checkers’ side of the firewall is already drowning
The UK shows what happens when nobody prices the readers’ queue. The Guardian’s weekend investigation documented councils, MPs’ offices and transport bodies swamped by chatbot-written complaints — pages becoming ten-page documents, as our own report carries. The government’s own statistics confirm the direction beyond anecdote: single employment tribunal claims rose 28 percent year-on-year in the April–June 2026 quarter — the highest open caseload since 2013/14 — while disposals fell 6 percent. Claims made in seconds; hearings scheduled in months. The Information Commissioner’s Office guidance on such requests rests on a blunt premise: they still must be answered. The machine side pays nothing per additional claim; the human side pays per additional page.
Agentic flooding is a market failure in miniature: machine output is nearly free to make, no cheaper to consume, and the consumer never consented. An economy that ships output to the unconsenting books a bill it never counted.
The referee’s office was next
The squeeze reaches the office that reads the test results. Per the AP on November 8, staff at the Commerce Department’s frontier-model testing centre, CAISSI, were asked to report informally to a White House office — two directors gone since April, a budget one-eighth of its UK counterpart’s, results unpublished since August, as our reporting set out. The UK’s AI Security Institute runs at arm’s length and publishes its findings. The three instances — a hiring mix, a backlog, a governance dispute — share one shape: the caseworker reading ten-page complaints, the engineer supervising agents half the week, the public never seeing test results. Our reporting on the invisible frontier documented capability moving behind closed doors while capital prices an older public reality; apply that gap to verification and you get a checkless economy — output everywhere, assurance nowhere in the price.
What it means practically for New Zealand
New Zealand is downstream of all this but not insulated. First, the job ladder: the mapped NZ market clusters at $74,000 to $220,000, but the growing end is oversight, not creation — reskilling should target supervision skills, not assume AI fluency converts to checking capability.
Second, flood defence. The UK’s “AI clog” will arrive in NZ inboxes — council submissions, OIA requests, tribunal claims — and the preparation is triage capacity, not ban-and-block. Meeting it with an explicit, funded triage protocol, as the Information Commissioner’s Office guidance does, turns the flood into a workflow; silence inherits the backlog.
Third, domestic verification: NZ’s early AI deployments show the same deploy-first pattern the UK is scaling up, a critique made from home ground months ago. The scarce skill in 12 to 24 months is not prompting; it is priced, professionalised verification.
❓ FAQ
Why does Anthropic’s 6:1 ratio matter if it is just one company? The topology is the finding: the human job became supervising machine output. Anthropic published the report itself, flagging its own limits — runtime hours, not effort, no external audit.
Is the tribunal surge actually AI’s fault? The official quarterly data shows the surge, not the cause: single claims up 28 percent year-on-year, with the bodies processing them reporting AI-written submissions as the visible driver. Direction confirmed, attribution still argued.
Who is accountable when the checking office stops publishing? CAISSI’s evaluations are voluntary — nothing compels the labs. As our CAISSI reporting set out, a testing unit whose reviewers answer informally to the White House weakens the signal those agreements were meant to create; the UK institute still publishes.
What should an NZ worker actually do with this? The shift is toward judgment-heavy oversight — verification, review triage, incident escalation — exactly where NZ market growth is concentrating.
🔍 THE BOTTOM LINE
Three datasets, one shape. The labs’ own data says machines now outproduce the humans supervising them; a government’s tribunals data says machine-written claims are landing on unpriced queues; an administration’s restructuring says formal checking can be swapped for informal attention when convenient. The 12–24 month direction is not more AI output — that is already cheap at the margin. It is scarcity repricing everything else: the readers, the checkers, the adjudicators, the testers. The companies selling machines have priced tokens, seats and compute. Nobody has priced the inbox it all lands in. In that gap, the arbitrage belongs to whoever turns verification into a product — and the first movers will not be the biggest labs, because the biggest labs have the largest unsupervised surface to defend.
📰 Sources
- Anthropic — Employment at Anthropic (internal report, October 2026)
- Startup Fortune carrying AP — US AI safety center staff now report informally to the White House (8 November 2026)
- Ministry of Justice — Tribunal Statistics Quarterly, April to June 2026
- The Guardian — ‘AI clog’: UK public bodies battle onslaught of chatbot-generated complaints (8 November 2026)
- ICO / Civil Service World — guidance on AI-generated FOI requests (May 2026)
- SEEK NZ — September AI skills gauge (September 2026)
