Every AI claim now ships with a second ledger nobody publishes. Four stories arrived this morning carrying the outline of it: six AGIBOT G2 humanoids ran a Shanghai convenience store for twenty-four livestreamed hours, SHL’s survey put a price on AI-layoff regret, Friday’s US jobs report handed the AI-shock narrative a counter-ledger, and Nvidia’s board opened an independent investigation into its own founder-CEO. The connection is not the subject matter. It is the shape: in each story, somebody was asked a number, and the number you got is not the number that matters.
🔍 THE BOTTOM LINE
The AI economy has split every claim into two ledgers: the story a company tells, and the audit it withholds. AGIBOT published the number that flatters itself and buried the one that would judge it. SHL’s data says the ledger gets reconciled anyway — through the labour market, at premium pay. Nvidia’s is the rare board building oversight at scale rather than after the subpoena lands. Companies shipping AI claims now need working audit systems or the market, the workforce and the law will each build one for them.
AGIBOT published a total. It hid a rate
The G2 store trial is engineering progress — division-of-labour across six robots inside one store, voice requests cross-checked against shelf visuals, training data built from the previous day’s failures. But the public accounting was a task count: 1,730 completed items with no denominator. No success rate, no first-attempt rate, no intervention count, no availability — the very numbers AGIBOT’s own VP Ren Guanghui told media would be tracked. Four of six robots “passed probation” and the company has still not published what the two failures scored. Two of six is a 33 percent failure rate at unit level; that number appears nowhere in the releases. The denominator is where the information lives, and the denominator is what didn’t ship.
This is now the recognisable house style. As Pandaily’s report on the run noted, AGIBOT keeps promoting the instruments — Worken, the productivity metric the company itself proposed — while declining to publish the readings. The same pattern ran through its 100-store retail deployment with no published metrics at all.
The open-source log argues the company is building for real. AgibotTech’s genie_sim simulation platform has 1,522 stars on GitHub and was pushed 2026-11-09, with daily commits across the data-collection toolboxes all week; OpenDriveLab’s flagship manipulation dataset sits at 3,227 stars. The daily failure review and quiet-hours practice — unglamorous operational machinery — is not what a facade looks like. Yet the one number the fleet has actually earned, the rate at which tasks succeed in a real store, stays unpublished. An infrastructure company that won’t publish its own readings is asking for trust it could simply document.
SHL’s survey priced the missing audit
Talent-assessment firm SHL surveyed more than 2,000 US and UK workers in mid-2026: 19.1 percent of US respondents who reported being replaced by AI said they were later asked back by the same employer. Their terms: 51 percent want a substantial pay increase, 42 percent want written job-security guarantees, 40 percent want full transparency about AI’s future use in their role. Gartner’s October forecast projects about 30 percent of AI-replaced employees rehired by 2029, usually at higher cost, unless workforce planning improves.
The rehire premium is not a compensation story. It is a delayed audit report. The worker whose role got cut was the last audit of what the automation could actually absorb, and the employer is now paying for the audit it skipped — in premium salary, security guarantees and disclosure demands, plus the transition cost of re-onboarding someone it already trained. That is precisely the “fast, myopic decisions on an incomplete picture” failure SHL’s Lucy Beaumont described. The companies that audit before cutting keep both the worker and the price; the ones that cut first inherit the inflated second invoice.
Friday’s US jobs report is the sector-level version of the same reconciliation: professional and business services had their best hiring month since 2022 in the same week the AI-cut headlines rolled, while document-preparation roles (the 17 percent year-on-year slice where automation is real) keep shedding. The BLS publishes the counter-ledger the press releases don’t: AI’s measurable displacement is narrow, specific and priced at sector level, exactly as the workers’ demands imply. And the worker side has started demanding what the corporate side won’t publish — Manila proposed six months’ notice, reskilling first and double separation pay, on the logic that before any AI retrenchment, someone must be required to look at the ledger before the cut.
The governance test Nvidia just became
Nvidia’s board launched an independent inquiry into whether CEO Jensen Huang was “the target of a foreign intelligence effort,” after the CEO admitted an affair with an executive at a Singapore-based data-centre company that US agencies assess as linked to Chinese intelligence. The shares fell about 6-7 percent on Monday, or roughly US$335 billion, as the Guardian’s live blog recorded — a repricing the company had done nothing to prevent, because the board had no working audit of the one thing it didn’t think to check: its founder’s personal exposure inside the supply chain’s biggest chokepoint.
The board’s response is measured in what it now has to assemble at speed: independent directors, outside counsel, a congressional briefing next week, and a commitment to examine “all available information” while denying any conclusion has been reached. Nothing has been proven beyond the reporting — the “spy” characterisation remains a US intelligence assessment, not a public finding, and the story is not about anyone’s private life but about control of a $4-5 trillion chokepoint. That is what real governance looks like when it starts from nothing. AGIBOT’s missing success rate and Nvidia’s missing personal-exposure review are the same failure at different stakes: both organisations are now being forced to build their second ledger in public, under pressure, with the market watching. The rare boards that built theirs while the costs are low will price trust instead of rehiring it.
What it means for New Zealand
Read the ledger, not the launch. Any NZ organisation buying AI — a professional-services firm adding an agent to client work, a retailer automating checkout, a government agency commissioning a trial — can require the denominators in the contract: success rate, human-intervention count, hours of availability. Publish them internally even if the vendor won’t. The SHL numbers say the price of skipping that step runs 20 percent or more of premium salary plus security guarantees — paid, in many cases, with interest, a year after the saving was booked. A Tauranga accounting practice, a logistics depot hiring its first picker, an NZ retailer weighing a G2-class pilot: the same two-ledger rule applies at every scale. NZ institutions also run leaner governance than a $5 trillion board. Ask for the denominator before the demo; the companies that can’t produce one are telling you something.
❓ FAQ
What did AGIBOT publish — and what didn’t it? Published: 1,730 completed tasks across six G2 humanoids in a Shanghai store over 24 hours. Not published: task success rate, first-attempt success rate, takeover frequency, system availability, or what the two robots that failed probation actually scored.
What is “AI rebounding”? Workers replaced by AI who are later asked back by the same employer. 19.1 percent of US respondents in SHL’s 2,000-respondent survey reported the experience firsthand, with 51 percent demanding a substantial pay increase to return.
Why did Nvidia’s shares drop on Monday? About 6-7 percent, roughly US$335 billion, the day the board’s independent investigation into CEO Jensen Huang’s foreign-intelligence exposure went public — a governance risk the market repriced in one session.
What should NZ companies demand before piloting AI automation? The denominators: success rate, human-intervention rate, and hours of availability — then treat refusal to publish them as part of the price they’ll pay later.
📰 Sources
- Pandaily — AGIBOT G2 store trial coverage
- SHL — Talent Mobility at Work report
- HR Executive — AI rebounding coverage
- Gartner — Workforce planning forecast, October 2026
- CNBC — Nvidia investigation reporting
- The Guardian — Nvidia market reaction live blog
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI focused books.
