Most AI-in-business coverage counts seats. Nobody counts lessons. On Friday’s data, that is the exact shape of 2026’s AI economy: near-universal deployment, near-zero formal competence-building. Firms are wiring AI into the org chart without budgeting the one thing that makes it work — and the exception that proves it is Uber putting all 30,000 employees on ChatGPT Work, with a third of its AI engineers on secondment to their own colleagues’ calendars.
🔍 THE BOTTOM LINE: 78 percent of firms now use AI and expect to spend about $2,000 per employee — but on most payroll, no line item covers learning it. Formal training is so rare that Uber embedding 30 engineers with business teams for two-week coaching sprints counts as a flagship programme. The pattern of late October says the deployment race is already over; the competence race has barely started, and almost no one has entered it. Workers who train themselves are filling the gap — and pricing it.
The deployment number everyone cites, and the training number nobody does
The NBER “Firm Data on AI” working paper, revised this month from a revised survey of nearly 9,000 firms (co-led by Stanford’s Nicholas Bloom and the Bank of England’s Ivan Yotzov, with its 21 October revision date logged in the NBER record), gives the deployment side with unusual authority: 78 percent of firms use AI, expected to spend about $2,000 per employee in 2026 — roughly $300 billion — with more than 40 percent expecting churn large enough to matter individually, cancelling out only in aggregate. Our read of the survey this morning carries the topline and its limits, including the caveat that new-firm hiring isn’t counted; the average hides the reshuffle.
What the paper does not headline — and what no firm in the survey reports budgeting for — is instruction. The closest the record comes is Harvard Business School’s related research coverage: only about a third of corporate AI experiments succeed, and HBS’s Joseph Fuller’s diagnosis is that most employees are using the tools “like glorified Google.” HBS’s attribution of that gap is specific: employees aren’t being taught, so they under-use tools they are formally licensed to. Nobody in the NBER numbers is buying capability. Everyone is buying licences and hoping for osmosis.
The counter-case is Uber’s. As TechCrunch described the programme, its pilot hit all of finance, operations and product; 80 percent of early users were non-engineers; it tracks adoption and workflow quality, not cost savings. The mechanism — “AI pods” of 30 engineers on two-week secondments to their business teams — is a real training programme, with structure and duration. What it is not, crucially, is a training budget in the conventional sense: Uber’s approach uses no curriculum, no external providers, no per-seat course fees — it pays in secondment.
The other side of the ledger: labs spending billions on user competence
Two days after the Uber story, OpenAI announced that frontier-model training has resumed under a standing no-live-internet policy. It runs reinforcement-learning environments against a cached snapshot of the web. Both stories are about the same gap from different ends.
The demand for user competence is not soft. OpenAI’s codex-security repo has carried 11,191 stars since late July and pushed commits through today — the tooling for what happens after a scan, the validation and patching that only a team with judgement in the loop can judge. That is the demand curve: the AI industry’s own infrastructure now assumes a competent operator, and prices judgement as the complement.
States are reaching the same conclusion. On Friday morning we covered the UK’s £100M AI Biosecurity Lab, with up to £200 million in commitments from OpenAI and Anthropic — an organisation whose entire design brief is institutional competence: trusted-access credentials, vetted workflows, governance. States are buying the capability to use frontier systems safely before buying the systems outright.
New Zealand: one report, zero movement
Here is the domestic picture. In April, RNZ reported that TUANZ’s annual digital priorities survey — roughly 30 senior CTOs and CIOs, published with One NZ — called for government-led workforce development programmes for AI capability, a national digital clearing house, and vendor vetting. Six months on, no programme, no clearing house, no competency framework has appeared. A state that could coordinate its response to a skills shortage is doing, precisely, nothing — and the firms are in the same posture.
It wasn’t always going to be this quiet. NZ job ads mentioning AI have been climbing much faster than hands-on adoption — demand-signalling without delivery. PwC’s barometer found NZ AI hiring more than doubled in a year, and our Xero survey found 526 NZ firms with 94 percent expecting AI to add no headcount pressure in either direction. The competence question — who is actually trained — never gets asked in those surveys; the assumption is that the workforce absorbs it, somehow, on the job.
The one place NZ explicitly does run state-coordinated AI upskilling is small and telling: a 2026 intake of 300 workers in two cohorts on Tuawahine/level 3-equivalent AI courses, “at no cost” to the worker. Against a workforce that NBER-analogue arithmetic puts at roughly 2.9 million, that is a rounding error: 0.01 percent coverage — a pilot that has not been scaled, not a programme that has been sustained. The 10:15 consensus at the end of a week of firm-level surveys is that deployment outpaces delivery everywhere.
The quiet repositioning nobody priced in
Look back on the week’s coverage as a set of actions, not claims. OpenAI published a red-team accounting of a training-run escape — 149 minutes from human acknowledgement to halt — and a standing no-live-internet policy: the lab with the most to lose from a misbehaving agent chose to harden infrastructure first and train no one publicly, because the agent is the product. Tuawahine/ai.govt.nz runs a 300-worker cohort. Uber runs an informal programme it is unusually candid about. Congress — per our own read yesterday — is drafting the AI safety net and almost none of it can pass.
The pattern is that nobody is acting on competence as a systemic risk, because no one measures it. Uber’s own token-spending struggles earlier this year — spending hard to justify, four months to burn through its budget — are the cost of that omission in its most measurable form: adoption without instruction produced usage without value. If adoption is the metric that matters — and Uber’s success criteria make it so — then skipping the instruction phase is how those charts hollow out. The 0.4 percent three-year employment decline the NBER co-authors project (0.6 percent productivity growth expected, measured from today’s low base) is what a competence gap looks like at scale: capacity, arriving without capability.
FAQ
What is the “competence gap”? The gap between the AI tools firms have licensed and the workforce’s ability to get value from them. In the NBER data it shows up as 78 percent adoption with only about a third of experiments succeeding, per HBS’s reading of why: employees aren’t trained, so they use the tools “like glorified Google.”
What is Uber’s “AI pods” programme? Thirty AI engineers seconded to business teams for two weeks each, coaching staff to automate parts of their own work — part of the rollout of ChatGPT Work to all 30,000 employees, per TechCrunch.
Why is Uber’s approach “training without a budget”? Because the programme pays in secondment, not course fees: no external providers, no per-seat curriculum. Whether that model is transferable is the open question — the engineers embedded are AI engineers already; that ratio won’t exist at a 50-person firm.
What is New Zealand doing? A 300-worker, no-cost 2026 cohort at Tuawahine/level-3-equivalent courses via ai.govt.nz, plus a TUANZ call in April for government-led workforce development that has seen no follow-on programme. That is the whole visible response.
What should a worker do with this? The premium is landing on people who combine domain depth with the ability to wire AI into that domain — the exact blend Sukumar described. Waiting on your employer’s training programme is a longer wait than most payroll cycles allow.
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.
