Ninety-one per cent of New Zealand organisations are now using artificial intelligence in some form. That is up from eighty-seven per cent last year and sixty-six per cent the year before, according to Datacom's fourth annual State of AI Index. It closes a debate that has run through most of this series since February: whether New Zealand businesses would catch up. They have. The adoption gap is gone.
In the same survey, the share of organisations using artificial intelligence to transform their core operations fell, from eight per cent to four. Eighty-one per cent remain stuck in exploratory or implementation stages. Fifteen per cent report scaling the technology organisation-wide, up from twelve. Read the two figures together, the way Datacom's survey presents them, and a different story appears. Adoption rose by a quarter in a year. The number meant to show that adoption was changing the business fell by half.
A figure that halves while the figure beside it rises is a cohort that tried something, counted it honestly, and stopped. As I wrote in The Manifesto of the Humanist Party: "That is not a headline number. Nobody is going to build a keynote presentation around a 17 per cent reduction in administrative burden. But buried in the same briefing was a second pilot, and this is where the story becomes instructive." The 91 per cent figure is this year's keynote number. The four per cent figure, and what sits underneath it, is the second pilot.
What sits underneath it is a barrier ranking, and the order is not the one most boardrooms assume. Datacom's respondents named internal capability gaps as the leading barrier to scaling AI, at twenty per cent. Second, within two points of the leader, is employee resistance or fear of job displacement, at eighteen per cent. Third is implementation cost, at seventeen per cent.
Capability gaps are the barrier every organisation already has a plan for: hire, train, contract, repeat. Resistance is not that kind of problem. An organisation that answers a four per cent transformation rate by appointing a head of AI has addressed the barrier worth twenty percentage points, the one with the obvious procurement response. It has left the barrier one rank below, worth almost as much, entirely untouched. You cannot hire your way past a workforce that does not trust what it is being asked to use.
It helps to be precise about what "transform core operations" is actually asking. The four per cent figure is easy to misread as a verdict on how much AI is in use. It is not that. An organisation can sit inside the ninety-one per cent and still answer no to this specific question. Using a chatbot to draft an email is not the same claim as rebuilding a claims process, a supply chain decision, or a credit assessment around a model's output. Datacom's own fifteen per cent scaling organisation-wide and four per cent transforming core operations are measuring two different depths of the same word. Most of the ninety-one per cent sits at the shallow end: individual tasks, individual tools, no change to how the organisation actually decides anything. That shallow end still does real work, just not the work the keynote number claims.
Datacom's own AI director, Lou Compagnone, has a name for the pattern underneath the resistance figure. It is a better description than anything this series has produced for its own concept of ad hoc deployment. Sixty-eight per cent of organisations in the survey are relying on general-purpose assistants such as ChatGPT, Gemini and Claude rather than any coordinated rollout. Compagnone calls the result "AI confetti": individually chosen, scattered across desks, with nobody above them able to say what the organisation, taken as a whole, is actually doing with the technology. It is a useful correction to a concept this series has reached for before under a different name. That name belongs to a different series and a different problem: concealed, unauthorised use. What Compagnone is describing is the opposite. Nothing here is hidden. It is uncounted, which from a governance seat looks exactly the same.
The same pattern shows up outside New Zealand. Gartner's IT Symposium and Xpo opening keynote in Orlando, on 19 October, reported that AI project success has stayed flat at twenty per cent, one in five, unchanged from last year. Spending against it has risen eighteen-fold over two years. Eighty-six per cent of the CIOs surveyed say risk is growing faster than value. Seventy-one per cent expect to deploy agentic AI within twelve months. Twenty-one per cent say they have governance structures in place for it.
That gap, seventy-one against twenty-one, is the headline most outlets covering the keynote reported. Neither survey nor most of the Gartner coverage led with the figure that explains it. Only twenty-five per cent of CIOs say their technology risk management and their business risk management are actually aligned. That figure comes from a separate survey of 303 CIOs and IT risk leaders within the same release. Three in four are not. That is the mechanism behind the eighty-six per cent figure. When the people who own the technology risk and the people who own the business consequence are not the same conversation, a portfolio can spend eighteen times more than it did two years ago. It still succeeds on exactly one project in five. Nobody with the authority to stop a failing pilot is the one measuring whether it is failing.
Gartner analyst Kristin Moyer put the mechanism in a single sentence at the same keynote: "It's not possible to create value faster than AI can create risk." Gartner's own answer is something it calls the Value-Risk Flywheel, described in its release as balancing rapid adoption against governance the way a racecar balances speed against drag. Four per cent is what a portfolio looks like when nobody owns that balance. Datacom's eighteen per cent resistance figure and Gartner's eighty-six per cent risk figure are two different instruments, on two different continents, describing the same organisation from two different seats. Neither is wrong on its own; together they agree.
A stopping rule is three decisions written down before a pilot starts, not argued over after it has already cost money to kill. First, the metric: what number, measured on what cadence, tells you this specific pilot is or is not working, agreed before launch rather than debated once the invoice arrives. Second, the threshold: the value of that metric below which the pilot stops. State it as a number, not as "when we are not happy with it." Third, the authority: one named role, not a committee, who can act on that threshold without a second approval cycle. A stopping rule that needs a steering group to convene before it fires is a delay with better paperwork. Gartner's own twenty-five per cent risk-alignment figure is, underneath the label, a count of organisations that have put the technology owner and the business owner in one room. That room answers all three questions together. New Zealand's own capability-gap barrier, the one sitting at twenty per cent, is usually read as a hiring problem. Read against that structure, it looks different. It is a decision-rights problem dressed as a skills shortage: the people with the AI capability are rarely the people with the authority to stop a pilot, and the people with that authority rarely have the capability to know when the metric has crossed the threshold. Fixing that pairing costs nothing to pilot and nothing to measure, and it needs no procurement cycle to test.
This figure proves less than it seems. It does not prove that artificial intelligence is delivering value for New Zealand business, and it does not, in the other direction, prove that New Zealand's AI adoption is hype. Eighty-one per cent of organisations sitting in exploratory or implementation stages are doing the slow, unglamorous work that scaling actually requires, and some will get there. The specific claim this survey supports is narrower than either reading: core-operations transformation, by Datacom's own measure, moved from eight per cent to four. Datacom has a commercial interest in selling AI leadership and implementation services, but that should not be read into the barrier data itself. The twenty, eighteen and seventeen per cent splits are an independent finding, whatever conclusion the vendor publishing them might prefer.
The employee-resistance figure fits a wider, rational pattern. Gallup's most recent sentiment research, fieldwork from early May 2026, found thirty-nine per cent of Americans now say artificial intelligence does more harm than good, up from thirty-one per cent the year before. That figure rises to forty-seven per cent among eighteen to twenty-nine-year-olds. Seventy-nine per cent expect the technology to reduce the number of jobs available over the next decade. Two countries, two different instruments, and the same resistance shows up from two different angles. No better internal memo fixes it. On the technical side the same week, Google's Gemini 4 Argon reached general availability on 13 October at two US dollars per million input tokens through the end of November, doubling from 1 December. The model-routing decisions behind these deployments carry their own dated deadlines.
The Hamberger Report: Generative AI 2026 put a version of this same shape on the page before either of this week's figures existed. Late in 2025, twenty-three per cent of organisations had scaled agentic workflows while only eleven per cent maintained validated oversight frameworks. This week's numbers show the identical gap. It is one tier further on. The distance between adoption and oversight has widened, because adoption was always the easier number to move.
The same gap between what is adopted and what is governed shows up in the tooling teams reach for when resistance and risk need an answer. The OWASP Foundation's Top 10 for Large Language Model Applications, now in its third annual revision, is a community-maintained, openly licensed risk taxonomy built by practitioners across dozens of organisations. It exists because two teams naming the same failure differently cannot share one risk register. The same community habit built the CVE numbering scheme for conventional software years before generative models existed. That open vocabulary, not a proprietary scanner, lets a governance team and an engineering team read the same incident the same way. An organisation on the wrong side of Gartner's seventy-one to twenty-one split does not need to commission a framework before it starts. A peer-reviewed one is already published, free, and in its third revision.
The same openness is what makes a model checkable. A model's weights, once published, can be audited by anyone with the compute to run them. That is why the European Union's AI Office treats open-weight general-purpose models differently under the AI Act from closed systems only their vendor can inspect. Its systemic-risk function is built around compute thresholds rather than national borders, a governance answer to the same split Gartner measured. New Zealand has no domestic frontier lab and no equivalent oversight body of its own. An open framework like the Office's raises a practical question: how a smaller economy inspects the control plane of systems built elsewhere before trusting them with its own critical functions. Four per cent is a domestic figure; the governance gap behind it is global.
This week's two surveys leave one question on the desk. It is not whether to adopt; that question is closed in both countries. The real question is whether anyone in your organisation can name the specific condition under which a live AI pilot gets switched off. It is also whether anyone has the authority to pull that switch before the spending finds out the hard way. If the answer is no, the four per cent figure already applies to you. You have not measured it yet.
If your organisation is moving AI agents from pilot to production and nobody outside the vendor has inspected the control plane, message me. I will send the scope and the fixed fee for an independent review.
The views expressed in this article are entirely my own, informed by morethan 30 years of professional experience in architecture, security, andtechnology leadership in New Zealand. I write as director of Te PonoLimited; the views are personal and do not represent the position of anyclient, any government agency, or the New Zealand government. My commentaryon legislation and policy is analytical, drawing on publicly availablesources and my professional expertise in architecture, security, and AIgovernance, and it is politically neutral.
Andreas Hamberger is a New Zealand leader in Architecture & Security and Associate Member of the Institute of Directors. The Hamberger Report: Generative AI 2026 provides enterprise leaders with evidence-based analysis of the AI landscape. Through Te Pono he provides independent reviews of agentic AI control planes for organisations moving from pilot to production; contact andreas@thehambergerreport.com for the scope and fixed fee.
This article was produced with AI assistance under my direction. Research,drafting and images pass through a pipeline I built and govern: automatedgates for source verification, forbidden language and political neutrality,and my own review before anything is published. The tools include Claude,Gemini and Openart. The frameworks, arguments and editorial judgements aremine and are the same discipline I apply to the AI systems I audit forclients. AI accelerated the work; the thinking, and the responsibility forit, are mine.
[1] Datacom. "2026 State of AI Index." Reported in Management Magazine NZ, "AI uptake rises but leadership gap a barrier to AI at scale," 3 September 2026. https://management.co.nz/technology/ai/ai-uptake-rises-but-leadership-gap-a-barrier-to-ai-at-scale
[2] Datacom. "2026 State of AI Index." Reported in B2B News NZ, "NZ AI adoption 91 percent but only 4 percent transform core ops," 29 August 2026. https://b2bnews.co.nz/news/nz-ai-adoption-91-percent-but-only-4-percent-transform-core-ops/
[3] Gartner. IT Symposium/Xpo opening keynote, Orlando, 19 October 2026. Reported in CIO Dive, 19 October 2026. https://www.ciodive.com/news/gartner-keynote-symposium-2026/833398/
[4] Gartner. IT Symposium/Xpo opening keynote, Orlando, 19 October 2026. Reported in InformationWeek, 19 October 2026. https://www.informationweek.com/responsible-ai/gartner-cios-must-balance-ai-value-with-risk-and-fast
[5] Gartner. IT Symposium/Xpo opening keynote, Orlando, 19 October 2026. Reported via Business Wire, syndicated on Yahoo Finance, 19 October 2026. https://finance.yahoo.com/technology/ai/articles/gartner-says-ai-risk-rising-150000028.html
[6] Google. "Gemini 4 Argon." General availability 13 October 2026. Reported by Lindy, 2026. https://www.lindy.ai/blog/gemini-4-argon
[7] Gallup. "Americans Cool Toward AI." Poll 712751, fieldwork 4-11 May 2026. https://news.gallup.com/poll/712751/americans-cool-toward.aspx