Return to Part 0: Table of ContentsPrevious Article: Part 39, AI Governance: The Number Gets Worse Before It Gets Better
Forty per cent of organisations with more than a billion US dollars in annual revenue now say they are scaling AI agents. That is up thirteen points from twenty-seven per cent a year earlier. It is the figure every intelligence stream carried this week. Almost none of them carried the line directly beneath it in McKinsey's own report: organisations below that revenue threshold stayed exactly where they were. Twenty-two per cent, flat.
The Hamberger Report: Generative AI 2026 named this shape before McKinsey re-measured it. Writing from 2025 data, the book put it plainly: twenty-three per cent of organisations had scaled agentic workflows. Only eleven per cent maintained validated oversight frameworks for them. The scaling number and the oversight number were moving apart. A year on, at a size tier up, the same shape has reappeared.
What McKinsey Actually Measured
McKinsey's State of AI 2026 survey ran from early May to early June 2026. It drew 1,719 responses across ninety-seven countries, and published on 25 August 2026. Among organisations with more than a billion US dollars in annual revenue, forty per cent now report scaling AI agents, up from twenty-seven per cent the year before. Among organisations below that threshold, the figure did not move. Twenty-two per cent a year ago, twenty-two per cent now.
That flat twenty-two per cent is the finding. The entire thirteen-point movement sits inside a cohort that almost no New Zealand organisation belongs to. New Zealand's economy is overwhelmingly built from firms with annual revenue well under a billion dollars. A handful of listed companies and state-owned enterprises clear the threshold, and most of the readers of this series do not run one of them. The headline everyone quoted describes a world most New Zealand leaders are not standing in.
McKinsey's own report carries a second figure that belongs beside the first, read at a different resolution. Across the whole survey, the share of organisations reporting that AI is scaling across the enterprise rose from thirty-eight to forty-four per cent year on year, a six-point move. Over the same period, the share reporting any positive contribution to earnings before interest and tax from that scaling did not move at all. Thirty-seven per cent, in both years. The organisations McKinsey calls "AI high performers" held flat too, at roughly six per cent.
Put the two readings together and one argument appears twice, at two different resolutions. At the scale of the whole survey, deployment outran return: more organisations scaled, and the share actually converting that scale into earnings did not grow with it. At the scale of the revenue split, that same gap between deployment and return belonged almost entirely to the largest firms in the sample. The six points of enterprise-wide movement and the thirteen points inside the billion-dollar cohort are the same six points, counted two different ways.
The forty per cent figure is accurate, but it describes a concentration inside one cohort, not a universal trend. A reader who hears "forty per cent of organisations are scaling AI agents" and converts that into urgency for their own organisation has misread the survey. The correct response to a concentration finding is a specific question, not acceleration: which of the things that let the largest firms move does your organisation actually have?
The Measurement That Is a Different Kind of Evidence
Everything above is a survey. Organisations were asked what they are doing, and they answered. That measures intent and self-report, not what is actually running in an organisation's deployed systems.
Check Point's own September 2026 threat landscape, published 8 October 2026, measures something else entirely: an installed base, not a sentiment. Across the organisations it monitors, one in every thirty-nine enterprise prompts to a generative AI tool carries a high risk of sensitive data leakage. Eighty-nine per cent of organisations that regularly use generative AI tools are affected by at least one such exposure. This sits against a backdrop of 2,803 attacks per organisation per week, up forty-eight per cent on September 2025.
The categories of data most commonly exposed are not exotic. Network and IT infrastructure data turns up in risky prompts at seventy-one per cent of affected organisations. Financial data, seventy per cent. Legal and regulatory material, sixty-eight per cent. Employee and HR records, sixty-two per cent. Personally identifiable information, sixty per cent. These are the categories every organisation already has, sitting in the systems every organisation already runs, moving through a tool most staff now use without a second thought.
McKinsey and Check Point measured the same gap, from opposite sides. McKinsey measured who has scaled. Check Point measured how much of what has already been deployed sits outside any governed control, regardless of who deployed it. One is a survey of intent. The other is a measurement of exposure. Read together, they show that capability and deployment are diffusing fastest into the cohort best resourced to govern it. The same capability is also diffusing fastest into that cohort's own estate, ahead of the controls meant to hold it.
The Binding Constraint Was Never the Price
The cost of agentic tooling fell for organisations of every size over the same period. That rules out price as a big-company-advantage story. Smaller organisations did not stay at twenty-two per cent because the software got more expensive for them. They stayed flat while the price came down.
If price were the constraint, a falling price should have moved the smaller cohort too. It did not move at all. The constraint sits somewhere other than the invoice: implementation capacity. Four things separate an organisation that can scale an agent from one that can only pilot it.
The first is an evaluation harness: a repeatable way to test whether an agent does what it is supposed to do before it is trusted with real work. That test has to keep running as the underlying model changes under it. The second is identity management for non-human actors. Agents, service accounts and API credentials already outnumber human identities inside most enterprises by a wide and growing margin. Estimates range from twenty-five to fifty times the human count, depending on the estate measured. An organisation that has never built identity infrastructure for people it does not employ is starting that problem from zero. The third is workflow integration: the unglamorous work of rebuilding a business process around what an agent can actually do reliably. The alternative is bolting an agent onto a process designed for a person. The fourth is a named person who owns the risk when the agent is wrong, which is a governance decision, not a technical one.
A large organisation accumulates all four almost as a side effect of its own scale. A compliance function already runs evaluation-like processes for other systems. An identity team already manages machine credentials for infrastructure. A change-management capability already exists for enterprise software rollouts, and an executive layer already holds the authority to assign risk ownership. A smaller organisation has none of this as a going concern. It has to build all four from nothing, at the same time, to get the same result.
No New Zealand-specific survey, regulatory instrument or published dataset measures agentic adoption by organisation size, implementation capacity or non-human identity governance directly. No unrelated number can stand in for that gap.
The closest available New Zealand evidence is a structural analogue rather than a direct measurement. Deloitte New Zealand's Tech Trends 2026 reports that roughly ninety-three per cent of New Zealand AI spend goes to technology procurement, and roughly seven per cent to people, skills and workflow redesign. That measures where the money goes, not the pace of agent scaling. The report itself is about physical AI and what Deloitte calls the hybrid human and silicon workforce, not agentic adoption.
Even read with that caveat attached, the parallel carries real weight. The four-item capacity gap described above, evaluation harnesses, non-human identity management, workflow integration and risk ownership, is a people-and-workflow problem. A ninety-three to seven split toward procurement and away from exactly that side of the ledger illustrates why the implementation-capacity gap persists in a New Zealand setting. That holds even without a New Zealand-specific agent-scaling percentage to cite.
Four Questions Before You Accelerate
The honest response to a forty per cent headline is an audit against the four items that separate the organisations that moved from the ones that did not.
Start with the evaluation harness. Before an agent touches a real workflow, does your organisation have a repeatable way to test it, and will that test still mean something after the next model update changes the agent's behaviour underneath you? If the answer is a one-off pilot review rather than a standing process, you do not have this item yet.
Then identity. Does every agent, service account and credential your organisation runs have an owner, an access scope and an expiry, or does "the system" simply have standing access to everything it has ever needed? An inventory that cannot name who or what holds a given credential is a gap that is already open.
Then workflow. Has the business process actually been redesigned around what the agent reliably does? Or has the agent been bolted onto the process a person used to run, inheriting that process's assumptions about judgement, escalation and exception handling that the agent does not share? Bolting on gets a pilot working fast and breaks at scale.
Finally, ownership. When the agent is wrong, who is accountable, by name, and did that person agree to the role before the agent went live or only discover it afterwards? Naming a risk owner only after an incident proves one was needed is too late to count as ownership.
None of these four questions requires a billion dollars in revenue to answer honestly. They require time, attention and a willingness to find out that the answer is currently no.
The open-source dimension of this capacity question is about who already governs non-human identity in the open, rather than behind a vendor's dashboard. It is not about which model happens to be open-weight. SPIFFE, the Secure Production Identity Framework For Everyone, and its reference implementation SPIRE, are Cloud Native Computing Foundation graduated projects. Together they specify an open way to issue short-lived cryptographic identity to a workload, a service or an agent, rather than trusting a stored secret indefinitely. Adopting SPIFFE means adopting a community-governed standard an organisation can inspect, audit and run itself, rather than buying a vendor's proprietary answer to non-human identity management. That is a different capacity question from the one McKinsey measured, because it does not require a revenue threshold to install.
The implication on the sovereignty side is that defence procurement already uses implementation capacity to decide who gets access. The United States Department of Defense's Cybersecurity Maturity Model Certification programme requires a contractor to hold a certified level of cybersecurity maturity. That level is assessed against established federal standards appropriate to the sensitivity of the information involved, before the contractor may handle controlled unclassified information at all. The parallel is structural. A government buyer that already refuses to take a contractor's word for its own readiness is applying, formally and for years, the same test McKinsey's revenue split applies informally to enterprise AI agents. Capacity is verified before access is granted, not assumed because the invoice cleared.
Of the four items, an evaluation harness, non-human identity management, workflow integration and a named risk owner, which one would your organisation fail first if someone asked honestly today? And what happened the last time somebody actually asked?
This is one of seven weekly series in The Hamberger Report. 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 more than 30 years of professional experience in architecture, security, and technology leadership in New Zealand. I write as director of Te Pono Limited; the views are personal and do not represent the position of any client, any government agency, or the New Zealand government. My commentary on legislation and policy is analytical, drawing on publicly available sources and my professional expertise in architecture, security, and AI governance, 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: automated gates 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 are mine and are the same discipline I apply to the AI systems I audit for clients. AI accelerated the work; the thinking, and the responsibility for it, are mine.
[1] McKinsey & Company. "The State of AI in 2026: On the Road to ROI." 25 August 2026.
[2] Civic. "Big Companies Scaled Agents. Smaller Ones Didn't Get Cheaper Help." Field Notes, 2026. https://www.civic.com/field-notes/big-companies-scaled-agents
[3] Check Point Research. "September 2026 Cyber Threat Landscape: Global Attacks Jump 48% as Phishing and GenAI Data Exposure Rise." 8 October 2026. https://blog.checkpoint.com/research/september-2026-cyber-threat-landscape-global-attacks-jump-48-as-phishing-and-genai-data-exposure-rise
[4] Deloitte New Zealand. "Tech Trends 2026." 2026. https://www.deloitte.com/nz/en/Industries/technology/perspectives/tech-trends-2026.html
[5] United States Department of Defense, Office of the Chief Information Officer. "Cybersecurity Maturity Model Certification (CMMC)." https://dodcio.defense.gov/CMMC/
[6] General Services Administration. "Get to Know the Cybersecurity Maturity Model Certification." 12 February 2026. https://www.gsa.gov/blog/2026/02/12/get-to-know-the-cybersecurity-maturity-model-certification
[7] Cloud Native Computing Foundation. "SPIFFE: Secure Production Identity Framework For Everyone." https://spiffe.io/