The gains are real but fragile
New Zealand businesses have moved fast on AI. 82% now report using it in some form, up sharply over three years. The productivity case is genuine too. The Deloitte Access Economics Productivity Propelled report, commissioned by 2degrees and published in May 2026, found a 10-point increase in AI adoption correlates with a 1.5% lift in labour productivity.
The catch is in the fine print. That lift only materialises where AI is meaningfully embedded into operations. Shallow adoption, meaning AI features bolted onto existing tools or staff dabbling in public platforms, produces no measurable benefit. The report describes most NZ adoption as exactly that, still relatively shallow, fragmented, and with limited alignment to core strategy.
The macro picture confirms the gap. Treasury found no evidence of aggregate productivity growth at the national level in its August 2025 analysis, despite positive firm-level effects, with a 10% increase in AI investment associated with only a 0.04% bump in firm growth. The headline prize, a Microsoft estimate of $76 billion added to GDP by 2038 cited in MBIE’s July 2025 AI strategy, assumes deep, economy-wide integration that does not yet exist.
The cost the productivity metrics miss
Here is the part boards are not engaging with. A cluster of mid-2026 research is converging on an uncomfortable finding: the people getting the most out of AI are also the ones most at risk.
Psychology Today reported in June 2026 on workplace research, including an eight-month study at a tech company, finding employees using AI worked faster, took on broader scope, and were measurably more productive. They were also 88% more likely to be burned out and twice as likely to quit. Cognitive overload was significantly more common among professionals juggling more than three AI tools at once.
As Paula Davis, who reported the findings, put it: “When technology accelerates work, it also accelerates decision-making, information flow, client expectations, and the cognitive demands placed on employees.” A 1News piece in April 2026, drawing on BCG and Harvard Business Review research, described workers experiencing mental fog, headaches and slower decision-making. The logic is brutally simple: if staff are now managing 40 things instead of 20, and must validate all 40, their mental load has doubled even as output rises.
Cognitive surrender, and why it is invisible
The deeper risk is what AI does to judgment over time. Psychology Today described “cognitive surrender” in June 2026 as what happens when AI consistently handles the thinking. Because AI returns results in your own voice, shaped by your prompts, the cognitive loop feels closed. But there is a critical difference between approving a thought and generating one, and the technology is designed to make that difference invisible. The first capacity to erode is tolerance for not knowing, the discomfort that actually drives genuine judgment.
The loss, the research warns, announces itself only after it has already happened.
The reviewer bottleneck nobody priced in
There is a structural version of this problem too. RNZ described the inverted workload of the human in the AI loop: a generative tool can produce a plausible 50-page report in half an hour, but domain experts must spend hours reading and rewriting it. Creators now use less than 20% of the total effort, reviewers more than 80%. One described it as drinking from a firehose.
When those reviewers burn out and leave, they get replaced by more junior staff who trust AI output and sign it off faster, producing professional-looking content of uncertain quality. That compounds with a warning in TUANZ’s 2026 Digital Priorities Report: AI is automating the entry-level tasks that traditionally trained junior staff, removing the pathway through which expertise gets built. TUANZ chief executive Craig Young said in April 2026 that “AI is no longer a future concept. It’s operational today in many large businesses around Aotearoa.”
The board question
The accountability point is stark. As Newsroom noted in June 2026, AI tools are legally property, not persons. They cannot hold duties. Final accountability for every output falls to a human, one who is increasingly depleted, possibly junior, and under pressure to produce more, faster.
The Deloitte report is blunt that governance and risk maturity have struggled to keep pace, and that the firms succeeding are those investing in workforce capability and governance alongside the technology. The practical question for directors is whether they are measuring AI by outputs, speed, volume and cost, or also by what it is doing to the people responsible for quality. Most are doing the former. The cognitive debt accumulating from ignoring the latter will, like the research says, announce itself only once it has already done the damage.
Sources
- AI, productivity and the leadership challenge: Turning adoption into economic impact (2026-05-20)
- Official Information Act Response 20250500 – Future of AI (2025-08-13)
- AI-Accelerated Work Demands New Leadership Questions (2026-06-25)
- The AI irony: why is something designed to make work easier leading to burnout? (2026-04-20)
- AI and the Psychology of Cognitive Surrender (2026-06-01)
- Like drinking from a firehose – what it’s like to be the human in the AI loop
- TUANZ seeks bold leadership from government over AI worker training, national framework (2026-04-29)
- How the govt’s wager on AI could backfire (2026-06-24)