Every document uploaded, process automated and workflow handed to AI is helping build systems that could eventually outperform the very businesses feeding them. Productivity gains are real, he told iStart, but so are the unintended consequences, particularly for companies deploying AI without fully understanding where their data, knowledge and intellectual property end up.

“Everyone disregards the unintended consequences until it bites.”

A recent report found 40 percent of Australian office professionals have entered customer data or information into public AI tools, with 28 percent admitting inputting financial information or disclosing confidential company documents or strategies. Shadow AI Survey also noted that 70 percent of those surveyed have used AI tools or services at work despite not being allowed under company policies – and over half (53 percent) have faced formal consequences, such as a warning or disciplinary action, due to unauthorised AI use.

But Robb argues the issue isn’t simply whether staff are using AI. It’s how organisations are using it, what information is being shared with AI systems and whether anyone has thought through the consequences.

“Everyone disregards the unintended consequences until it bites,” he says.

Robb believes most employees are acting with good intentions. They’re looking for faster ways to complete tasks, improve reports and eliminate repetitive work. In many cases, they’re right: AI can dramatically improve productivity.

The problem is that productivity gains can obscure risk.

One of Robb’s biggest concerns is the amount of organisational knowledge being fed into public AI platforms. As staff upload documents, workflows and business information to improve outputs, organisations may be unintentionally giving away IP and expertise.

“[The foundation model providers] own the IP and you are training them to be better than you are,” he says.

Last month, Microsoft CEO Satya Nadella warned about the unintended consequences of AI, including the ‘reverse information paradox’ where enterprises unknowingly give away their proprietary institutional knowledge and competitive advantage simply by utilising third-party AI systems to boost productivity.

That’s just one unintended consequence Robb says companies need to consider as AI adoption accelerates. Another is the ease with which well-intentioned employees can create data exposure risks while trying to be more productive.

That, however, isn’t a new phenomenon. He recalls regularly dealing with incidents at a major bank where staff moved documents outside secure environments in order to work faster – sending a PDF to their home PC to convert it and send back to the work PC, for example. The intention was never malicious. Employees were simply trying to get the job done. But once sensitive information left the bank’s systems, a security breach had occurred and it was a summarily dismissible offence.

Robb sees similar behaviour emerging around AI tools today.

“Now, if I’ve loaded company data in to get that, that company’s got a genuine exposure,” he says.

For tech leaders, that means the challenge is no longer just identifying unauthorised AI use. It is understanding what information employees are putting into AI systems and whether appropriate guardrails exist around that behaviour.

Robb says regulated organisations, particularly financial institutions, have already invested heavily in governance and controls. “Banks have already stopped their people sending data to their iPads. It’s really hard to get data out of a bank or an insurance company.

“But when you’re a smaller business, the data’s everywhere.”

At the same time, companies are moving beyond AI experimentation.

Many of the early AI projects focused on incremental improvements – better reports, improved customer service or small efficiency gains. Robb says the next phase is different as businesses start to redesign entire processes around AI.

“It’s moved from incrementalism to step changes in processes.”

That shift creates new opportunities but also new risks. He cites the example of a plumbing business that has automated significant amounts of administrative work using AI. Compliance documentation and other repetitive tasks that once required multiple staff are now largely automated. The business gains efficiency, reduces effort and lowers its cost to serve customers.

The broader question, however, is what happens when competitors follow.

“If you have a lower cost to serve, you will kill your competition. Ai is a lower cost to serve.”

For business leaders, that’s a reminder that AI governance isn’t just about risk management. It is also becoming a competitive issue. Organisations that fail to adopt AI may find themselves at a disadvantage. Organisations that adopt it recklessly may create security, compliance or intellectual-property problems they don’t yet understand.

So what should companies be doing now?

Robb’s answer, unsurprisingly given he runs an AI governance consultancy, is governance. Not governance as a brake on innovation, but governance as a way of avoiding the ‘we’ll deal with that later’ approach that has accompanied many AI deployments. The question, he says, isn’t whether organisations should use AI. It’s whether they’ve thought through what happens when productivity gains collide with compliance obligations, security risks or unexpected business outcomes.

He says every organisation deploying AI into production should have a governance plan appropriate to its size and risk profile. For larger organisations, that means asking practical questions:

Does the AI initiative have board approval and a defined business outcome?

Is a human accountable for customer-impacting decisions?

If you lost control of your AI model, could you regenerate using using paper and pen, the data and the answers to get to the same position within 24 hours?

Does the business understand what data is being shared with AI tools?

An AI governance pack Robb’s company developed for banks and listed companies warns that AI governance is no longer optional and ‘we have a policy’ is no longer an answer – and ‘we are not doing anything yet’ is not a defence.

The pack outlines six guardrails:

• Strategic alignment – is AI use tied to risk appetite and corporate strategy),
• Accountability – who owns each AI system end-to-end?
• Explainability – can you explain a decision to a regulator or customer?
• Human oversight
• Monitoring and adaption – are models monitored for drift, bias and performance, and
• Organisational ethics

“If your board cannot score itself against these six, you do not yet have AI governance,” the pack says.

Smaller businesses? They’re exempt to some degree, he says. “If you’re say a five person plumbing business and you’re using AI to do training or something, honestly, congratulations, you’re curious, you’re using the tools. You don’t need a governance plan because you’re aware of the boundaries of what you’re doing.”

Putting in the wrong parts won’t be the end of the world – and it’s probably something you sometimes did pre-AI anyway.

Robb isn’t advocating caution at the expense of innovation. In fact, he believes many organisations across Australia and New Zealand have been too slow to realise AI’s potential.

Instead, his message is to stay curious, move quickly and prepare for the downsides, rather than ignoring them.

“AI is more valuable than it is detrimental,” he says. But organisations also need a plan for what happens when something goes wrong.

The post Unintended consequences: Are you training AI to eat your lunch? appeared first on iStart keeping business informed on technology.

The scathing ‘rapid review’ – commissioned by Public Service Commissioner Brian Roche after the Government Digital Delivery Agency (GDDA), which previously sat under the Department of Internal Affairs, was moved under the Public Service Commission – found widespread concerns about the way technology projects are governed, funded and managed across government. It concluded the system lacks visibility into the nature and value of technology investment, struggles to coordinate major projects and has a central digital function that is focused more on administration than strategic leadership.

“This review gives us a clear picture of what’s working and what’s not.”

The findings are unusually blunt for a government review.

Among the more damning observations were comments from agency leaders who said the GDDA often failed to provide the expertise or leadership needed to deliver projects.

One agency leader quoted in the report says the central digital function had simply helped them ‘admire the problem’ without making project delivery any easier. Another says the organisations focused on process, rather than outcomes, while others questioned its ability to broker common solutions across agencies.

The review, which was led by former Auckland Airport chief executive Adrian Littlewood, Justin Gray (former managing director of Datacom) and transformation advisor and former Kāinga Ora CEO Matt Crockett, found the problems ran deeper than any single agency. It identified a fragmented system, lacking complete and timely information and missing the urgency to drive change, despite investing $42 million a year in the central digital function. The system was characterised by fragmented accountability, duplicated investment, weak prioritisation and poor information. Agencies were found to be pursuing similar projects independently, competing for scarce technology resources and in some cases building systems that could have been shared across government.

Examples cited in the review include multiple agencies pursuing digital identification initiatives with limited coordination, an agency developing its own payments system rather than reusing existing and mature payment systems operated by another agency, and two related organisations – in the same building – tendering for new systems independently before wider integration issues were considered.

The report notes these issues have emerged despite the existence of a central digital foundation intended to identify common opportunities and drive all-of-government outcomes.

Perhaps most concerning is the review’s conclusion that government lacks reliable information about its own technology investments.

“It was hard to get fundamental metrics to judge performance of tech investment across the agencies and there was limited understanding on baseline tech spend funded through appropriations vs project spending,” the report notes.  Existing reporting was described as largely focused on expenditure and activity, rather than business outcomes and public value. Limited central visibility into agency spending – with ‘skunk-works’ projects hidden in budgets and financial reporting, is also creating accountability problems, with some tech initiatives effectively hidden within operational budgets.

The review says this lack of information makes it harder to prioritise projects, assess risk and direct investment to the areas likely to generate the greatest benefit.

While the report focuses on public service IT, many of the problems it highlights will sound familiar to enterprise leaders. Large organisations frequently face similar challenges around technology sprawl, competing priorities, duplicated systems and limited visibility of investment outcomes. The challenge of balancing business unit autonomy against enterprise-wide standards is hardly unique to the public sector.

But the review highlights how much more difficult the issues become when multiple agencies, funding streams and ministerial priorities are involved.

It’s also highly critical of the current funding and approval process, saying the existing model is designed for large capital infrastructure projects rather than the fast-moving, iterative nature of modern digital delivery. Agencies often struggle to obtain funding for smaller technology initiatives, leading some to seek approval for larger projects than necessary, hide smaller discretionary projects within IT operating budgets, delay sensible upgrades or allow tech debt to grow.

The review argues current settings can actively discourage investment and innovation.

It notes agencies face lengthy approval processes while governance structures often lack the authority to make meaningful decisions or resolve trade-offs across government. Interviewees described a culture where avoiding mistakes is rewarded more than delivering value and where activity can become a substitute for outcomes.

The review also notes that engagement with tech providers is inconsistent and does not support long-term partnerships or co-investment.

Resetting for digital delivery

Roache says the review provides a solid base for a reset as part of the work to transform the public sector.

“This review gives us a clear picture of what’s working and what’s not,” he says.

To address the issues the review proposes a three-part reset.

The first focuses on prioritisation, including an assessment – led by GDDA with support from Treasury and key agencies – of digital projects across government and the reallocation of resources towards higher-value initiatives. Low-value or conflicting projects could be paused or retired, with investment concentrated on programs deemed most likely to deliver impact. Identifying ways to ‘self-fund’ new investment by reprioritising existing funding and making short-term operating and process changes that deliver ahead of new technology is also a key output identified.

The second focused on repositioning the GDDA itself. Rather than acting as a project delivery organisation, the agency would become a smaller, but more capable, strategic authority responsible for standards, assurance, architecture, procurement and system-wide capability. Delivery responsibility would sit with lead agencies, while the GDDA would be expected to provide expertise, direction and accountability, retaining lead on strategic procurement via a panel model for common/functional services such as cloud.

The third track targets structural reform across government, including changes to funding models, governance arrangements, accountability frameworks and reporting requirements. Legislative changes could also be considered to provide clear boundaries or support for critical enabling tech projects, such as ‘bundled consent’ options for citizen data.

The review also recommends more consistent measurement of technology costs, performance and benefits across agencies.

Underlying the recommendations is a belief that technology should be treated less as an IT function and more as a driver of productivity and service transformation.

The report cites Inland Revenue’s business transformation as evidence that major gains are possible when governance, funding and delivery settings are aligned. It also highlights potentially significant productivity opportunities in areas such as health, where technology-enabled improvements could generate substantial efficiencies if barriers to delivery were removed.

The message from the review is straightforward: New Zealand does not necessarily have a technology spending problem. It has a prioritisation, governance and execution problem.

The government’s proposed digital reset aims to fix that. Whether it succeeds may depend on whether the public sector can do something the report suggests it has struggled with for years: Stop talking about transformation and start delivering it.

Roche says he will consider the findings of the review before making decisions on ‘a way forward’.

The post NZ digital govt reset after damning review appeared first on iStart keeping business informed on technology.

The initiative marks the global financial messaging network’s first production use of a blockchain-based ledger and is being positioned as a way to enable funds to move overnight and on weekends while remaining within the regulated banking system – and compete with the emerging stablecoin market.

“We see strong potential to help customers move funds in real-time and manage liquidity more flexibly.”

Banks from six continents are taking part, including ANZ, Citi, HSBC, Lloyds Banking Group, UBS, Wells Fargo and Standard Chartered. Swift (the Society for Worldwide Interbank Financial Telecommunications) says participating banks will use the shared ledger as an orchestration layer for bank-issued tokenised deposits on their own infrastructure, allowing institutions to move value before final settlement is completed through existing banking systems.

For ANZ, the trial expands a digital assets strategy the bank has been pursuing for several years, including work with tokenised assets and blockchain-based settlement technologies.

Lisa Vasic, ANZ managing director transaction banking, says the bank is working with Swift and global partners to ‘securely scale next-generation payments infrastructure’ and deliver ‘more efficient, always-on payment capabilities’.

“By combining Swift’s trusted network with this new infrastructure, we see strong potential to help customers move funds in real-time and manage liquidity more flexibly,” Vasic says.

Beyond crypto and speculation

Blockchain has struggled to find a use case within business. For much of the past decade, the technology, initially hyped as the foundational technology for Bitcoin and decentralised digital currency, has been associated more with cryptocurrencies, speculative trading and fraud than practical enterprise applications. Despite billions of dollars invested globally, and plenty of attempts, relatively few large-scale business deployments have emerged outside digital asset markets.

The distributed ledgers have been trialled for everything from smart contracts that automatically execute when predetermined conditions are met, to supply chain tracking systems designed to create tamper-resistant records of product governance and movement. But while some have delivered niche benefits, few have achieved widespread adoption at scale – and there have been several high profile failures. Maersk and IBM shuttered the TradeLens blockchain-enabled global trade platform in 2022 after five years work, blaming an apparent lack of interest and trust from the industry. The ASX’s blockchain-powered Chess settlement and clearing system project meanwhile went down in a ball of flames – labelled a ‘profound failure’ by the chair of a parliamentary joint committee on corporations and financial services, and leading to an AU$250 million writedown.

The payments difference

The Swift initiative sees some of the world’s largest banks using the technology to tackle an issue that has frustrated businesses for decades: The limitations of cross-border payments.

International transfers remain constrained by time zones, banking hours and fragmented settlement processes. While businesses have become increasingly digital and globally connected, moving money internationally often remains slower than moving information.

Swift believes tokenised deposits operating over a blockchain-based ledger could help bridge that gap. The organisations says the new capability would allow banks to support 24/7 cross-border payments while maintaining the compliance, risk management and control frameworks already embedded into the global financial system.

“It allows tokenised value to move across borders with the velocity and flexibility modern commerce expects, while maintaining the same high levels of resiliency, security and compliance global finance requires,” says Thierry Chilosi, Swift chief business officer.

The approach differs significantly from public cryptocurrency networks.

Rather than creating an alternative financial system, Swift’s ledger is designed to sit alongside existing banking infrastructure. Participating banks continue issuing and managing deposits, while the ledger provides a common mechanism for coordinating transactions between institutions. Final settlement still occurs through established banking channels.

That distinction may prove critical.

Many organisations have shown little appetite for holding cryptocurrencies or exposing core treasury processes to public blockchain networks. The involvement of the 17 banks suggests growing interest from major financial institutions in tokenised deposits as a way to improve payment availability and liquidity while remaining within established regulatory and banking frameworks.

Swift says banks will be able to move tokenised funds around the clock, including outside normal business hours, helping businesses access capital more quickly and manage liquidity more efficiently.

Testing blockchain’s real value

The launch also appears to reflect a broader shift taking place within the banking sector, with many major financial institutions increasingly focused on determining where exactly blockchain can deliver measurable value. Cross-border payments have emerged as one of the strongest candidates because they involve multiple organisations, multiple jurisdictions and significant coordination requirements – all characteristics that align naturally with distributed ledger designs.

In effect, the trial represents blockchain being used for what it was arguably always designed to do: Maintain a trusted and shared record between parties that do not directly control the same systems.

Swift says the shared ledger, first announced last year, was designed and built with feedback from international financial institutions.

“The strong support from banks shows the practical value of this approach,” Chilosi says, adding that it will help scale benefits globally while creating a foundation for future innovation in areas like programmable money and agentic commerce, where automated systems execute transactions and payments on behalf of users.

The scale of the network involved also gives the initiative significance. Belgium-based Swift underpins the vast majority of international bank messaging and says its infrastructure connects more than 200 markets worldwide.

While enterprise blockchain initiatives have often generated considerable hype before falling short of expectations, Swift’s initiative targets a clearly defined operational problem and is backed by institutions already responsible for moving trillions of dollars through the global financial system.

For ANZ and its banking peers, that may ultimately prove blockchain’s most natural home: Not replacing banks, but helping them move money faster.

The post ANZ joins Swift blockchain payments push appeared first on iStart keeping business informed on technology.

The prediction reflects one of the less discussed consequences of the generative AI boom. While early conversations focused on productivity gains, code generation speeds and developer efficiency, software engineering leaders (and indeed, all IT leaders) are now facing a new challenge: Controlling the cost of AI itself.

“Token discipline will not emerge through developer choice alone.”

At the same time, vendors are increasingly moving away from traditional seat-based licensing models towards consumption-based pricing, where organisations pay for the tokens consumed by AI models. As AI deployments scale, token consumption is emerging as a growing cost consideration across the industry. The combination of higher usage and metered billing is creating a cost trajectory that Gartner believes many organisations are underestimating.

On the coding front, Gartner says 40 percent of software engineering leaders are reporting that more than half of their teams are using AI tools to augment software development processes. As adoption grows, token consumption – the units increasingly used to measure and bill AI model usage – is rising rapidly.

Early signs of the cost impact are already emerging with Gartner Peer Insights data showing 23 percent of software engineering leaders are reporting costs of US$200 to US$500 per developer per month in token costs for AI coding agents such as Claude Code, Cursor and OpenAI Codex. For five percent of organisations, that figure has already soared to more than US$2,000 per developer per month.

“Organisations are rapidly moving from experimentation to scaled deployment of AI coding agents, but many are underestimating the financial impact of rising token consumption,” Nitish Tyagi, Gartner senior principal advisor, says.

““AI coding costs will continue to rise as infrastructure investment and profitability challenges push model pricing higher,” Tyagi says. “At the same time, as more developers adopt AI tools, light users are expected to rapidly become mainstream users as familiarity and reliance increase, driving further growth in token consumption and overall spend.”

The concern isn’t simply the price of large language models. Gartner argues that three common behaviours are driving much of the overspending: Giving AI agents too much autonomy, providing excessive amounts of context, and failing to establish feedback mechanisms to identify and prevent wasteful usage patterns.

“Token discipline will not emerge through developer choice alone,” Tyagi says, noting that developers will naturally optimise for speed and convenience, rather than token efficiency. Instead, organisations should treat token consumption as a governed part of their engineering operating model, the analyst company says in its How to Optimise Token Consumption for AI Coding Agents report.

That recommendation may be particularly relevant as organisations race to adopt increasingly capable AI coding agents. Unlike traditional coding assistants that suggest code snippets or provide developer support, agentic platforms can take on broader development tasks autonomously. Gartner points to increases of 100x over coding assistants experienced by software engineering leaders.  

Gartner says organisations need to establish clear rules around when AI agents should be used and how much autonomy they should be granted.

The report recommends categorising software development activities into three tiers: Developer-only work for sensitive, operationally risky work or work not worth the token cost; AI-assisted development where humans remain in control, and fully autonomous agent-led development. AI-assisted development should be the default zone for most day-to-day engineering because it improves throughput while maintaining human oversight and avoiding some of the introducing costly agentic loops which can occur when autonomous systems repeatedly call models and tools without adequate controls, Gartner says. Those with lower software engineering maturity and limited governance frameworks are also advised to prioritise assistive AI over autonomous development.

Governance is another area Gartner says requires urgent attention.

Less than one in three engineering organisations currently has formal AI governance policies in place. Gartner says engineering leaders should assess both the maturity of their software development practices and existing AI controls before expanding the use of autonomous coding agents.

“Organisations should introduce mechanisms such as token thresholds, escalation policies and automated monitoring to manage growth. Embedding these controls into engineering workflows ensures consistency and prevents uncontrolled costs.”

The report also highlights what Gartner calls context engineering as a critical discipline for controlling costs.

Every file, document, code repository extract or tool output included in an AI prompt increases token consumption. Gartner advises development teams to treat context as both a quality input and a cost driver, requiring developers to evaluate what information is essential, what can be summarised and what can be removed altogether.

“A mature context engineering practice should teach developers how to structure context so that it is not merely comprehensive, but optimised.”

Another recommendation is to avoid defaulting to premium AI models for every task. Instead, Gartner advocates model-routing approaches where smaller, lower-cost models handle routine work, with tasks escalated to more powerful and expensive models only when complexity or risk justifies the additional spending.

“AI coding agents are most cost‑effective when work is broken into smaller tasks that can be handled by smaller models, with escalation only when complexity demands it. Model choice should be driven by task complexity, context size, application criticality, and developer maturity.”

Also recommended is better visibility into how AI coding tools are being consumed, with regular reviews of high-token-consuming workflows as part of sprint retrospectives to identify inefficiencies, refine practices and promote knowledge sharing across engineering teams.

“Most organisations still lack the maturity and frameworks to effectively measure cost versus business impact,” says Tyagi.

The report stresses that higher costs aren’t necessarily bad if they bring measurable business benefits, such as profit, reduced risk, improved growth or support strategic goals.

“The goal is to use AI coding agents efficiently but not to restrict the use of them.”

The post AI coding costs poised to overtake developer pay appeared first on iStart keeping business informed on technology.

The government-backed investment agency is promoting New Zealand as a destination for large-scale data centres and AI infrastructure, arguing the country offers a rare combination of renewable electricity, available land, grid capacity and political stability at a time when global demand for computing power is surging.

“The question is whether global investors will be persuaded that renewable power outweighs and room to expand outweigh the advantages traditionally offered by larger markets.”

The strategy is built around a simple proposition. As AI drives unprecedented demand for data centres, access to power is becoming a critical constraint. And if electricity is becoming the scarce resource, New Zealand believes it has an advantage. Earlier this year Boston Consulting Group flagged data centres as a key growth opportunity for New Zealand, proviing a $70 billion opportunity for the country.

Invest NZ describes New Zealand as ‘the secure, green solution’, highlighting that more than 88 percent of the country’s electricity generation comes from renewable sources. It says established data centre hubs are increasingly facing limits around power, land and environmental capacity, while New Zealand still has room to grow.

The pitch, launched last week, comes as global spending on Ai infrastructure reaches unprecedented levels. Microsoft, Amazon, Alphabet, Meta and Oracle collectively spent US$437 billion on capital expenditure in 2025, up 68 percent year-on-year, with spending forecast to continue rising as hyperscalers build out AI capacity.

Boston Consulting Group, which has been advising on New Zealand’s data centre opportunity, says global computing capacity is expected to more than double by 2030, with demand continuing to outstrip supply. Its research argues New Zealand has a ‘right to win’ a larger share of international investment because of its renewable energy resources, fibre connectivity and stable operating environment.

But New Zealand is entering a competitive market.

Australia has been attracting large-scale investments from AWS, Microsoft and Google for years, supported by a significantly larger domestic market and a more established cloud footprint. Those investments have largely been concentrated around Sydney and Melbourne, where providers can serve large enterprise and government customer bases.

Invest NZ is effectively arguing that, in the AI era, the equation is changing. Rather than being built solely where customers are located, some future AI infrastructure may be built where long-term electricity supply is available and affordable. It’s an argument BCG also supports. It says a growing share of workloads, particularly AI model training, are not latency-sensitive, so distance is no longer a constraint and computing capacity can increasingly be located where power and infrastructure are most competitive.

56 and counting

The country isn’t starting from scratch.

Last year a Tech New Zealand report into data centres noted there were already 56 operational data centres in the country, with another 20 planned or under construction, claiming the sector ‘directly employs over 1,000 people’. That report also painted a glowing picture of data centres in New Zealand, saying the sector underpins $93b in economic activity, with $16.5b in ICT GDP and $76.5 billion in knowledge-intensive services. And, like BCG and Invest NZ, it also pushed the sustainable power card, noting local data centres achieve on average power usage effectiveness. (PUE) of 1.3 – ‘significantly’ outperforming the global average of 1.54, and are rapidly transitioning to 100 percent renewable energy sources.

Recent years have also seen major investments from local operators including Spark, Datacom and CDC, alongside expansion by international cloud providers. Amazon Web Services has previously announced plans to invest NZ$7.5 billion in its Auckland cloud region over 15 years.

But the clearest example of the proposition Invest NZ is now selling offshore may be Datagrid’s Southland project. A subsidiary of Singapore’s BW Digital, and headed up by founder and CEO Rémi Galasso who was also the founder of the Hawaiki Cable (Callplus founder Malcolm Dick is Datagrid’s other co-founder), the company has secured approvals for a $5 billion ‘AI factory’ at Makarewa, north of Invercargill. The 78,000m² 

development will comprise six large data halls and is designed to support hyperscale AI and cloud workloads.

Galasso has described the development as having the potential to transform Invercargill into a digital destination, while the company has consistently promoted Southland’s renewable energy resources, cool climate and international connectivity as key advantages.

But the project also highlights the scale of infrastructure required to support AI ambitions. When fully operational, the facility is expected to draw up to 280MW. It’s been granted approval however, to scale up to an eye-watering 1GW capacity – twice the power consumption of the Tiwai aluminium smelter, whose future has been a key factor in New Zealand’s energy planning. Even at 280MW, the facility will be the second-largest electricity user after Tiwai and accounting for six percent of national electricity consumption. (Globally, a UN University report – which notably doesn’t include New Zealand suggesting the country’s current limited visibility in global AI infrastructure development – says by 2030, AI could be consuming three percent of the world’s electricity.

Earlier this year Datagrid secured a 15-year 140MW/year power purchase agreement with Mercury to help underpin future supply requirements.

The infrastructure challenge extends beyond electricity. Hyperscale AI facilities require extensive cooling systems, and Datagrid plans to use groundwater and on-site stormwater for thermal management at its Southland facility. As New Zealand pursues large-scale AI investment, energy, water and connectivity are all becoming part of the equation.

The company is also expected to benefit from the Tasman Ring Network, a 6,000km trans-Tasman subsea cable linking Invercargill with Sydney and Melbourne that has received approval and is intended to improve international connectivity for cloud and AI workloads.

Supporters see such projects as evidence New Zealand can compete internationally.

University of Waikato AI Institute director Albert Bifet has previously said facilities such as Datagrid provide the computing power and connectivity needed to train and run modern AI systems, noting that other regions are investing heavily in comparable infrastructure to support innovation and economic growth.

Others point to broader economic questions around who ultimately benefits from such investments.

University of Auckland researcher Angus Dowell has noted that major data centres can be physically located in one region while remaining economically integrated into global cloud and AI networks, where customers, decision-making and commercial returns are often concentrated elsewhere. “The key issue is therefore not just whether the project brings investment, but how the benefits of that investment are structured and distributed,” he said earlier this year.

For now, Invest NZ is focused on making the investment case.

At an Auckland event last week, business and energy leaders were presented with a vision for turning New Zealand into a significant destination for AI infrastructure investment. Contact Energy chief executive Mike Fuge described the opportunity as a ‘once-in-three-generations opportunity’.

The question is whether global investors will be persuaded that renewable power, available capacity and room to expand outweigh the advantages traditionally offered by larger markets.

With hundreds of billions of dollars continuing to flow into AI infrastructure globally, that is the bet New Zealand is now taking to market.

The post NZ’s $30b dollar AI data centre dream appeared first on iStart keeping business informed on technology.

“The program is the first major step in an ambitious program designed to strengthen connection between education and industry.”

The Industry Accelerator – Teachers to Industry Program will see educators undertake professional development and industry immersion with organisations including Adobe, Macquarie Technology Group and Omron Automation and Robotics Australia. It’s designed to give teachers hands on experience and direct exposure to current technologies, workplace practices and employer expectations, before returning those insights to the classroom.

The launch follows a successful pilot involving 34 advanced manufacturing teachers, which drove the program to progress to full implementation last week with an initial cohort of 50 digital and finance teachers beginning their initial professional development phase before beginning industry placements.

TAFE NSW says participating teachers will focus on priority skills areas and then share learnings across the wider organisation to broaden the program’s impact.

The move comes amid ongoing concern across Australia and New Zealand about the availability of digital skills.

In New Zealand, technology industry body TUANZ recently issued another call for ‘urgent, coordinated investment in home-grown digital capability’, including encouraging diverse pathways of training, expanding industry-led reskilling programs with tax credits or subsidies, and mandating technology upskilling in the core curriculum.

While employers are continuing to invest in AI, cybersecurity, automation and data-driven technologies, industry groups on both sides of the Tasman have continued to warn that demand for digital capability is outstripping supply. At the same time, employers on both sides of the Tasman have long complained that graduates often arrive with qualifications but lack exposure to the tools, platforms and technologies being deployed in business environments, with the technical skills learned in education settings often not matching those required in the real world.

In announcing the initiative, TAFE NSW said the program would help educators remain at the forefront of industry practice by providing hands-on experience with ‘leading employers across NSW like Adobe, Macquarie Technology Group and Omron’. The aim is to strengthen expertise in priority skills areas and improve the relevance of training delivered to students, ensuring it better reflects workplace expectations.

The organisation says industry partnerships will play a critical role in ensuring training reflects current workplace requirements and future workforce needs, and representatives from Microsoft, Google, SAS, Adobe, Macquarie Technology Group, Palo Alto Networks and Omron on hand for last week’s launch at the Institute of Applied Technology, Digital at TAFE NSW Meadowbank.

There is, of course, a difference between learning how modern organisations use technology and learning how tech companies would like organisations to use technology. TAFE NSW has framed the program around industry practice and workplace readiness, but it remains to be seen whether educators return with deeper insight into employer needs, vendor ecosystems or a healthy dose of both.

Fiona Watts, TAFE NSW senior manager education standards, says the program is the first major step in an ambitious program designed to strengthen the connection between education and industry while providing direct exposure to emerging technologies and workforce requirements.

Watts, who is leading the initiative, says the program is about creating genuine partnerships to help educators understand where industry is heading so they can prepare learners for careers that may not even exist today.

The new scheme forms part of a broader push by TAFE NSW to strengthen tech capabilities across its training network. Other recent initiatives include the launch of microskills programs covering AI, cybersecurity and data centres, and specialists training facilities and centres of excellence focused on emerging technologies. The Meadowbank facility, into which the Albanese and Minns Labour Governments have invested $11 million, is one on those new centres of excellence and is expected to train more than 50,000 Australians annually, upskilling and reskilling students and workers in cybersecurity, AI, big data, cloud and software.

The launch also follows the New South Wales Government’s announcement of a A$2.7 billion investment in TAFE NSW, including funding for network upgrades, classroom technology modernisation and new devices for staff and students.

The post TAFE NSW sends teachers back to industry to tackle tech skills gap appeared first on iStart keeping business informed on technology.

One Gartner client began with more than 600 potential AI use cases competing for investment. Unsurprisingly, Gartner’s Luke Ellery recommends companies don’t pursue everything. Instead, he recommends creating a ‘AI funding funnel’ that ranks initiatives based on business value and technical feasibility before significant budgets are committed.

“It’s a great way of helping CFOs understand AI investments.”

For CIOs facing pressure to ‘do more with AI’ the message at Gartner’s recent Data and Analytics conference in Sydney, the message was surprisingly simple: Stop launching pilots and start managing AI investments like a portfolio.

Ellery, Sydney-based Gartner VP analyst, says while organisations have been spending billions on AI, ‘you could say there is no R in ROI, only investment’. A Gartner survey showed 80-95 percent of Gartner clients saw limited financial returns from their AI investments and only 11 percent of CFOs were able to concretely measure financial returns from big AI investments.

Ellery’s assessment? Based on other Gartner surveys, he says the number one thing is picking the wrong use cases. Deploying Microsoft Copilot might help employees save time and feel happier and more productive, but that’s not financial value.

Value vs feasibility

Gartner’s funding funnel requires ideas to be vetted and prioritised across two dimensions – business value and feasibility.

Business value focuses on the outcomes executives care about – higher revenue, lower costs, reduced risk or improved service delivery. It’s an area that will vary for each organisation, Ellery notes. Time horizon is also an important factor here, he says. While last year CFOs were very focused on getting AI returns within the year, that stance has softened somewhat with the realisation achieving AI goals, and returns, will be harder than anticipated and returns will come on a longer time frame.

Feasibility, meanwhile, looks at the practical realities of implementation, including data readiness, technology requirements, skills availability, security controls and governance obligations.

The sweet spot, of course, is funding initiatives that score highly on both measures and will provide easy wins.

But Ellery notes: “Often organisations won’t have many.” All, however, is not lost. Ellery says leaders need to think about how projects work together, evaluating AI initiatives as a portfolio, rather than in isolation.

“In some scenarios it is very hard to get any of the individual projects off the ground by themselves, but if we collect a number of projects together, we can combine the cost and value and risks.

“It’s a great way of helping CFOs understand investments so we can actually create those larger returns.”

Why boring projects matter

He used customer personalisation as an example of an initiative with potentially high business value. The ability to tailor products, offers and interactions to individual customers has long been viewed as an AI holy grail, with obvious implications for revenue growth and customer retention.

The challenge, however, is that many organisations lack the foundations required to make personalisation work. Customer data may be fragmented across multiple systems. Governance controls may be immature. Data quality may be poor. Technology platforms may not be integrated.

As a result, what appears to be a high-value opportunity often scores poorly on feasibility.

By contrast, a much less glamorous project – such as automated client call summarisation – may deliver only modest business value on its own. It is unlikely to transform an organisation’s growth prospects or dramatically improve profitability, yet it may be significantly easier to implement.

More importantly, the project may capture and structure customer information essential for more advanced personalisation initiatives later.

Seen that way, the lower-value project becomes a stepping stone to something much bigger.

“We need to look at all these initiatives as a group in a broader view,” Ellery says. “The problem is that a lack of consistency prohibits us from actually achieving differentiation,” he says, arguing that organisations often need to fix underlying operational and data problems before they can use AI to create competitive advantage.

Building foundations before differentiation

Many AI projects fail not because the models are inadequate, but because organisations do not have accessible, well-governed data available for those models to use.

Technical debt presents a similar challenge. Legacy systems often store valuable information in formats that are difficult to access, integrate or analyse. Organisations may also need to invest in security, governance and compliance frameworks before deploying AI more broadly.

These investments can be difficult to sell with boards and executive teams naturally attracted to customer-facing use cases and ambitious transformation projects. Few get excited about data architecture, governance frameworks or system modernisation. Yet those investments can determine whether more ambitious AI initiatives succeed or stall.

Ellery pointed to the banking sector to illustrate that, noting years of core banking modernisation efforts as an example of organisations improving consistency across technology environments.

“That’s probably why some of the banks are more ahead in the AI race than others, in terms of realising the value of AI.”

In some cases capability-building projects lacking immediate business impact may be the highest-value investments an organisation can make.

Making the financial case

After vetting and prioritising – providing the litmus test for funding – Gartner’s funding funnel focuses on the crux of the issue – the funding. He noted there are different types of value for organisations. ‘Green money’ provides direct, measurable financial outcomes that appear in profit and loss statements, such as increased revenue or reduced budgeted costs. These are the benefits CFOs tend to care most about because the cause-and-effect relationship is clear.

‘Blue money’ by contrast, includes measures such as customer satisfaction, employee experience and Net Promoter Score. While these metrics may contribute to future financial performance, they don’t guarantee a return and can be difficult to directly link to revenue or profit with less direct correlation.

Ellery also warned against relying heavily on ‘cost avoidance’ arguments, such as delaying a system upgrade, noting that financial leaders often place less weight on savings that have never formally appeared in a budget.

“What we’re doing is putting all this information together on a slide so we can make a determination with the CFO.”

When cumulative costs are higher than the benefit – as with the earlier customer personalisation example – there’s a chance to flip the conversation and argue it as a cost leader to enable you to do other projects – or to bring it into a wider program.

Ellery noted the success of a multinational DIY and home improvement goods company. “They used a system dynamics model and did scenario simulations over three and five years to actually figure out how to order, how to do promotions with feedback loops.”

Targeting 10 percent revenue growth, they nearly got it, coming in at nine percent.

“But they also stabilised their supply chain, which unleashed about $120 million in working capital, and they also deferred the expansion of their processing facilities, which saved them about $25 million in capital costs.”

A courier company increased on-time deliveries by 25 percent, cut idle time by 20 percent and reduced fleet expansion to three percent through use of ‘a very interesting agent-based model, where they had agents represented for vehicles, drivers, customers as autonomous agents’ and micro-hubs.

And a hospital, aiming to reduce emergency department wait times by 15 percent in three, months used AI in its demand forecasting and rostering. Their wait times dropped 77 percent and – as a side benefit – they also saved $100,000 in overtime through data-driven staffing.

“If we can target the value that we’re trying to optimise, it makes it so much easier for us to build these AI systems to actually realise those outcomes and to optimise them.”

The post The AI prioritisation problem appeared first on iStart keeping business informed on technology.

In its first determination on tracking pixels, the Office of the Australian Information Commissioner (OAIC) has found that two health providers – Medmate Australia and Monash IVF – interfered with user privacy by using the technology to track website visitors and target them with advertising on social media platforms.

“Advanced technology used for tracking and targeted advertising in the online realm still has to be used in compliance with the Privacy Act.”

Privacy Commissioner Carly Kind’s decision establishes that the use of tracking pixels to track website visitors to health-related websites and to subsequently target them with advertising on social media platforms, amounts to a collection of sensitive information for which website providers must obtain users’ consent.

Tracking pixels are small pieces of code provided by third parties – typically platforms like Meta (with Meta Pixel) or TikTok (with its TikTok Pixel, integrated into TikTok Ads Manager, and used either natively or via TikTok partners including Shopify). They’re embedded into webistes to collect data about user activity. When a visitor loads a page, the pixel transmits information about that interaction back to the platform, enabling functions such as campaign measurement, audience profiling and ad targeting. Kind notes tracking pixels can be configured for a variety of purposes, tracking webpages users visit, clicks, what is put into carts and, in some cases, information entered into forms. Unlike cookies, they can’t easily be cleared or blocked entirely.

While such tracking is widely used across sectors, the regulator’s determinations focused on health-related websites, where browsing behaviour can reveal – or allow inferences about – an individual’s health status. This is classified as sensitive information under Australia’s Privacy Act.

The Commissioner found that using pixels in this context to track visitors and retarget them with ads amounted to collecting sensitive information without consent, in breach of privacy obligations.

“Today’s decision establishes that the advanced technology used for tracking and targeted advertising in the online realm still has to be used in compliance with the Privacy Act,” Kind says.

The ruling extends beyond just health, with Kind noting that website providers using tracking pixels to collect any sensitive information, such as data on political opinions, race or ethnicity, as well as health, must obtain consent.

Widespread use, limited disclosure

Alongside the determinations, the OAIC released findings into a ‘scan’ of 50 health service provider websites, highlighting how common the use of tracking pixels has become – and how poorly it is disclosed.

The regulator found that more than half of the sites examined used a third-party tracking pixel, yet 77 percent of those did not mention the technology in their privacy policies.

Overall, 96 percent of the websites used some form of tracking technology.

The OAIC says the results point to a broader issue across the digital ecosystem, where tracking technologies are embedded into websites but remain largely invisible to users.

A deeper dive into 12 websites found 50 percent used more than one tracking pixel provided by social media platforms, with all using Meta’s pixel.

Among the information being sent to the social media platforms was the full URL containing information about the page visited, website searches, button clicks, and device information.

Instances were form fields, including hashed name, address and phone numbers were shared with social media platforms were also observed, enabling matching of information to individuals and their profile, even when they are not logged in.

Despite that, the majority did not disclose use of the pixels or that information was being shared with social media platforms.

In the case of Medmate, a telehealth provider, the company was using both Meta and TikTok pixels and fed full URLs to TikTok which revealed information on health conditions, including urinary tract infections, bacterial vaginosis and the need for emergency contraception, and medication sought by individuals.

The company, which stopped using all tracking pixels on its site as of December 2025, paid for online advertising campaigns, primarily on Facebook, Instagram and TikTok, using custom audience lists to retarget individuals based on their website interactions.

The company had Meta Pixel’s Advanced Matching feature enabled from October 2021 to December 2024, enabling collection of more granular information about individuals.

“For periods in which the Advanced Matching feature was not in use, Medmate could retarget ads to those individuals who visited the website on the pixel provider platforms,” Kind says. “I am therefore satisfied that Medmate was able to distinguish individuals from others in a way that affected their rights or interests by using tracking pixels to target and treat individuals on an individualised basis, even without their identity being known.

Monash IVF meanwhile used five active tracking pixels, including Meta, Google Ads and google Analytics in December 2024. It also used two others, including Pinterest, at different times.

It used Meta Pixel for advertising campaigns relating to egg donor and freezing, endometriosis, fertility and sperm donation, among others and created custom audience lists using other sources of customer information, uploading the lists to Meta Pixel Provider Dashboard to retarget advertising and ‘layer, build and further refine individuals’ they wanted to retarget.

Broader compliance signals

The determinations are the first of their kind from the OAIC on tracking pixels and signal increasing regulatory attention on digital tracking and targeted advertising practises.

Kind has emphasised that existing privacy laws apply to modern tracking technologies, regardless of how they are implemented or which third-party platforms are involved.

The OAIC’s guidance on tracking pixels notes that while the Privacy Act does not prohibit the use of the technology, organisations deploying third-party tracking pixels on their websites should conduct appropriate due diligence to ensure they are used in a way that is compliant with the Privacy Act and Australian Privacy Principles, and should ensure sensitive data isn’t disclosed to third-party platforms and is only collected with consent.

“Organisations must ensure their privacy policies and notifications contain clear and transparent information about the use of third-party tracking pixels,” the guidance says.

It says organisations should adopt a data minimisation approach and conduct regular, ongoing reviews of tracking technologies deployed on their website to ensure use remains appropriate and complies with privacy obligations.

The post Invisible tracking tools face privacy crackdown appeared first on iStart keeping business informed on technology.

The IBM Institute for Business Value’s Calculus of AI Sovereignty study, which surveyed 1,000 senior executives responsible for AI, data, tech or related enterprise capabilities globally, found that 71 percent of them already expressed concern that switching AI vendors or models would be difficult. At the same time, the report shows a gap in understanding of what exactly that dependency looks like, with 91 percent of respondents admitting they didn’t fully understand their dependencies across vendors, models and infrastructure.

“Sovereignty is not an on/off switch. It’s more like a set of dials, each tuned to a specific business impact.”

Meanwhile, 72 percent of respondents said they’d be willing to pay a 20 percent cost increase to maintain multiple AI vendors for strategic flexibility.

IBM says its research, which was carried out in collaboration with Oxford Economics, shows the value of that flexibility and leverage. Organisations in the study who had the greatest control across their AI stack protected 55 percent more operating profit from AI-driven disruption than those with less control, the report says.

The study also highlights the operational implications of those dependencies as AI moves into core business functions. Executives reported an average of six AI-related disruptions over the past two years, largely linked to vendor services. In addition, 81 percent said a seven-day outage at a primary AI provider would cause severe or critical disruption to operations, and effectively halt operations.

“In AI, dependency extends into the model layer and the services around it, introducing new forms of volatility,” the report says. Models can change behaviour without formal release cycles and service terms and safety controls can be updated with little notice. Respondents cited unexpected changes across the AI ecosystem, from price increases and usage restrictions, to model deprecations, changes to privacy and data handling terms, performance degradation and new geographic access limitations aka the Anthropic shutdown of Fable and Mythos.

Control tightens and regulators respond

The report comes hard on the heels of warnings from Forrester that changes to SAP’s API policies, which came into force earlier this month, will restrict third-party AI agents, large-scale data extraction to non-SAP environments and workarounds through proxies, gateways, customer code or impersonation.

The changes are being enforced through platform updates and supported by SAP’s own AI and data services, which define the approved pathways for access.

The restrictions affect how organisations connect external AI systems, including third-party models and agents, to SAP data and processes. The vendor made its agentic Joule Studio 2.0 free – at least until December 31 2026, or in Forrester’s words ‘free to adopt, priced to entrench’. Agent runtime is free as is A2A interoperability – with no cap.

“This is the most aggressive commercial move SAP has made in a decade,” the Forrester analysts – Faram Medhora, Mark Moccia and Stephanie Balaouras – say.

“The 2027 pricing for Agent Gateway throughput at scale, A2A consumption, BDC egress, and Joule Studio post-promotion remains undisclosed. Most customer 2026 budgets do not model this cliff. The free-through-year-end window is SAP’s path to getting the customer on the highway, knowing that the customer will inevitably pass through a metered tollbooth down the road in 2027.”

Forrester has gone so far as to say SAP are attempting to become the gatekeeper of enterprise AI, urging CIOs to ‘push back’.

At the same time, regulators are examining whether similar dynamics are emerging across other enterprise software ecosystems.

In the United Kingdom, the Competition and Markets Authority has opened an investigation into Microsoft’s business software ecosystem as part of the digital markets competition regime. It’s accessing whether Microsoft’s licensing practices and product integration could reduce competition as AI becomes embedded in businesses.

“The UK will benefit most where customers can access the best tools in the market, and mix and match software and AI services from a broad range of competing suppliers. It is therefore important that competition in business software is working well,” the CMA said as it launched the investigation in May.

It says it has heard that UK businesses may not always be able to effectively combine Microsoft’s software with that of other providers, limiting the ability to get access to the best products at the most competitive prices. The investigation will include how AI competitors are able to integrate with Microsoft’s business software, giving customers access to AI software across suppliers to best suit their needs. The results of the investigation must be announced February 2027.

Those developments align with the dependency trends identified in the IBM study, which found organisations are already operating across multi-layered AI environments, but lack full visibility into how those dependencies are structured. Dependencies extend across vendors, models and infrastructure, creating challenges in assessing risk, managing change and responding to disruptions.

‘Selective’ AI sovereignty

The report positions ‘selective’ AI sovereignty – the ability to apply control across data, models, and AI infrastructure in proportion to business risk and strategic value –  as the key to maintaining control over AI.

“In reality, enterprises don’t control AI. They control a tangled web of contracts, architectures, data flows, and operational choices spread across a complex estate,” the report notes.

Seventy-five percent of executives who have switched, or attempted to switch, AI vendors in the last two years say the process was difficult due to data portability, model revalidation, compliance requirements and technical lock-in.

“Sovereignty is not an on/off switch. It’s more like a set of dials, each tuned to a specific business impact.”

The report recommends organisations classify AI systems into tiers – mission critical or differentiating systems, important but non-differentiating capabilities, and operational or commodity services – and apply different sovereignty expectations to each.

The report says the tiered approach establishes a shared language for decision-making, enabling CIOs, COOs, CFOs and business leaders to align on how much control is required, where vendor dependency is acceptable, and where optionality must be engineered. It also creates a consistent way to coordinate decisions across the enterprise – managing policies,

dependencies, and trade-offs with shared visibility.

“Not all AI is equal, and the operating model must reflect that.”

The post AI promise meets lock-in reality appeared first on iStart keeping business informed on technology.

Amanda Williamson, director of Deloitte’s New Zealand Artificial Intelligence Institute, says most businesses are seeing productivity gains from AI use, but few can link that to revenue gains.

“Experiments do seem to be quite easy to do.”

“AI’s value is hitting the cost line, where it’s easy to see, but not the revenue line. So CFOs are not really clear on how to book it as growth,” she told iStart.

At the same time, research company Omdia says organisations are still approaching AI as a technology investment, rather than defining clear business outcomes – creating a ‘blind spot’ in measuring real returns.

One of the consistent themes across both sets of research is a disconnect between perceived and measurable value. On the ground, teams are using AI tools extensively and seeing benefits, particularly in areas like back office processing and customer interactions. But those gains are rarely tracked in a way that ties back to financial performance.

“We’re really bad at measuring AI right now,” Williamson says. “We’re not tracking value like we would other investments.”

The Omdia research, which included Australia and New Zealand and was conducted for Boomi, found 34 percent of organisations were unable to effectively measure the success of their AI initiatives.

The ROI blind spot

Michael Barnes, chief analyst enterprise IT Asia at Omdia, argues the problem starts with how organisations approach AI in the first place.

“There’s the erroneous assumption, or misguided belief, that somehow technology has value in and of itself,” he told iStart. “That’s simply not the case.”

AI is being treated as an extension of existing IT spend, rather than a strategic business transformation, meaning each new initiative adds complexity, rather than value.

He says organisations need to start with outcomes – something most are not doing.

“Experimentation with no clear end goals isn’t going to cut it. Think in terms of actual business outcomes and then work backwards towards the AI strategy.”

He admits it’s not easy. “It’s hard to engage the business, or have the business lead discussions with a focus on outcomes, when they don’t fully understand the capabilities of the technology.”

Without that outcomes-led approach, AI initiatives remain disconnected from the metrics that matter. And that, in turn, makes it difficult to justify investment and even harder to scale projects and reposition AI as a growth driver.

“Organisations need to have a better sense of outcomes in order to justify budgets for AI within particular business units,” Barnes says.

“The more different business units can clearly understand the value of AI and link that to their measurable outcomes, the more we will see AI budgets linked to those particular business units,” he says.

A/NZ decision makers were the most pragmatic and least enamoured of the potential value of AI and slightly more focused on the challenges that need to be overcome in Omdia’s research, with New Zealand decision makers being even more so than Australian respondents.

In growth markets such as the Philippines and Malaysia, there was a significantly higher expectation that technology would drive business growth or innovation.

That was also clear in the Deloitte research where just 23 percent of Kiwi CFOs saw application of technology, including AI, as a top three growth driver. That’s compared with 30 percent Apac-wide.

Acquiring new customers, increasing sales to existing customers, operational efficiencies, price increases and innovation all ranked ahead of technology for New Zealand CFOs.

Stuck in productivity mode

A key issue remains where organisations are focusing their AI efforts.

Many AI deployments are aimed at improving individual productivity – helping staff generate content, process information more quickly or automate small tasks. While useful, those gains don’t always translate into measurable business impact.

“Focusing on individual level productivity is only going to give us so much in terms of actually getting a return on investment from pilots and AI in general,” Williamson says. “We really need to focus on the whole of the workflow – where is AI really going to make a difference?”

Instead of a scattergun approach, putting efforts into many different ‘hobby projects’, she’s calling on local organisations to focus on where AI might really shift the dial.

“I’m seeing a lot of good effort and good time being spent on projects that you probably can tell upfront would not have been worth the time and effort. So get to the core of value.”

The inability to prove value is also showing up in the gap between experimentation and scale.

While AI pilots are widespread across Australia and New Zealand – covering everything from invoicing to customer-facing tools – many stall before delivering enterprise-wide impact.

“Experiments do seem to be quite easy to do,” Williamson notes. “But getting it into something that scales can be a real challenge because scaling forces you to confront clean data, a motivated team, a real workflow.”

The measurement gap is being reinforced by two practical challenges: Data and cost.

On the data side, both Williamson and Barnes point to long-standing weaknesses in data quality, integration and governance, that were often deprioritised in favour of more visible initiatives. As Barnes notes, AI has made data quality issues ‘very obvious, very quickly… up to board level’.

On the cost front, AI introduces a new level of unpredictability compared to traditional IT investments, with Williamson noting AI is opening up a whole new era of calculating costs. The shift – or tokenomics – makes it harder to build a reliable business case. (You can read more on tokenomics and how to address the issue in our earlier story.)

What needs to change

For CIOs, CTOs and CFOs the message from both Williamson and Barnes is consistent: The issue is not lack of capability, but lack of discipline.

Organisations need to define outcomes clearly and at a function level, with broad concepts like ‘productivity’ or ‘automation’ not sufficient.

“Outcomes for AI need to be context specific – they’re going to be very different by function,” Barnes says. “You can’t talk about process improvement or productivity improvements, that doesn’t necessarily mean anything in business terms. The question is always towards what end?

“Yes, you’ve improved productivity, but what does that mean for your staff, whether it’s your call centre or your fraud detection unit? What are you enabling staff to do that they can’t currently do – that needs to be thought through because in most cases, it isn’t simply a cost saving exercise.”

Measurement needs to be built from the outset, including clear baselines and post implementation tracking. “Measure what is the efficiency is without it and what it is afterwards,” says Williamson.

She says organisations also need to focus on end-to-end workflows, rather than isolated use cases, saying that’s where AI can begin to drive meaningful, measurable impact.

And finally, foundational issues, particularly around data, must be addressed if organisations want to move beyond experimentation.

The post AI value lost in translation appeared first on iStart keeping business informed on technology.

The admission relates to a February 2022 market update in which the ASX said the upgrade was ‘progressing well’ despite internal classifications at the time showing significant risks and unresolved issues.

“As market operator and a steward of critical market infrastructure, our words matter.”

Under the proposed resolution, agreed to by the ASX and ASIC (the Australian Securities and Investments Commission), the two organisations will ask the Federal Court to find the ASX breached provisions of the ASIC Act, impose the $20.5m penalty and order ASX to pay an additional $3m towards ASIC’s legal costs. The settlement is subject to court approval and will avoid a trial in proceedings first filed by ASIC in August 2024.

Other allegations of misleading statements – including claims by ASX that the project was ‘tracking to a published plan’ and ‘tracking to go-live in April 2023’ have been dropped as part of the settlement.

At the centre of the case is the Chess (clearing house electronic subregister system) replacement project, a long-running effort to modernise the clearing and settlement infrastructure underpinning Australia’s equities market, with a distributed ledger based platform. The system sits at the core of how trades are processed, settled and recorded, making it critical financial market infrastructure.

ASIC says the ASX misled the market by overstating the health of a project which, internally, was already experiencing significant delivery challenges. The ASX has since admitted that as early as December 2021 the project was no longer on the critical path required to meet its planned April 2023 go-live date. At the time of the February 2022 announcement publicly claiming the project was progressing well, the project was classified ‘red’ internally – denoting significant unresolved issues or risks – and had been at that status since December 2021. Industry test environments had been opened, and others were planned, but were unable to do all that the system had been scoped to do.

About six weeks later, in March 2022, the exchange disclosed that go-live would likely be delayed. The project, which began in 2016 with initial expectations of rollout around 2020, then moving to a planned go-live of April 2023, was subsequently paused and ultimately abandoned after six years work and $245-$250 million in spend.

A review by Accenture in 2022 was scathing, finding significant challenges and deficiencies in the project.

In 2023 the ASX offered $70m in incentives, including rebates on clearing and settlement fees, to encourage brokers to help redesign the system.

Market-wide implications

Sarah Court, ASIC chair, says the ASX’s admission of a misleading statement went to the accuracy of disclosures about a major technology initiative with market-wide implications.

“ASX has admitted to making a misleading statement in relation to critical market infrastructure at the centre of Australia’s financial system,” Court says.

She added that accurate and timely disclosures are fundamental to maintaining trust in Australia’s financial markets, particularly from entities that operate core market infrastructure.

David Clarke, ASX chair, says the settlement reflected the exchange’s responsibility to ensure the market can rely on its communications about major operational programs.

“The market must have confidence in what ASX says about its operations as these statements can be relied upon to make decisions,” he says.

“When we stopped the Chess project in November 2022 to reassess our whole approach, that tested market confidence in ASX and called into question the nature of statements previously made.

“As market operator and a steward of critical market infrastructure, our words matter. I am sorry ASX fell short.”

“Firmer footing” for new system

Clarke says the project is now on a ‘firmer footing’.

ASX interim CEO Darren Yip says Chess remains a ‘critical priority’.

“Just two months ago, the team successfully delivered release 1 of the new system, providing clearing services on a modern, cloud-aligned platform,” Yip says.

That system is based on Tata Consultancy Services’ BaNCS for Market Infrastructure and Quartz Gateway offerings.

Release 1 replaced the clearing component and introduced financial information exchange messaging for trade registration. Release 2, focusing on post-trade modernisation, is targeted to go-live in 2029 and will replace the settlement and subregister functionality, deliver improved corporate action functionality and make further enhancements to clearing. The introduction of global standard ISO20022 messaging interfaces will also be part of release 2.

The case highlights the regulatory scrutiny that can be applied to communications about large-scale tech transformation programs, particularly where those programs underpin core business or market operations.

The Chess replacement project’s scale, complexity and integration across market participants meant delays and performance issues had implications beyond the ASX itself, affecting brokers, investors and other market infrastructure participants.

ASIC says the February 2022 statement exposed market participants to the risk of financial harm, underscoring the extent to which disclosures about technology delivery timelines can affect planning, investment decisions and operational readiness across dependent organisations.

ASIC says it has obtained commitments from the ASX to strengthen oversight, governance and oversight of the replacement program.

The post Chess project failure lands ASX $23.5m penalty appeared first on iStart keeping business informed on technology.

For enterprise technology teams in Australia and New Zealand, the impact was immediate: Tools they expected to use were suddenly unavailable, and outside their control.

“There are real and valid concerns for enterprises in terms of stability.”

Anthropic was forced to ‘abruptly disable’ its Fable model – launched on June 9 – along with Mythos, for all customers after an export-control directive restricted access by foreign nationals over ‘national security concerns’ – reportedly over concerns of jailbreaking. That category was impractical to enforce technically according to Anthropic, leading to a full shutdown.

The shock goes beyond a single model release, highlighting the confused state of AI regulation and signalling a shift in how risk needs to be managed.

Sam Higgins, Forrester vice president, principal analyst, says the Anthropic shutdown fits a broader pattern he flagged several years ago, saying the 2020s were a decade of systemic risk, including geopolitical risk.

“Geopolitics was going to drive a wedge between tech markets,” he says, pointing to a longer-term trend toward fragmentation.

That fragmentation may now be accelerating, with traditional assumptions that allied counties would have continued access being challenged.

“There are regularly carve outs for Australia, New Zealand… I was surprised not to see them,” Higgins says, of the lack of exemptions in the export controls.

The result is that enterprise access to frontier AI is no longer guaranteed – even in aligned markets.

“If traditional allies are cut off we will start to look elsewhere,” Higgins says, suggesting diversification away from US providers.

The shutdown comes against a backdrop of tension between Anthropic and the Trump administration. The company very publicly disagreed with the Pentagon, refusing to allow the US military to use its models for domestic surveillance and fully autonomous weapons systems. It was blacklisted by the Department of Defense as a ‘supply chain risk’.

Higgins notes that ‘it seems odd’ that Anthropic seems to be targeted again. “This feels like strike two,” he notes.

While the underlying motivation for the shutdown might be unclear, the outcome is not.

“There are real and valid concerns for enterprises in terms of stability,” Higgins says, noting the impact of policy decisions that extend beyond technical issues.

Cybersecurity experts have also been quick to criticise the ban and shutdown, describing it is ‘dangerous’ and warning that it has taken the best models away from defenders while failing to meaningfully limit attackers.

In an open letter, they also argue the capabilities cited – such as identifying software vulnerabilities – are not unique to Anthropic’s models and can be replicated using other tools. Removing the capabilities from security teams could increase overall risk while adversaries continue to develop similar capabilities unchecked, they say.

Local impacts

The impact has landed quickly with regional businesses.

Amanda Williamson, director of the Deloitte Artificial Intelligence Institute told iStart the move is an inflection point for organisations. “It’s a very big deal,” she says.

“With a large technology vendor essentially switching off their AI… I think we’re about to see many leaders wake up to the huge dependence that we have on our fundamental systems.

“As we build more and more on AI, we need to be more resilient if we’re going to have our fundamental operations relying on it.”

She notes locally we’ve been working with two AI superpowers – China and the US, highlighting reliance on external providers. “This will help organisations really think about this and raise the question of AI sovereignty.”

The shutdown forces a rethink of that dependency, bringing the issue of AI sovereignty to the forefront, she says.

“The definition of sovereignty isn’t necessarily clear and consistent from an organisational standpoint,” Michael Barnes, Omdia chief analyst enterprise IT Asia, told iStart. “ What they’re thinking about is how to mitigate risk and ensure they are able to continue to operate and have access to whatever underlying capabilities – including models for AI – that they need.

“This announcement is a stark reminder that we need to be more aware of control, and that some of that access is outside their control – and that level of uncertainty is likely only to increase. This isn’t just a blip.”

Barnes says he’s ‘absolutely’ seen a growing concern among decision makers in New Zealand in particular. “We’ve had some conversations where it’s very much a case of we need to act now, we need to accelerate our level of access or even creation of our own models – things that give us an increased level of control, or if not control, then certainly visibility so we can mitigate risks when and if they happen.”

Other regions are already investing in alternatives. Higgins notes Singapore has made significant investments in their own frontier models pouring ‘truckloads’ of money into sovereign AI.

“This will drive some real thinking about what sovereignty means,” Williamson says.

A recent Forrester Sovereignty Forecast put Australia at the bottom end of the tech sovereignty index. Higgins believes New Zealand – which wasn’t included in the report – would be slightly lower than Australia, given the lack of hyperscale cloud capacity.

Reframing enterprise risk

Williamson says many organisations have simply been switching on the AI model. Provided in whatever technology stack they have access to.

“But there are other things which can be done like using open weight models.”

Those open weight models – essentially open source – changes the cost structure, but does require having some hardware.

“Now is a good moment to start exploring what that could look like and if there are certain use cases where it makes sense to switch out the model and approach and get CFOs and engineers to come up with smart plans for doing AI efficiently”, she says.

It’s not all up to business however. Higgins says governments need to step up too.

“This goes to the bigger question of do we really understand where we have agency as it relates to our sovereignty and where we should and shouldn’t invest? And if we’re not going to invest, what’s the economic continuity plan?

“This is a risk assessment at the end of the day. If you can’t control the hazard and remove the risk, then you need at mitigation strategy – and at the moment it doesn’t feel like governments on either side of the Tasman have any mitigation strategy in place.

“We should ask ourselves some hard questions about what arrangements we have in place with the US so that this doesn’t happen again,” he says, while acknowledging that governments are currently dealing with some unprecedented policy shifts by an often challenging administration to deal with. “And if we can’t guarantee that, what’s our Plan B? Every business has a business continuity plan. What’s our ECP – economic continuity plan? That’s what’s missing and voters should be demanding that from our governments.”

The post A/NZ’s Anthropic sovereignty wake-up call appeared first on iStart keeping business informed on technology.

That warning from the Gartner distinguished vice president, analyst, encapsulates a turning point for enterprise AI, and comes as the Linux Foundation moves to tackle AI token cost management and address rising AI costs.

“Get clear and understand those token costs and alternative methods organisations have to remedy the ills of the token cost burden.”

The Tokenomics Foundation, launched by the Linux Foundation this month, will focus on establishing open industry standards, benchmarks and best practices for the economics of AI infrastructure. Its launch comes as enterprises grapple with rising costs and uncertain returns from generative AI and agentic AI.

At the centre of that shift is a fundamental change in how AI is priced – and how organisations experience its costs.

Unlike traditional enterprise software, which is typically priced per user or per licence, AI is increasingly consumed on a usage basis. Costs are increasingly tied to how much compute is used, measured through tokens – the basic, fundamental unit of data that LLMs process and generate. The more complex a task, the more tokens it requires.

And while per-token costs fell heavily during 2023-2025, the Linux Foundation says they have now levelled off and new model token prices are rising, making AI the largest and fastest-growing line item on enterprise technology budgets.

Dr Amanda Williamson, Deloitte New Zealand AI Institute Director, told iStart AI is opening a whole new era of calculating costs and introducing a level of unpredictability unfamiliar to many enterprise buyers.

“What we’ve been observing around New Zealand is a lot of leaders have been rolling out AI tools, and AI is a tricky little beast because you roll it out and suddenly you get surprises with consumption costs around it,” she says.

“As usage continues, sometimes we have no idea how much it’s going to cost over the next six months or even the next six days because it’s based on usage as opposed to seats,” Williamson says.

This shift – commonly referred to as tokenomics – means organisations are no longer buying fixed capacity. Instead, they’re paying for every interaction, query and workflow processed by AI systems.

For some, that’s hurting. Uber’s CTO Praveen Neppalli Naga recently revealed that the company had exhausted its entire 2026 budget on AI coding tools in just four months, driven by ‘token-maxxing’, where engineers were encouraged to maximise usage. Another, unnamed, enterprise business reportedly accrued a US$500 million bill for Anthropic’s Claude AI in a single month after giving employees unrestricted access and uncapped usage of API tokens for agentic AI workflows.

The economics of AI are also being shaped by supply-side constraints.

Chandrasekaran notes that generative AI was a ‘consumerisation phenomenon’. Tools used in personal lives have moved into the enterprise. The companies making those tools have an incentive to continue to serve the consumer market, because it’s a big brand factor and a pull factor for them into the enterprise.

Vendors prioritised growth and adoption, offering relatively low-cost access in order to build user bases – and the all important data advantages: the more users, the more data to train models on, the better your model becomes

“We were getting access to AI often at a subsidised rate, and we were often able to just purchase it with a per seat approach,” Williamson says.

That, however, was the old world. Williamson notes a few weeks ago, things started to shift.

“There are a few factors that are making it really hard for them to continue to subsidise AI,” Chandrasekaran says.

Alongside the token maxing trend, he points to the rise of AI agents. Unlike AI chatbots, which are more asynchronous workflows, AI agents have more autonomy and are able to make sequential or parallel requests to AI models running in the back end – leading to AI models being hit with more requests from AI agents. Goldman Sachs has estimated agentic AI could see token use increase by over 24 times by 2030.

A scarcity of compute, with construction of many of the data centres planned to support the AI boom now delayed, has also driven the shift from surplus to scarcity, with direct implications for pricing.

Organisations that previously relied on predictable, low-cost access to AI are now facing the prospect of rising and variable costs tied to usage patterns, as vendors GitHub with AI Copilot, start transitioning to a token-based – or consumption – pricing and Anthropic and OpenAI move to pay-as-you-go models, charging business users for compute resources.

GitHub’s move prompted plenty of Reddit discussion and angst, with one user noting their bill would be going from US$25/month to US$750/month.

“Suddenly enterprises are worried because they don’t know what their downside is with the token-based pricing model because they don’t have tools, they don’t have instrumentation to really measure that token usage within the organisation,” Chandrasekaran says. “So there is a lot of worry, and rightfully so, around what the changing consumption pattern, the change in pricing model really means for long-term TCO of AI within the enterprise.”

Adds Williamson: “AI’s value is hitting the cost line… not the revenue line.”

While AI-driven productivity gains are widely acknowledged, many organisations are still struggling to quantify those gains in financial terms – particularly when it comes to revenue growth.

“Everyone’s using AI and feeling more productive, but very few can yet measure AI lifting revenue,” she adds.

As a result, AI is often being assessed through a cost lens, rather than as a driver of top-line growth.

At the same time, token-based pricing means those costs are becoming more visible—and, in many cases, harder to control.

For CFOs and technology leaders, this is creating a new area of focus: Managing token consumption.

Token usage can increase rapidly as organisations adopt more advanced AI capabilities, particularly those involving multiple chained interactions or automated workflows.

Even relatively simple use cases can drive higher-than-expected costs if usage scales across teams or business units, making cost visibility and measurement critical.

Engineering for efficiency

As token costs rise, organisations are being pushed to adopt more disciplined approaches to AI deployment.

One key lever is efficiency in model selection and system design.

“You don’t always need to use the best, most token-hungry AI models,” Williamson notes, highlighting the importance of matching model capability to use case.

“This is a great moment to start thinking about what models we’re using and when.”

Until now, most organisations’ approach has been to switch on the model provided in whatever tech stack they have access to, but Williamson notes use of openweight models – effectively opensource models you can get without paying normal fees – changes the cost structure, though they do require having hardware.

“There are certain use cases where it might actually make sense to switch out the model and switch out the approach and get the CFOs and engineers together to really come up with a smart plan for how to do AI efficiently,” she adds.

Optimising workflows – reducing unnecessary queries and improving system design – can have a significant impact on token usage.

Ultimately, tokenomics represents a shift in how organisations think about AI – from a relatively predictable tech investment to a dynamic, consumption-driven service, requiring a new level of financial discipline.

“Get clear and understand those token costs, how the world of token costs is shifting and alternative methods organisations have up their sleeve to be able to remedy the ills of the token cost burden,” Williamson says.

“Learn about token costs of AI and how to do it efficiently. There’s always going to be a cost-benefit analysis to decide to go forward or not with AI, and if the costs cannot be managed, then it’s going to be a non-starter.”

For many organisations, that means building closer alignment between finance and technology teams, improving measurement frameworks, and developing a clearer understanding of how AI usage translates into business value.

Standards emerge as costs rise

The Linux Foundation’s move to establish the Tokenomics Foundation reflects the growing importance of these issues.

As enterprises navigate shifting pricing models, rising infrastructure costs and evolving usage patterns, the need for standardisation and best practice is becoming more acute.

By focusing on benchmarks and economic frameworks, the initiative aims to bring greater transparency and consistency to how AI costs are measured and managed.

“Measuring and benchmarking token efficiency across different models and vendors is critical to how organisations make business decisions, but until now, there was no neutral home to develop the standards needed to measure token economics transparently across the entire supply chain,” says Jim Zemlin, Linux Foundation CEO. “The Tokenomics Foundation provides that neutral home, ensuring these standards remain open and community-driven.”

For enterprises, that could help reduce uncertainty—and provide a clearer path to scaling AI sustainably.

The post AI’s free lunch ends as token costs bite appeared first on iStart keeping business informed on technology.

Fei, a Gartner director analyst, says in many enterprise environments, agentic analytics can introduce additional cost, complexity and risk – particularly where decisions are routine, tightly regulated or based on immature data foundations.

“Bad insights become bad decisions at scale. And that’s a much bigger problem.”

Analytics and business intelligence platforms are increasingly integrating AI agents to transform how data is analysed and insights are delivered, and businesses across Australia and New Zealand are increasingly experimenting, often layering new capabilities onto existing data environments. Many already have well-established reporting platforms in place and are now looking to extract more value – particularly through natural language interaction and automated insights.

But Fei told attendees at this week’s Gartner Data and Analytics Summit in Sydney that the push to add agency to analytics workflows needs tighter discipline. She says AI agents fundamentally change how analytics operates, and that shift is not suited to every use case.

At its core, agentic analytics moves organisations away from the traditional linear BI workflows – where data is cleaned and transformed, queried and visualised in dashboards, step-by-step, before insight is interpreted manually – into systems where agents handle much of the heavy lifting. These agents can set multi-step processes, automating data preparation and generate insights with minimal human intervention and in a more conversational based interface.

“The biggest trend here is that Ai agents are going to proactively detect changes in context and adjust insights and actions,” she says, noting that the key thing about agents – differentiating them from traditional automation – is they are goal-driven, not task-focused.

That shift has clear appeal, but also introduces new requirements around governance, trust and data readiness that many organisations are still grappling with.

Where agents don’t belong

Fei was clear that agentic analytics isn’t a universal fix, and is ‘overkill’ in many cases and actively counterproductive.

“There are some situations where it makes sense to just adopt a traditional reporting tool or BI, rather than agentic analytics.”

Among those are highly regulated environments, which need strict compliance and auditability. Traditional analytics workflows remain better suited where decisions must be fully traceable, deterministic and auditable. In areas such as large financial reporting, organisations need to understand exactly how outputs are produced – something that can be challenging with systems built on large language models.

Similarly, routine business processes offer little justification for introducing agents, Fei says. Where the objective is simply to view static metrics or track recurring KPIs, without the need for adaptive decision, agentic analytics adds unnecessary overhead without delivering additional value.

“These are low ROI use cases,” Fei says, noting that agentic systems typically require significant investment across compute, integration and governance.

In these scenarios, organisations risk over-engineering solutions and introducing complexity where existing tools already meet business needs.

Data maturity still dictates success

Beyond use case selection, the success of agentic analytics hinges on the underlying data environment.

Fei is clear: Organisations with immature data foundations are unlikely to succeed. Instead, she recommends addressing data integration and quality issues before attempting to deploy AI agents at scale.

“AI won’t solve your data issues. It only amplifies them,” she says.

“If the data is immature itself and the overall integration readiness is low… you won’t be successful with agentic analytics.

That has implications for organisations attempting to build more advanced capabilities on top of existing reporting systems. While there is clear momentum to make data more accessible and interactive, the underlying governance, semantics and quality still need to be in place.

The cost of getting it wrong

Agentic analytics doesn’t just change how insights are produced, it changes where risk sits.

Fei notes once organisations introduce AI agents into analytics workflows, the stakes move beyond reporting errors. “The truth is AI will make bad analytics organisations fail faster.”

Impact will be amplified. “The impact is not just bad insights but bad decisions at scale. That’s a much bigger problem.”

She says that helps explain why 32 percent of respondents in a recent Gartner survey believe truly usable AI agents remain a distant reality. The same survey showed only 19 percent of those surveyed are using AI agents to some extent on a daily basis.

Avoiding agent washing

Fei also called out agent washing. “Not everything offered or claimed as an AI agent is indeed an AI agent.”

She says agent washing is emerging as a prevalent trend, with vendors rebranding legacy features as AI-powered or agentic capabilities, often without meaningful innovation. In some cases products rely on basic natural language queries, scripted workflows or pre-defined reports, while presenting them as autonomous systems.

The absence of key features such as multi-step reasoning, hypothesis testing or proactive recommendations, signals the tools are not agents.

That widespread agent washing is threatening to undermine customer trust and stall true innovation, and is creating market friction and pervasive AI fatigue among B2B buyers.

Fei urged organisations to adopt an evidence-based evaluation when it comes to agents, evaluating them on real-world scenarios, and implementing task-driven acceptance tests. That includes assessing how systems handle enterprise data volumes, complex integrations and governance requirements. She also recommended engaging in detailed roadmap discussions with vendors and urged prioritising explicit transparency features, such as code visibility, confidence scores and semantic grounding, when selecting agentic offerings.

The path forward

Rather than pursuing broad deployment, Fei points to a more structured, refined, AI adoption model.

The first step is defining where agentic analytics truly add value, focusing on use cases with clear decision logic, strong data foundations and a need for adaptive insights. She suggested picking three to five low-risk high-value use cases as starting points.

From there, Fei says organisations can take a stage approach, establishing cross-functional teams, testing shortlisted platforms, running proof of concepts with task-driven acceptance tests and ultimately scaling wins, raising autonomy where trust is earned and adopting multiagent platforms.

Fei’s underlying message is clear: Agentic analytics represents a significant evolution in how organisations interact with data – but it is not a replacement for all existing analytics practices.

“It’s not everywhere for all your use cases,” she says.

The post Where agentic analytics becomes costly overkill appeared first on iStart keeping business informed on technology.

The Hays Salary Guide FY26/27 shows a market caught in an uncomfortable middle: Pay rises are modest or non-existent for many workers, yet employers continue to report widespread skills shortages across industries.

Despite softer salary movement, organisations are still struggling to find the capabilities they need. Hays reports that 82 percent of organisations experienced skills shortages over the past year, up from 79 percent the year prior, with Australia (79 percent) feeling it harder than New Zealand (75 percent). Tech, however, isn’t feeling the pain as badly as many other sectors – Hays has the ‘technology’ category experiencing skills shortages seven percent lower than average, with ICT 14 percent lower than average and software and services three percent lower. Hurting the most is the engineering sector – it is +12 percent above average when it comes to shortages.

“Budgets may be tighter, but critical tech roles don’t wait.”

At the same time, the guide highlights that pay growth has been modest, with hirers reporting an average salary increase of 3.3 percent in New Zealand and 4.0 percent in Australia across all roles. Many (42 percent), however, are reporting either no increase, an increase below 2.4 percent or even a decrease.

The result is a market where demand remains strong, but pricing power is constrained. That’s forcing a shift in how organisations compete. Instead of relying purely on salary increases, employers are being pushed to look at broader workforce strategies, particularly in high-demand areas such as technology and digital.

Robert Half’s 2026 New Zealand Salary Guide reinforces the trend, noting that organisations are rethinking compensation strategies to attract professionals with the specialised skills needed to ‘maintain a competitive edge’ particularly as digital transformation continues to drive hiring decisions.

What roles are earning

While overall salary growth is muted, pay for technology roles, particularly specialised ones, is holding up.

The Hays salary guide includes detailed benchmarks across dozens of tech roles in Australia and New Zealand.

Technology salaries in New Zealand ranging from around $55,000 to $80,000 for entry-level service desk roles, while the same role in Australia garners from $51,000 to $80,000.

Those with SAP expertise can also command higher salaries, with the difference between Australia and New Zealand also more stark – and the differences between Australian states also more apparent. While an SAP Functional Consultant will see a ‘typical’ salary of $200,000 in NSW, in Queensland it drops to $160,000, and in Tasmania and NT its $145,000 and $150,000 respectively. Across the ditch, Kiwis in the same role will see a typical salary of $160,000.

More senior Dynamics 365 professionals are also commanding big salaries, with functional consultants on par with their SAP counterparts, as are Oracle and Salesforce functional consultants.

Australian AI engineers are seeing typical wages of $155,000 (Tasmania) to $180,000 (NSW and ACT), while their Kiwi counterparts are typically receiving $165,000 (Christchurch) to $180,000 (Auckland). Senior AI engineers meanwhile are commanding up to $250,000 in NSW. It’s the GenAI engineers, however, who are reaping the big returns, with typical ACT salaries $220,000 – and up to $285,000. NSW follows closely with a typical salary of $215,000 and a high end of the range of $280,000. In New Zealand, it’s much more muted, with Auckland genAI engineers commanding the highest typical salary of $170,000, and up to $210,000.

Data and analytics roles are also showing strength, as are cybersecurity roles. CISOs in Victoria are seeing between $200,000 to $350,000, with a typical salary of $280,000; NSW and ACT are seeing between $200,000 to $340,000, while New Zealand CISOs are seeing between $180,000 to $350,000.

Robert Half data meanwhile shows entry-level IT support roles sit around NZ$60,000 to $70,000, while more experienced systems administrators can command up to $95,000. More specialised roles attract significantly higher salaries. An AI engineer can command $120,000 to $160,000, with an AI tech lead, driving the design, implementation and delivery of AI solutions, commanding up to $220,000.

Hiring pressure and retention risks

Robert Half’s IT-focused data shows talent shortages, digital disruption and economic uncertainty are all shaping the tech hiring environment, with organisations recognising that skilled IT professionals are critical to maintaining operations and competitiveness.

Even in a more cautious market, hiring urgency remains a factor. A significant proportion of tech leaders say the need to fill roles quickly is influencing their willingness to increase salary offers during negotiations. In other words, budgets may be tighter, but critical tech roles don’t wait.

That tension is reflected in how organisations are allocating compensation. While overall salary growth is moderating – Robert Half puts increases for most employers at three to five percent – employers are still prepared to pay more for candidates with in-demand skills, particularly those tied to ‘rare or emerging skills’ particularly in AI, data engineering and automation. Those three areas are also seeing counteroffers becoming more frequent as companies compete for talent.

With salary budgets constrained, both Hays and Robert Half point to a broader shift in how organisations compete for talent. Non-monetary factors, including flexibility, training opportunities and career development, are becoming increasingly important in attracting and retaining skilled employees.

The rise of hybrid work also reflects this shift. Hays reports hybrid arrangements are now the norm for a large proportion of employees, signalling that flexibility is no longer a differentiator, but an expectation.

At the same time, employers remain willing to negotiate for candidates who bring specialised skills and can deliver immediate impact.

While a topline glance might suggest a relatively stable labour market, the data suggests pressure is building.

Hays’ findings show many professionals are feeling underpaid, though despite the rising cost of living, dissatisfaction scores are largely in line with those seen last year, with 29 percent feeling dissatisfied or very dissatisfied (seven percent) and 29 percent feeling neutral. Just nine percent feel very satisfied, with the most satisfied tending to be more senior, higher income earners who had had a greater salary increase in the past 12 months.

Among tech workers, that isn’t translating into increased job hopping, however, with IT managed service providers, ICT and technology workers among those most likely to have longer tenures.

AI demand rising faster than capability

Compounding the issue is the acceleration of AI adoption.

Hays data, which is drawn from surveys with more than 7000 people across Australia and New Zealand, shows that 60 percent of employees are already using AI at work, but only a minority (22 percent) have received formal training. That flies in the face of what employers are saying, with just 27 percent of employers reporting no AI training or support is provided.

That gap is creating a new layer of demand, not just for technical roles, but for employees with the practical capability to apply AI effectively. Hays notes that on the hiring side, no agreed standard exists for evidencing AI capability. “[Employers] are split: 69 percent point to portfolio or practical examples, 52 percent to internal assessment, 49 percent to professional references, while 43 percent say none of the listed credentials suffice. Forty percent are seeking university or formal academic qualifications.

“The competency question is open and that ambiguity is itself an opportunity to bring structure.”

The post The hot tech skills demanding big dollars appeared first on iStart keeping business informed on technology.

The deployment, built on Amazon Connect, has shaved more than two-and-a-half minutes off average calls for the 150+ year old company, one of New Zealand’s largest publicly listed general insurance companies, removing more than 796,000 minutes of customer time across sales, service and claims interactions.

“By bringing interactions into a single platform and integrating customer data, we’ve enabled real-time meaningful AI support for our frontline teams.”

But the more telling detail sits behind those numbers.

Paul Johnston, Tower CEO, told iStart the project centred on consolidating customer interactions into a single platform and combining that with integrating customer data to deliver AI support as conversations unfold.

“By bringing interactions into a single platform and integrated customer data, we’ve enabled real-time meaningful AI support for our frontline teams,” Johnston says.

That support is delivered during live calls, with real-time transcription across every call and AI-assisted guidance removing the need for agents to pause interactions to take notes or search for information, and there’s less chance of missing important details.

Rather than building bespoke AI capability, Tower leaned on what was already available within its platform, with Johnston saying most of the capabilities were delivered out-of-the box, with effort focused not on customisation, but integrating those tools into day-to-day operations.

The core of that work sat in data and knowledge.

“Our focus has been on strengthening our core knowledge base and integrating this seamlessly with the platform, ensuring agents receive accurate, up-to-date guidance,” Johnston says.

That meant ensuring information feeding AI responses was well-governed and current, so it could be surfaced in real-time to support customer conversations.

The deployment was designed to sit inside existing workflows rather than alongside them. Johnston says embedding AI into how work was already done reduced friction during rollout and allowed the system to evolve based on frontline use.

Real-time feedback loops

“We’ve been able to continually refine the platform based on what actually works best for our teams,” he says. Real-time feedback loops ensure continual improvements in responses. Involving frontline ‘change champions’ throughout also aided the process, he says.

As a result, Johnston says teams quickly saw benefits, with AI supporting conversations, reducing complexity and improving consistency in responses.

“Our ability to adapt and adopt has been a real strength, supported by a strong culture of innovation and a focus on bringing our people on the journey from day one.”

Alongside the reduction in handling times, he says employee and customer net promoter scores (NPS) have improved, while quality assurance coverage has increased.

The system is also generating a richer data set to inform future product and service decisions.

Monitoring is built into the deployment. Johnston says quality is assessed – and improved – using automated evaluation alongside agent feedback, providing real-time visibility and enabling issues to be identified and addressed as they emerge.

Across the project, Johnston points to a combination of strong data, well-governed knowledge and workforce engagement and empowerment as underpinning the outcome.

“A big part of our success has come from getting those foundations right, while continuing to build and improve over time.”

Johnston describes the work as part of a broader digital transformation, with further opportunities to expand how AI supports both customers and frontline teams.

“We see this as part of our ongoing digital transformation, with continued opportunities to scale and evolve how AI supports both out people and our customers,” he says.

“Looking ahead, we’re poised to invest further in technology and AI-enablement throughout FY26. This is a critical part of our strategy that will drive greater efficiencies and continue to enhance our customer experience over the medium- and long-term.”

The post Tower’s AI playbook: Embed, don’t bolt on appeared first on iStart keeping business informed on technology.

“Successful pilots are satisfying, but real value comes from end-to-end transformation,” the company warns – and most companies haven’t achieved that yet.

“Optimising a single activity, such as code generation, test creation or requirements drafting, simply shifts bottlenecks elsewhere.”

The warning comes as AI reshapes the fundamentals of software development. What began as incremental productivity gains from coding assistants is accelerating into something far more structural: A full redesign of how software is conceived, built and deployed. According to Bain & Company, AI is ‘creating a seismic shift in software development’ moving organisations from isolated use cases toward an integrated, AI-native development model.

At the centre of the shift is the move from AI-assisted to AI-led development. Today’s tools may still sit alongside developers, but Bain & Company argues the trajectory is clear with AI systems increasingly capable of executing entire workflows, not just individual tasks.

“This evolution isn’t incremental; it is redefining what’s possible,” the report says, pointing to the emergence of hybrid human-agent teams delivering dramatically higher output.

The scale of that change is already filtering through to executive expectations. In 2024, executives were forecasting productivity gains of 20-30 percent from AI in software development. Now those expectations have ‘surged’ with many now anticipating improvements of ‘five times to 10 times’ over the next several years, with exponential gains no longer framed as theoretical, but instead apparently becoming the benchmark organisations are planning for. By late 2026 more than half of the global executives surveyed for the report expect more than half of their engineering efforts to be agent assisted. By around March 2027, that number hits around 90 percent.

But while speed is increasing, reports suggest value is not automatically following. Bain & Company’s report, which is based on surveys of global executives and market research conducted across 2025 and 2026, makes it clear that faster coding alone won’t deliver those outcomes – and some companies are missing out on many of the productivity benefits.

“It’s not enough for engineering teams to deliver code five times faster; business teams must generate demand at the same pace, and operations must match that speed to deploy, scale, and support solutions in production,” the report says. Writing code faster in isolation, or indeed optimising a single activity, such as code generation, test creation or requirements drafting, simply shifts bottlenecks elsewhere.

The State of Engineering Excellence 2026 report from software platform Harness, reinforces that point. It found 89 percent of engineering leaders say productivity has improved with AI coding tools, but 81 percent say developers are spending more time on manual code review. The self-reported impact is ‘overwhelmingly positive – but the cost is accumulating in places organisations aren’t watching,’ Harness says, with nearly a third of the work now ‘invisible work’ – reviewing code, fixing bugs and context switching between tools.

This dynamic aligns with Bain & Company’s argument that traditional software development thinking, where improvements are made within isolated phases, is no longer sufficient.

“Rolling out lots of pilots may also feel like success, but pilots don’t necessarily translate into real usage or business impact. Without new workflows, measurement, and guardrails, companies are likely to see adoption plateau and only minimal new value,” Bain & Company says.

Instead it points to the rise of an ‘integrated AI development lifecycle’ which breaks down the traditional divide between product development (defining what to build based on customer needs, market opportunity and strategic roadmaps) and software development – two streams of work that often occur across different teams.

“AI shatters these boundaries, and it can define requirements, generate code, test, and iterate all within a more continuous flow.

“Companies are moving toward an AI development life cycle in which AI is embedded across the entire process and product and engineering operate as a more integrated system rather than sequential steps. Instead of product development defining the objective and engineering building it, AI-enabled teams continuously define, build, test and refine together.

 It’s a process, Bain & Company says, that requires a redesign of organisational structure with tech teams making adjustments to all levels of their engineering operations, with developers moving from code executors to agent architects and new roles created.

“The shift to an AI-led development life cycle is not just a technology upgrade; it’s a full-system transformation that rewires how organisations build, operate, and compete. “Companies that move decisively, redesigning workflows, redefining roles, and anchoring on measurable outcomes, will capture disproportionate value.”

The post AI coding runs into real-world limits appeared first on iStart keeping business informed on technology.

Unveiled at a company conference in London, the Meta Business Agent operates across WhatsApp, Messenger and Instagram, responding to customer enquiries, recommending products and booking appointments inside chat conversations. Meta says the agent can also qualify sales leads and complete transactions.

“The agent can also qualify sales leads and complete transactions, extending its messaging tools into revenue‑generating workflows.”

While Meta is dominant in online ads which account for the vast majority of its revenue, it has struggled when it comes to actually selling its products. The company’s messaging ecosystem, however, is already used as a communications layer for many businesses of all sizes.

Meta has famed the new offering as part of a broader enterprise push, with the company saying the goal is to provide businesses of any size with an always-available digital agent to handle customer engagement and operational tasks.

The company is also launching a ‘Business Agent Platform’ to allow organisations to build and customise their own AI agents, including integrations with third-party systems such as Shopify, Zendesk and Shopee.

“The platform provides larger businesses with enterprise-grade controls, guardrails and measurement built in so they can define rules and offer personalised experiences, starting with the messaging apps their customers already use,” Meta says.

The move also hints at Meta’s AI ambitions and desire to compete with the likes of OpenAI, Anthropic and Google with enterprise versions of their AI tools.

It also comes as competition intensifies in the market for enterprise AI agents. Companies including Google, OpenAI and AWS are developing tools designed to automate customer interactions and business workflows, targeting similar use cases in sales, support and operations.

The Business Agent places Meta directly into this category, with executives explicitly positioning it as an enterprise play, rather than an incremental update to existing messaging products. The launch builds on earlier chatbot capabilities, but extends them to enable automated actions, aligning with broader industry shifts towards ‘agentic’ AI systems that can take on tasks, rather than simply respond to queries.

Unlike some competitors, Meta’s strategy is to deploy the capabilities inside its existing messaging platforms, where businesses and consumers are already interacting at scale.

Meanwhile the Business Agent Platform places Meta in competition with vendors offering standalone AI agent frameworks and enterprise automation platforms. (As an aside, the Business Agent launch came just days after Meta floated the idea of becoming a hyperscaler and renting out its compute infrastructure.)

Chatting with customers, closing sales

A free test version of the Meta’s agent service was released in select markets, including Mexico and India, late last year under the ‘Business AI’ name.

At the launch of Business Agent, Meta said more than one million businesses were already using a Meta Business Agent – presumably the free test offering – on WhatsApp and Messenger.

The new Business Agent provides automated capabilities to allow companies to handle customer conversations at scale. It’s designed to answer questions specific to a business, make product recommendations from business catalogues, book appointments and qualify incoming leads, close sales and escalate more complex queries to human staff where required – with the business setting the parameters for when a team member will step in.

It’s also offering the ability to generate summaries of chats that occurred overnight and provide insights on threads.

Meta says the agent can be deployed ‘in minutes’ within its platforms or integrated directly into existing enterprise infrastructure and, while initially available for free, will be a paid subscription, believed to be a business-focused subscription tier for Meta One, the umbrella brand for Meta’s paid subscription offers, launched earlier this month. Larger enterprises are expected to be charged based on token usage, similar to existing pricing models for business messaging on WhatsApp.

While the initial rollout focuses on customer engagement and sales tasks, Meta is positioning the Business Agent as part of a broader toolkit for businesses.

It says it is developing additional features that would expand the agent’s role into areas such as conducting market research, surfacing product insights, connecting with tools to manage your calendar and providing competitive intelligence.

Challenges ahead

For enterprises, rolling out autonomous agents in customer-facing channels raises practical implementation questions, particularly around integration, governance and customer experience.

Meta says the agent can connect to third-party tools such as Shopify and Zendesk and can be embedded into existing business workflows.

For enterprises, deeper integration with systems such as CRM and ERP platforms would be required to align the agent with customer records, order management and service processes. Whether businesses are willing to connect those systems to Meta’s platform could be a factor for enterprise adoption, particularly given the sensitivity of enterprise data and existing system architectures.

While Meta brings scale through its messaging platforms, adoption at enterprise level will depend on how effectively those capabilities integrate with existing systems, operate within governance frameworks and deliver reliable customer experiences at volume.

The post Meta wants your CRM – via Messenger appeared first on iStart keeping business informed on technology.

Within two weeks of launching the feature in the realestate.co.nz app, more than 60,000 searches had been clocked up. It’s still a tiny percentage of the company’s searches, but  Simon Hargraves, realestate.co.nz chief information officer, is optimistic.

“We had just finished a project to get our data into a good state. Without that, it would have been extremely difficult to do this work.”

He told iStart users are increasingly comfortable working with AI and expectations for search are changing. “They want to be able to express their search intent more naturally, rather than just using predefined filters or having to use exact keywords.”

The AI-image search gives users the ability to search ‘at a completely different level’ – searching what’s actually in photos.

What users are searching for, however, hasn’t changed in early searches. “We have a list of traditional keywords people were searching for and generally it’s following the same trend,” Hargraves says.

What has changed is how effectively results are being surfaced. Traditional keyword search relies on exact inputs from both sides, with agents required to include the right terms in listings and users having to guess and input the same wording. AI image search removes that dependency, Hargraves notes, with the AI system examining the photos and pulling out key features.

The system also changes how results are presented, with the platform able to surface the image where the feature has been detected, rather than returning a listing with no obvious visual confirmation.

The accuracy challenge

Hargraves says the feature itself was built and released quickly – one month to build and another month spent testing, tuning and getting feedback from the business. “The really challenging part was more the focus on accuracy and making sure the results aren’t frustrating for users because the way the semantic search works is it expands what the user is searching for to understand intent and then expands the search out from there.”

As an example, early on a search for wine cellars associated them with luxury properties – returning a broad range of luxury properties rather than those with wine cellars. Worse, searching for a property with an accessible ramp expanded out to include staircases.

“Those kind of cases are obviously going to be super frustrating to users when you’re getting the opposite of what you’ve actually search for. So we spent probably half the project time actually testing it internally, across the whole company and getting feedback from users of examples where it was wrong. Then we spent quite a bit of time tuning the system.

“Making sure the experience wasn’t frustrating was the most important part.”

Data does the heavy lifting

Behind the feature sits a significant data effort.

Realestate.co.nz had just completed an 18-month re-platforming and centralisation of all of its data into a single Snowflake platform, standardising definitions and consolidating multiple data sources across the business.

“That was a huge foundational piece that allowed us to build it very quickly.”

The work enabled the rapid development of AI features, but also highlighted a clear dependency.

“You have to have that data foundation in place… if you have bad data, then you’ll get bad results out from the AI.”

In image-based search, that extends beyond structured data to the quality of photos themselves – something the company doesn’t control.

“The AI performs much better when the features [are] clearly visible,” Hargraves says.

But he notes the higher the quality the images and the larger they are, the more cost to process via AI. “There are quite a few variables that we had to take into account and tune to try and get the right kind of cost versus quality.”

He deliberately frames the AI image search as an experiment. “The technology has evolved really rapidly and these kind of features are becoming easier and easier to implement so it makes sense to build the feature and release it to users and see if it’s used and useful.”

The company is less focused on initial uptake than on whether users will return. As Hargraves notes, if users don’t find it useful, they won’t come back to it again – a key test for the future of the system.

To manage expectations, it has been clearly labelled as a beta. “Because… with AI, it’s never going to be perfect.”

Beyond search

The company’s broader approach reflects a pragmatic view of AI. Internally it is being used across the engineering process, and teams are encouraged to experiment through hackathons and rapid prototyping.

“These tools are making it super easy and super quick to build out things and test them internally and build proof of concepts,” Hargraves says.

At the same time, decisions to productise those ideas need to be made carefully, he says, noting the challenge of ensuring projects show ROI and aren’t constantly ‘experiments’.

“You have to make sure projects make sense… and as with anything its prioritising what you have in your roadmap case by case with cost versus value.”

In that context, AI image search is seen as a logical extension of existing capabilities, rather than a standalone transformation.

Hargraves says the company is already working to merge AI image search with traditional keyword search into a single, unified experience.

Longer term, there is potential to expand search into other areas, including accessibility requirements and proximity to amenities, though that will depend on the availability and quality of additional data sources.

“Finding a good data source and being able to ingest it and surface it via AI search is going to be the challenge. But I think you’ll see a lot of those features coming over time. It’s just about getting data into a good place first because without that solid foundation you can’t build high quality AI experiences on top of it.”

Beyond real estate

So what’s Hargraves advice for any other companies considering similar tools?

“Getting the data into a good place is probably the most complex thing. We’re lucky because we had just finished a project to get our data into a good state. Without that, it would probably have been extremely difficult to do this work.”

The data needs to be high quality, well-structured and accessible and good governance around it is also required as well.

“Without that, it’s going to take a lot longer and there’s going to be much higher levels of complexity.”

And be prepared to spend time tuning. “The tuning and accuracy piece is going to be the most difficult part to get right from a product perspective,” Hargraves warns.

“With AI image search you don’t have so many of the concerns you’d have with AI in other areas, especially where you have chatbots and people interacting directly with the AI. AI search is an easier place to start but as we get more into potential AI features in future – chatbots and that kind of thing – it becomes more and more concerning from a risk perspective.”

On the cost front, he says there wasn’t ‘massive’ cost. The time spent tuning was the bigger cost, with the company now spending around US$1000 a month for processing all of its listing images.

“Costs are increasing in AI across the board though, so it’ll be interesting to see how it changes over time and it’s something we’re going to have to keep an eye on.”

Hargraves also acknowledges that for some companies, AI image search could surface unexpected content.

“With our platform, the photos have all been taken and verified by vendors and agents, so personally identifying information or anything sensitive is usually scrubbed from the photos anyway. On different platforms, you would definitely have concerns though,” he says.

Despite that, Hargraves says it’s important for all local companies to be experimenting with AI.

There are huge changes coming across every industry. There’s a massive shift in the way people are working in businesses and how they are interacting with AI and that’s changing their expectations of how they use other products too.”

The post Realestate.co.nz’s lessons in AI-powered image search appeared first on iStart keeping business informed on technology.

A study in the Journal of Service Management flags ‘AI complacency’ – a silent erosion of critical thinking that occurs when people stop questioning automated systems. Or, as defined by the report, an employee’s tendency to neglect validating AI-generated outputs, even when errors are present.

“The biggest risk of AI isn’t that it gets things wrong, but that nobody notices.”

The research is based on six experimental studies involving more than 1,300 participants and identifies a clear root cause: Lack of accountability for monitoring AI outputs. When responsibility for checking results is unclear, employees are significantly more likely to accept outputs at face value, even when errors are present. Over time, this leads to what the report authors – Khanh Bao Quang Le from Auckland University of Technology’s Department of Marketing and International Business and Werner Kunz from the University of Massachusetts Boston’s College of Management – describe as a diminished willingness to evaluate AI-generated outputs critically, along with an increase in work-related errors.

“The biggest risk of AI isn’t that it gets things wrong, but that nobody notices,” the pair say.

“It’s not about lacking technical skills; rather the machine’s fluent presentation creates a false sense of security. Over time, employees may fall into a ‘human-out-of-the-loop’ routine.”

The study suggests this behaviour reflects how organisations are deploying AI into workflows without clear ownership or oversight. Where monitoring is treated as optional, rather than embedded, complacency takes hold.

The way AI presents information compounds the issue. Outputs are typically fluent, structured and authoritative, even when incorrect. This aligns with well-documented ‘automation bias’ where users over-rely on automated systems and accept recommendations without sufficient scrutiny. The report notes that as AI become more deeply embedded in day-to-day work, that bias is strengthened by operational pressures for speed and efficiency, particularly in environments dealing with high workloads or complex tasks. Where tasks were cognitively demanding and employees were juggling multiple competing demands, complacency increased. Similar was seen when working in interdependent teams with people assuming ‘someone else will catch it’.

The shift was subtle but significant with employees moving from actively interrogating information to passively accepting it, and in some cases, stopping output reviews altogether.

High-stakes consequences

The consequences are already visible in high-stakes environments. Case in point, the Australian lawyer sanctioned in late 2025 after submitting documents containing AI-generated false citations. He admitted he didn’t verify the contents.

Similar patterns are emerging more broadly across organisations in Australia and New Zealand. A mid-2025 report from KPMG and the University of Melbourne found that AI is already in daily use in organisations but often without structured oversight or training. Globally, Trust, Attitudes and the Use of Artificial Intelligence found two in three reported relying on AI outputs without evaluating the information it provides, with other half saying they have made mistakes in their work due to AI. New Zealand and Australian respondents rated among the lowest for AI knowledge, efficacy and training.

The consequences aren’t limited inaccurate decision-making, with inadvertent exposure of sensitive information also a possibility, the KPMG/University of Melbourne report notes.

And consequences are limited to governance or security either, with cost control emerging as a risk area. In one widely reported case, a company spent US$500 million in a single month on Anthropic’s Claude after failing to set limits on employee usage. The overspend was reportedly linked to a lack of controls over how the tool was accessed and used (one CTO noted employees were using AI for things they could easily do themselves, such as checking the weather).

Local evidence suggests the problem runs deeper than individual behaviour. Ethos Advisory, which runs AI governance reviews, says most New Zealand organisations score poorly in the assessments, with AI policies often little more than documentation.

It notes typical gaps include no defined accountability for AI decisions with no defined processes for identifying accountability, investigating the failure or notifying affected parties.

Designing in oversight

Banning AI or requiring manual checks aren’t required, Le and Kunz say, instead, they call for a more nuanced approach. Among their practical options:

Build accountability into the process, not just the policy ensuring employees know that someone, somewhere, will ask them to explain what they reviewed and why they approved it.

Set outcome expectations so people know they’ll be held accountable for the outcome, not just the process – this, Le and Kunz says, makes users stay more alert.

Design for high-complexity, high-load environments specifically. “These are the conditions where complacency is most likely to develop.” Rotating oversight responsibility, adding review checkpoints or simply reducing simultaneous task demands during AI-assisted work can help close the gap.

Take team design seriously. Assign clear, individual ownership for AI monitoring, even in collaborative workflows, to avoid a situation where no one feels fully responsible for AI outputs.

“These measures are not meant to slow down innovation,” Le and Kunz say. “Rather they help ensure that AI fulfils its promise without eroding human judgement.”

The post The silent threat of AI complacency appeared first on iStart keeping business informed on technology.