Deloitte: 35% of Australian firms see weaker AI returns than hoped

What Deloitte published

On 9 November, Deloitte's latest State of AI in the Enterprise report started appearing in the Australian business press. The report is subtitled "Built differently" and draws on a global survey of senior leaders. Australian Cyber Security Magazine puts the sample at 3,500 leaders across 25 countries. Deloitte's own Australian page uses the 3,500 figure in its introduction, but its methodology section lists 3,235 leaders in 24 countries surveyed in August and September 2025. Deloitte hasn't published the size of the Australian cohort, so read the local percentages as a directional signal rather than a census.

Even with that caveat, the Australian results show adoption becoming routine work at larger firms. According to Deloitte, 27% of Australian organisations now report revenue gains from AI, up from 20% in January, and reported efficiency or productivity benefits rose from 66% to 75%. Financial Standard reports that 39% of the Australian cohort are redesigning several key processes around AI, and 73% expect AI to handle most routine knowledge work within three years.

The returns gap sits beside the gains

The figure we'd pay most attention to is less flattering. Australian Cyber Security Magazine's summary says 35% of Australian respondents got less measurable ROI than they expected, against 19% globally. The same summary shows Australian firms running AI at scale in IT and cybersecurity more often than the global average (62% versus 54%). In strategy and operations, the pattern reverses: 23% of Australian firms operate AI at scale there, compared with 38% globally, and 69% are still piloting or experimenting.

We read those two numbers together. Australian organisations have pushed AI hard into the functions where IT teams own the tools and can measure the output, such as ticket handling, code and security monitoring. The functions where AI would change how the business runs, such as planning, scheduling, pricing and supply decisions, are stuck in pilots. Productivity gains in IT rarely show up as revenue, and leaders who expected a business-wide return are finding the benefits sit inside one department. The returns gap follows naturally from where the effort has gone.

Autonomy is growing faster than accountability

Financial Standard reports that 38% of Australian organisations let AI act without human approval to a moderate degree, while 18% say accountability is unclear when an AI system produces a bad outcome. Nearly one in five organisations can't say who owns a mistake made by software they've allowed to act on its own. When an agent issues a wrong refund, sends a wrong quote or closes a ticket it shouldn't have, someone has to answer to the customer, the auditor or the board. Firms that haven't named that person before go-live will name one during an incident, usually badly.

The cost picture is maturing faster. According to Deloitte, three-quarters of Australian respondents track total spending on AI inference and API consumption, but only 38% track token volumes and 25% track cost per token or API call. Most firms see the monthly bill but can't connect the spend to a particular workflow or outcome, so they can't tell which use cases pay for themselves.

What the Australian data means for NZ businesses

Deloitte's Australian page doesn't break out New Zealand, and no official NZ adoption figures have been released yet. The closest recent local data is Prospa's SME Sentiment Tracker, which IT Brief reported on 26 October. The tracker found 59% of NZ SMEs using AI in some form, up from 50% six months earlier, with only 7% treating AI as strategically important across several functions. Among NZ firms that haven't adopted AI, Prospa found uncertainty about return on investment was the top barrier at 34%, well ahead of upfront cost at 12%.

The Prospa and Deloitte surveys measure very different populations, so the percentages shouldn't be compared directly. The direction still matches. NZ businesses are adopting AI quickly at the level of individual tools, and the step into cross-functional, operational use is where both surveys show hesitation. Australian enterprises are a year or two ahead of most NZ mid-market firms, so their experience shows what NZ firms can expect. Running AI widely inside IT doesn't produce a business-wide return on its own, and handing decisions to agents without clear ownership leaves accountability gaps.

Most NZ organisations are smaller, which makes these problems easier to fix early. A 60-person distributor or a regional council team can map every AI tool in use and name an owner for each in a few weeks. A 40,000-person bank can't. Our AI consultancy work starts with that inventory because the decisions that follow (which workflows to redesign, where autonomy is acceptable, how to measure return) depend on knowing what's already running.

Practical next steps

• Set a return measure before the pilot starts. For each AI use case, write down the baseline (hours, error rate, cycle time or revenue per customer) and the figure that would justify keeping the tool. Australian respondents reporting weaker ROI than expected suggests many of them skipped this step.
• Pick one operational process to redesign. Choose a workflow outside IT, such as quoting, purchase approvals or job scheduling, and rebuild it around what AI does well, rather than adding a chatbot to the existing steps.
• Name an owner for every autonomous action. Any AI that can send, approve, refund, close or change records without a human click needs a named person accountable for its outcomes and a documented way to switch it off. The AI agent setups we design treat that owner and that off switch as part of the build.
• Track cost per workflow alongside the monthly bill. Tag API keys or projects by use case so spend can be matched against the return measure. Token counts per process are usually available from the vendor's console or gateway logs.
• Write accountability into policy and contracts. Update your AI use policy and any supplier agreements to state who carries responsibility when an AI-driven action causes harm or loss.
• Watch for official NZ data. Stats NZ's 2026 Survey of Business Operations is expected to give the first official adoption baseline. Compare your progress against that baseline instead of vendor surveys.

The Deloitte findings show that AI in Australia now delivers measurable gains for a growing share of firms, mostly in the functions IT controls. NZ businesses that measure returns from the start, put an owner on every autonomous action and choose one operational process to rebuild will avoid the returns gap Australian firms are now reporting.