Only 21% of Australian firms say AI earns more than it costs

Australian organisations are paying heavily for AI, and few of them believe the spending has paid off yet. Cloudera's State of Enterprise AI 2026 report came out on 27 October, and the Australian results followed on 3 November. Among Australian respondents, 76% invest at least US$500,000 a year in AI, while 21% say AI currently generates more value than it costs.

What Cloudera found

Australia looks cooler than its neighbours. Higher AI spending than a year earlier was reported by 43% of Australian respondents, against 58% across Asia-Pacific, where 28% said AI returns more than it costs. Australian respondents were both slower to add budget and less satisfied with the value.

The global figures show where the pressure comes from. Some 89% of respondents have limited or redesigned AI use cases because of rising token consumption. Tokens are the units of text that AI providers meter and bill for. Nearly a quarter (24%) have paused or cancelled a use case, and 19% continuously monitor their token and inference spending. The 19% figure stands out, because a company can't control a metered cost it only checks now and then.

Read the numbers with the seller in mind. Cloudera commissioned the survey from Wakefield Research, which polled 1,500 IT leaders across nine markets in September. The same article notes that Cloudera's core pitch is letting customers choose where AI workloads run, so a finding about workloads moving suits its sales message. The published coverage gives no Australian sample size, and New Zealand isn't broken out. We couldn't confirm whether NZ respondents were included at all. The percentages show direction and shouldn't be treated as precise measurements.

The direction does match other research. Elastic's survey, published in June, found one in three businesses went over their AI budget last financial year. Avanade's survey this week found 66% of Australian and NZ organisations haven't formally calculated their AI return. Both are vendor studies too, but they're separate samples reaching a similar conclusion.

Why the bill changes as adoption deepens

Early AI use is cheap because the tasks are simple. Commonwealth Bank chief executive Matt Comyn explained the shift in June: as models gain reasoning, tool access and larger context windows, spending climbs unevenly. In Comyn's words, "your token costs do not scale on a linear basis."

An agent that reads a contract, queries a system, retries a failed step and writes up the result consumes far more than a staff member's one-line question to a chatbot. Business cases built on pilot usage can therefore understate production costs by a wide margin. A Celonis executive's opinion piece cites Deloitte research on one healthcare enterprise whose token usage grew 8% to 10% a month, adding more than US$6 million in unplanned annualised cost. Compounded over 12 months, those growth rates multiply usage by roughly 2.5 to 3.1 times.

Governance adds a second cost line. Across the full sample, 55% said governance costs grew faster than expected, and 82% of Australian respondents said governance and compliance limit their ability to scale AI, compared with 75% globally. Reviewing outputs, logging actions and handling exceptions all take staff time, and we'd expect few AI budgets to list that time as a project cost.

What it means for NZ businesses

Most NZ organisations are probably earlier on this curve than the Australian enterprises in Cloudera's sample. Datacom's AI Index, based on 207 senior managers, found 91% of NZ businesses use AI, 68% rely on general-purpose assistants such as ChatGPT, Gemini or Claude, and 4% use AI to transform core operations. Our reading is that many NZ firms are still paying mainly for assistant licences, where the monthly cost is reasonably predictable, and haven't yet met metered agent bills at any scale. Datacom's report doesn't measure billing, so treat that as inference.

Being behind has a practical upside here. NZ firms moving from assistants to agents can set up cost controls before the bills arrive, while the Australian respondents are retrofitting them. The US$500,000 threshold won't apply to most NZ mid-sized companies, but the mechanics carry over. A smaller budget has less room to absorb a tripled usage bill, and fewer people available to investigate one.

Data location deserves the same early attention, especially for firms that work across the Tasman. In Australia, 74% of respondents say residency and sovereignty requirements influence where AI workloads go. Moving later is expensive, with 58% globally saying a move requires moderate or major redesign. NZ suppliers with Australian customers should expect those customers to start asking where AI processing happens.

Practical next steps

• Cost each use case per completed task. Work out the price of one invoice processed or one ticket resolved, including human review time, and give each use case a named owner.
• Set alerts and caps before a pilot goes live. Apply spending limits to each API key or workspace, and have finance see the numbers monthly.
• Stress-test the business case. Before approving an agent project, model what happens if usage grows 10% a month for a year, and decide in advance what the business would pause.
• Match the model to the task. The Celonis piece suggests sending routine steps to smaller, cheaper models and keeping expensive reasoning for hard problems. Celonis has a product to sell, but the routing principle is sound.
• Record where each workload runs. Note the provider, the region and the data involved at design stage, then check whether customer contracts restrict any of it.
• Read the pricing terms at renewal. Ask what's metered, what the limits are and what happens when you pass them, because assistant licences and agent features can be billed differently.

Firms without someone in-house to own this work can get outside help with an AI cost and governance review. DataForge's AI consultancy covers adoption and governance for NZ and Australian organisations. Stats NZ and MBIE's Survey of Business Operations includes AI questions, and no results had appeared by 19 October. Once published, the survey should give NZ firms a better baseline than vendor research can.