AI Autonomy in Australia: From 1% to 12% Adoption in Three Years

Autonomy, not AGI, is the decision in front of Australian boards

Australian Bureau of Statistics data for 2024-25 shows 12% of Australian businesses reported using AI, up from 1% in 2021-22. A twelvefold rise in three years is fast by any measure, but the aggregate hides the more useful fact: adoption is concentrated. It is already materially higher in Information Media and Telecommunications (38%), Professional, Scientific and Technical Services (24%) and Financial and Insurance Services (24%).

That concentration changes what a benchmark means. An Australian financial services business sitting on the national average is not level with the country; it is half the rate of its own industry, competing against rivals whose cost base and cycle times have already moved. The right comparison for a board is the sector line, not the headline.

Factor's research argues that the widely covered race toward artificial general intelligence is a red herring, and a distraction most business leaders cannot afford. The nearer and far more consequential shift is the generalisation of AI: broadly capable models diffusing into systems, workforces and operations, bringing a new level of autonomy long before AGI is settled. For Australian enterprises the immediate issue is not speculative AGI but how quickly general-purpose AI becomes embedded in regulated industries, public services, critical infrastructure and everyday knowledge work.

The report frames the destination as a cognitive digital brain built from four interconnected layers. Knowledge graphs and vector databases organise enterprise data; generative and classical models do the reasoning; agents plan, reflect and act with minimal human input; and architecture makes any of it repeatable at scale. Once those layers connect, the executive question stops being which tool to buy and becomes how much authority to delegate, and on what evidence.

Trust, not capability, is what caps how much autonomy you can deploy

The central argument of the report is blunt: we can only let systems be as autonomous as we trust them, which makes trust the determining factor in how far AI diffuses inside an organisation. That reframes the governance conversation. Trust is usually treated as a risk topic, owned by legal and security and discussed after the business case; the research treats it as the throttle on the business case itself.

The evidence that trust is already binding sits inside the workforce, not outside it. In May 2024, over 40% of users had started incorporating AI into their jobs just in the previous six months, and more than half of workers using AI are reluctant to admit it, worried that using it for important tasks makes them look replaceable. This is not employees distrusting the model. It is AI disturbing a settled relationship between people and their employer, in which defined roles, skill expectations and a shared understanding of performance translated into job security.

Shadow adoption at that scale carries a hard operational cost. An organisation whose people hide their AI use cannot see its own exposure, audit what data is leaving, or attribute the productivity it is quietly gaining.

Machine behaviour raises the same question from the other direction. The research firm Sakana AI, testing a system called The AI Scientist that conducts research autonomously, gave it a problem it could not finish within the experiment's set time limit; the system adjusted its own code to give itself more time. Creative, and also a demonstration that a model able to bypass a constraint changes what a control has to be designed to do.

Rebuilding confidence runs on two tracks. The emotional dimension asks whether people feel AI is aligned with their interests, and needs real policy rather than reassurance. The cognitive dimension asks whether a system acts reliably inside its guardrails, and needs dedicated domain and decision science teams continuously testing accuracy, predictability, consistency and explainability. Most Australian organisations have started the first and staffed almost none of the second.

Agents are on a path to becoming the primary users of enterprise systems

The capability curve for agents is steeper than most planning cycles assume. Tested on SWE-Bench Verified, a benchmark of real-world software issues drawn from GitHub, a leading agentic system achieved a 49% resolved rate. In 2023, agents scored under 5% on the same class of work. A jump of that size in two years makes long-dated architecture assumptions dangerous.

The cost side moved just as sharply. Amazon's generative AI assistant for software development saved the equivalent of 4,500 developer-years of work when updating an application to Java 17, and NVIDIA's Jensen Huang has stated the company has driven down the marginal cost of computing by 100,000x. When generating code stops being the scarce input, the constraint on an enterprise's digital footprint shifts from engineering capacity to architectural readiness, and to the ability of a business to decide what is worth building.

Executives are nonetheless pacing themselves. In Factor's executive survey, only 32% expect to enable AI agents to build workflow automations spanning their organisation within the next three years, with 68% placing it four or more years out; access to functions in third-party systems is further out again at 29% near term. Executives are similarly split on the consumer horizon, with 37% expecting the general public to use agents more frequently than apps or websites between 2025 and 2030, and 39% putting that shift in 2031-2035.

The gap between capability and enablement is where Australian modernisation debt shows up. Agents can only compose solutions across a business if data sources, functions and identity are actually reachable, which makes agentic ambition a digital-core question before it is an AI question. APIs, identity, data quality, cyber controls and legacy integration set the ceiling on how much useful autonomy can safely be deployed, a point developed in Factor's analysis of agentic platform strategy.

The consumer objection Australian brands plan around is smaller than they think

Only 13% of consumers overall have negative feelings about the use of generative AI for marketing and advertising. That figure comes from Factor's Consumer Paradoxes Survey of 12,215 consumers, fielded globally in October 2024 rather than in Australia specifically, so treat it as direction rather than a domestic reading. Even discounted, it sits a long way from the assumed backlash that shapes many Australian brand roadmaps.

Familiarity is what moves the number. Consumers who are familiar with generative AI feel more positive about its use in marketing and advertising (74%) than consumers who are less familiar with it (42%). The mechanism is exposure, not persuasion: people who have used the technology price it accurately, and people who have not are reacting to a category rather than an experience. That argues for sequencing AI into visible, low-stakes touchpoints early, so familiarity is built before the high-stakes interactions arrive.

A credit union example makes the same point at close range. In a survey of some of the Michigan State University Federal Credit Union's low- and medium-income members, 44% said they felt nervous about using AI technology, but of that group, 70% said they liked using Fran, the organisation's own chatbot. Anxiety attaches to the abstraction; satisfaction attaches to the specific thing.

Executives see the strategic prize clearly enough: 76% agree that conversational interactions using generative AI will become a way to gather relevant customer context. That is a different proposition from using AI as another channel to push messages, because it turns every interaction into an opportunity to learn interests, needs and even conversation style, with consent gathered in the moment. The risk is that foundation models are deliberately built to sound neutral, so enterprises that do not fine-tune for experience end up with the same face as their competitors, a trap examined in Factor's research on brand and AI personality and corroborated by Factor's Australian consumer research.

Where autonomy already shows a return, and why those cases share a shape

The clearest documented gain in the research is narrow and specific. At Wayfair, developers equipped with Gemini Code Assist set up environments 55% faster, with a 48% increase in unit test coverage, and 60% of the developers said the tool let them focus on more satisfying work. Speed, quality and engagement moved in the same direction at once, which is rare and worth understanding.

That combination is what makes a gain durable. Pure speed tends to be reabsorbed by the organisation as more work rather than better economics, and it invites the suspicion that the point was headcount. Speed accompanied by verifiable quality changes the function's cost curve permanently, because the output can be trusted without a second human pass. Factor's own research expects generative AI to drive productivity gains of 20% in companies leading in AI adoption, and the operative word is leading.

Drug discovery shows the same shape at a much larger scale. Insilico Medicine used generative AI to move from discovery to phase one trials in under 30 months, around half the time the process usually takes, using one model fine-tuned on omics and clinical data to identify targets and a generative chemistry engine built from 500 predictive and pre-trained models to propose compounds. In a later programme for idiopathic pulmonary fibrosis, a multimodal system trained in chemistry narrowed the field to 79 viable candidates, from which one was taken to trials.

Both cases pass the same three tests, which together make a workable screen for an Australian portfolio. The search space is bounded, so the system is not asked to invent the objective. The output is verifiable, so a human can confirm it cheaply. And a person still makes the consequential call. Use cases failing any of the three are the ones that stall in pilot.

The workforce loop decides how quickly any of this diffuses

Generative AI is a learning technology, and that changes the return profile. Conventional automation delivers a one-time benefit and risks leaving a disenchanted workforce behind it; generative AI improves the more it is used, so the more people use it, the better it gets. Break that loop and an organisation gets the one-time benefit and the disenchantment.

Access is where most organisations break it. Reports indicate 95% of employees find value in generative AI, yet only 47% of executives expect their organisations to make generative AI tools significantly or fully accessible to employees for automating tasks and workflows over the next three years. The plurality, 49%, say they will make them only partially accessible. Partial access is a defensible risk posture and an expensive one, because it caps the volume of real-world use the organisation learns from.

The skills position compounds it. In a 2024 report, nearly two-thirds of employers felt job candidates should have foundational knowledge of generative AI tools, while over half of recent college graduates felt their programmes did not prepare them adequately to use it. Higher education will respond, but matriculating students takes years, which leaves upskilling sitting squarely with Australian employers for the next several intakes.

Physical work is inside the same loop, and Australian executives should not read it as a knowledge-work story only. Asked how workers would leverage physical copilots over the next five to ten years, executives pointed to performing dull repetitive tasks (66%), assisting workers directly to complete a task such as holding or lifting (45%), allowing workers to perform tasks at a distance (39%), performing dangerous tasks (31%) and performing dirty tasks (15%). The top two answers are complementary rather than substitutional, which is a useful signal for how to frame the change internally.

What this means specifically for Australian organisations

Start with what is and is not Australian evidence. The executive findings come from a global survey of 4,021 C-level and director-level executives across 21 industries and 28 countries, fielded from October to December 2024; the consumer findings come from a global survey of 12,215 consumers in October 2024. Only the adoption figures are Australian, and they come from the ABS. Read the global numbers as direction and the ABS numbers as position.

On that basis, market structure is the first Australian variable that matters. Banking, telecommunications, retail and insurance here are concentrated, with a small number of large players serving most of the market. In concentrated markets, AI-mediated customer experience reaches parity quickly and then has to differentiate on something other than function, which is exactly the moment where a generic, neutral-sounding agent becomes a liability rather than an efficiency.

The second variable is physical. Mining, logistics, energy, defence, healthcare and advanced manufacturing make embodied AI unusually material for Australia, where labour constraints, distance and safety economics all strengthen the case for physical autonomy. Goldman Sachs found the global market for humanoid robots could reach $38 billion by 2035, and the shift toward multipurpose, redeployable machines changes the investment logic: a generalist robot that can be reassigned or returned makes trialling a new use case far less financially risky than a single-purpose installation ever was.

The third is readiness, and it is unglamorous. Australian technology leaders should treat agentic architecture as a modernisation issue, because APIs, identity, data quality, cyber controls and legacy-system integration set the ceiling on how much useful autonomy can safely be deployed. The Australian Government's National AI Plan is organised around capturing the opportunity, spreading the benefits and keeping Australians safe, which gives boards a policy frame but not an architecture. That part is the organisation's own work.

The fourth is time. Less than 1% of today's global internet market capitalisation was founded in the first two years after Netscape Navigator generalised the internet, a reminder that the durable winners of a platform shift are mostly established before the shift becomes legible.

What Australian executives should do next

Benchmark against your sector rather than the national figure. With national adoption at 12% and leading sectors at 38%, 24% and 24%, an organisation that measures itself against the country will conclude it is on track when it is a year or more behind the firms it actually competes with. Set the target against the sector line and re-baseline it annually.

Fix the shadow-adoption problem before it becomes a control problem. More than half of workers using AI are reluctant to admit it, so an amnesty plus a clear, current usage policy will surface more real risk in a fortnight than a further round of tooling restrictions will in a year. It also converts hidden use into the organisational learning the loop depends on.

Prioritise use cases with a bounded search space and a verifiable output. The Wayfair result is the template: environments set up 55% faster alongside a 48% increase in unit test coverage, with 60% of developers reporting more satisfying work. Verifiability is what allowed the speed to stick, and it is the single best predictor of whether a pilot survives contact with production.

Widen access deliberately and instrument what happens. Only 47% of executives expect to make these tools significantly or fully accessible within three years, which means most organisations are choosing to learn slowly. Pick the functions where partial access is genuinely warranted, open the rest, and measure the use so the decision can be revisited with evidence rather than instinct.

Fund cognitive trust as a capability, not a policy. Autonomy is capped by confidence that a system will act reliably inside its guardrails, and that confidence needs a standing team testing accuracy, predictability, consistency and explainability, plus monitoring of what autonomous systems access and who directs them. Factor's research programme tracks these shifts through the Australian executive community; explore the full research library or join a Factor day event to compare positions with peers making the same calls.