AI adoption is accelerating; operating discipline will decide who captures the value

The adoption curve is real, but the headline is not the outcome

The latest Australian Bureau of Statistics release shows a decisive change in business use of artificial intelligence. In 2024–25, 12% of Australian businesses reported using AI, compared with 1% in 2021–22. Adoption was materially higher among innovation-active businesses at 20%, versus 6% among businesses that were not innovation-active. The same data show a pronounced industry spread: information media and telecommunications recorded 38% use, professional services and financial services each 24%, mining 18%, manufacturing and retail 9%, and transport, postal and warehousing just 1% [1].

International evidence points in the same direction. Across OECD countries with available data, 20.2% of firms reported using AI in 2025, up from 14.2% in 2024 and 8.7% in 2023. Yet diffusion remains uneven: 52% of large firms used AI compared with 17.4% of small firms, while uptake was highest in ICT and professional and scientific services [2]. These are adoption measures, not audited productivity results. They establish that AI is spreading; they do not establish that every deployment is creating durable economic value.

Usage, time saved and realised productivity are different measures

A July 2026 IMF working paper adds a useful but carefully bounded estimate. Using observed AI usage data collected between January 2025 and February 2026, the authors value the time currently saved by AI at a labour-cost equivalent of US$2.7 trillion annually, or 3.4% of global GDP [3]. The paper explicitly describes this as an indicative measure of the labour cost of time saved. It is not a forecast of additional GDP, cash earnings or employment effects, and the paper represents research in progress rather than an official IMF policy position.

That distinction matters commercially. Time released by automation becomes productive capacity only when the organisation has somewhere valuable to redeploy it. A faster first draft, classification or search task can lower effort, but the benefit may be absorbed by additional checking, weak data, duplicated systems or unchanged decision rights. Our inference from the evidence is that the next competitive gap will be less about whether staff can access AI and more about whether management can convert isolated time savings into faster throughput, better decisions, lower error rates or improved service.

The practical bottleneck is workflow redesign

The ABS data provide an important adjacent signal. Only 7% of businesses reported measuring the contribution of digital activities to overall performance in 2024–25. Separately, insufficient staff skills and capabilities and uncertainty around ICT costs and benefits were the two most commonly reported barriers to ICT use [1]. These measures cover digital activity more broadly, not AI alone, but they challenge a common assumption: rapid tool adoption does not imply that most organisations have a mature value-measurement system.

The Australian Government’s National AI Centre launched AI.gov.au in May 2026 with guidance organised around identifying value, planning adoption, managing change, understanding risk and building capability [4]. The sequence is commercially sensible. A useful implementation starts with a defined workflow and baseline, not with a general instruction to “use AI”. It identifies the accountable owner, the data allowed into the system, the human review point, the acceptable error rate and the metric that will show whether the process improved.

• Select a recurring, bounded workflow with enough volume for improvement to be measurable.
• Record the current cycle time, labour effort, error or rework rate, and service outcome before introducing AI.
• Define data, security and escalation boundaries before moving from a pilot to operational use.
• Measure the full process after deployment, including review and correction time rather than model speed alone.

Uneven diffusion creates both opportunity and execution risk

The industry and firm-size gaps are not evidence that every lagging sector should imitate technology firms. Different processes carry different data, safety, regulatory and customer risks. They do, however, suggest that organisations with repeatable information-heavy workflows may still have substantial room to improve. Construction and accommodation were among the fastest-growing adoption sectors in the OECD data during 2025, despite starting from lower bases [2]. This is consistent with diffusion moving beyond early technology-intensive users.

The strongest alternative explanation is that survey definitions and respondent interpretation are inflating apparent momentum. The ABS itself notes that its question records whether a technology was used and is not designed to measure intensity or extent of use [1]. That caveat limits comparisons across datasets and prevents a clean conclusion about how deeply AI is embedded in operations. It also reinforces the central point: management should treat adoption statistics as evidence of direction, not as a substitute for its own operating and financial measures.

The 2026 decision is where to build repeatable capability

For operators, the practical priority is to move from scattered experimentation to a small portfolio of governed workflows. The best candidates are frequent enough to measure, sufficiently structured to verify and valuable enough that faster completion changes a commercial outcome. Examples include document intake, exception triage, internal knowledge retrieval, customer-service preparation and recurring reporting, with human authority retained where judgement or external consequence is material.

The evidence supports a disciplined conclusion rather than a universal forecast. AI adoption is accelerating across Australia and other advanced economies, and usage data indicate potentially significant time savings [1][2][3]. What remains uncertain is how much of that time will translate into realised productivity, how benefits will be distributed across workers and firms, and which deployments will survive security, quality and change-management tests. In 2026, durable advantage is more likely to come from measured workflow redesign than from tool access alone.