Atlassian reports that 6% of executives are certain they can cite clear examples of organizational ROI; ADP reports AI usage and worker sentiment. The observed gap is consistent with measurement limitations and revenue realization delays, but it alone does not prove structural inadequacy.

The Essential Points

  • In Australia, 9 out of 10 companies report zero measurable impact on productivity despite widespread AI adoption, according to ADP People at Work 2026 and Atlassian State of Teams 2026.
  • Real AI gains focus on quality, speed, and error reduction: three dimensions that accounting systems built to capture revenues cannot measure.
  • This problem is not Australian: NBER Working Paper 34984 documents a similar gap internationally between adoption and aggregate productivity gains.
  • Without adapted measurement frameworks, the distribution of gains between labor, capital, and consumers remains opaque, which weakens any evidence-based public policy.
  • Productivity econometrics has already experienced this type of lag: computerization in the 1980s-1990s took more than a decade to appear in statistics.

Solow’s Paradox Strikes Again

Robert Solow said it in 1987: you see computers everywhere except in productivity statistics. The phrase was provocative. American productivity accelerated from the mid-1990s onward; consensus on the contribution of information technologies solidified in the early 2000s.

Generative AI presents certain traits comparable to early episodes of information technology diffusion, but it is too early to conclude that it reproduces their exact trajectory. The 2026 Australian surveys show weekly AI adoption. Copilots, writing assistants, synthesis tools, support ticket automation: the uses are real, documented, daily. The NBER paper observes positive but modest productivity gains and a gap between perceived gains and implicit gains; it does not provide a “9 out of 10” result. Working Paper 34984 documents, in a sample primarily of American financial executives, a gap between perceived productivity gains and implicit gains measured from revenues; Working Paper 34836 provides a comparison across four countries.

Two readings are possible. The first: the observed AI gains remain heterogeneous or limited depending on contexts, and management enthusiasm may exceed measured gains. The second: gains are observed in certain contexts, but available estimates remain heterogeneous and declared gains exceed implicit measured gains. The Australian data show frequent use and contrasting perceptions, but do not make it possible to determine whether real gains exist without being measured. When 89% of executives claim their team works faster, 94% are not certain they have clear examples of organizational ROI; this does not mean they observe no trace of AI’s effect in their accounts.

It is in the spreadsheets.

The Limitations of Current Accounting Systems

Modern accounting frameworks were built on an industrial model: measuring units produced, revenues received, hours worked. This architecture works well for productivity gains that translate into more products or services sold per hour worked. It works poorly for gains that manifest differently.

AI uses can improve several operational outcomes, some of which are imperfectly isolated in national accounts or short-term revenue measures. First, execution speed: an analyst producing a report in two hours instead of six does not necessarily generate more revenue; they address more topics or free up time for higher-value tasks. Next, quality: fewer errors in a legal report, more accurate customer responses, code with fewer bugs. These improvements reduce invisible costs—correction, litigation, and support costs—that only appear in the accounts when they occur. Finally, reduction of cognitive load: less saturated teams make better decisions, which is hard to trace in a quarterly balance sheet.

Some operational gains may not immediately create a new revenue line, but primary sources also identify channels of innovation and demand associated with revenue gains. Some operational gains do not immediately translate into revenue, but can be tracked by operational or accounting indicators depending on the information systems used. The same measurement difficulty applies to other forms of work organization: the benefits of remote work also long resisted traditional metrics before finer longitudinal studies documented them.

Rapid Adoption Against a Backdrop of Untransformed Organization

The Australian adoption rate impresses. But Atlassian data reveal an important nuance: AI is widely used individually, while its integration into workflows remains limited; the source does not directly measure process transformation. A large share of knowledge workers use AI, while a minority report integrating it into workflows.

This distinction is central. The econometrics of technical progress suggests that productivity gains from a general-purpose technology materialize when accompanied by complementary reorganizations, new roles, new processes, new task allocation. Electricity did not make factories more productive by replacing steam power machine by machine: it did so when factory architects understood that workshops could now be arranged differently, freed from the constraint of centralized mechanical transmission.

Generative AI is probably at the pre-reorganization stage. 44% of Australian workers report using AI at least several times a week; the survey does not allow us to affirm that they produce more, faster, with unchanged organizational tools. Individual gains exist. They do not necessarily translate into organizational gains; workflow integration, coordination, and complementary investments constitute possible explanations among others. This gap between individual adoption and collective transformation is documented in other technological contexts, and it has its own temporal dynamics: studies on personal computing diffusion in 1980s companies show organizational absorption cycles measured in years, sometimes in decades.

Who Captures What, and How to Know

The measurement problem has a political dimension in the strict sense. Without more precise measures, it is difficult to rigorously estimate the distribution of AI gains between capital, labor, and consumers.

The exact distribution of gains attributable to AI remains difficult to estimate and insufficiently measured, rather than indeterminable.

The question of governance of digital infrastructures encounters here a concrete dimension: poor AI measurement limits documentation of its economic effects; its precise consequences on wage negotiations, working time, or taxation remain a hypothesis to support. Poorly measured productivity is harder to attribute empirically; the modalities of its distribution are an institutional and negotiation question, not a consequence directly demonstrated here.

Economists have tools to approach these questions: field studies in labor microeconomics, longitudinal panels on wages and margins, sectoral comparisons between adopting and non-adopting firms. But these tools require time, company data rarely available, and methodologies that will not be stabilized before several years of research. In the meantime, decisions are made without reliable quantitative guidance.

Daron Acemoglu and Simon Johnson posed the question directly in their recent work on technology and power: who decides which functions to automate, and who benefits from that automation. The opacity of AI gains makes this arbitration even harder to establish. Without measurement, there is no evidence, and without evidence, political arbitration happens by default rather than by deliberate choice.

Horizon 2030: Prospects and Obstacles

The history of general-purpose technologies makes plausible a gap between AI adoption and appearance of measured gains, without guaranteeing that an aggregate effect will materialize or specifying its magnitude. NBER Working Paper 34984 documents, in a sample primarily of American financial executives, a gap between perceived productivity gains and implicit gains measured from revenues.

A first scenario, which one might call silent absorption, would see AI gains accumulate over five to ten years in improved margins, stable workforce levels, and increased competition between firms. National statistics would eventually capture an improvement in multifactor productivity around 2030-2032, with a lag comparable to that observed after computerization. In this scenario, the distribution of gains would have already occurred, essentially toward capital, before measurement tools would allow it to be documented or corrected.

A second scenario, more favorable to sharing, assumes that companies, social partners, and statisticians develop new indicators earlier capable of capturing qualitative gains. Service quality metrics, error reduction, turnaround times, integrated into sectoral dashboards, would make gains visible where they occur and allow negotiation of their distribution before they dissolve into margins. Several initiatives point in this direction: Diane Coyle’s work on measuring progress in the digital age, OECD reflections on work well-being indicators, or certain technology companies’ experiments publishing internal impact metrics beyond mere revenue figures.

The third scenario, less optimistic, is that of retrospective crisis. AI gains remain invisible too long, income polarization intensifies without anyone being able to document its precise mechanism, and ultimately a social or political shock forces reassessment: mandatory impact audits, regulation on sharing productivity gains, or even specific taxation on automation benefits. This scenario is not inevitable, but it becomes more likely as adoption accelerates while measurement institutions lag.

Signals to watch to distinguish these trajectories exist. The first is the evolution of real wages in sectors with high AI adoption: if productivity gains are shared, nominal wages adjusted for productivity should advance. The second is margin dynamics: rising margins in high-adoption sectors without comparable wage progression would point toward capital capture. The third is institutional: national statisticians, Australian or otherwise, beginning to integrate qualitative indicators into their satellite accounts would be a strong signal of adaptation underway.

Paths That Allow Not to Wait

Australia presents a useful particularity: its labor market is relatively transparent, with accessible sectoral data and structured social dialogue in several industries. This context is favorable to experiments on qualitative productivity measurement, if stakeholders seize the opportunity.

Several concrete directions emerge from current debates. Companies adopting AI could systematize internal impact assessments measuring not only revenue but processing times, error rates, subjective workload. This data does not replace national accounting but creates a corpus enabling comparative studies. Australian trade unions, active in several service sectors, have the tools to negotiate access to this data in discussions on working conditions.

On the public statistics side, the Australian Bureau of Statistics has a tradition of methodological innovation. Its teams have already developed indicators on the digital economy and intangible assets. A research program on qualitative AI gains, in partnership with universities and a few large willing companies, would produce in three to five years data usable for public policy.

The difficulty is first political: measuring precisely where gains go creates transparency, and transparency redistributes power in negotiations between employers, employees, and the state. Some actors have an interest in gains remaining obscured. Others, betting on long-term legitimacy of AI adoption, have an interest in making them visible. The question of regulating technological infrastructures touches exactly this point: the stakes concern who has the information to arbitrate, not just whether to build or regulate.

Solow’s paradox was partly attenuated for information technology diffusion in the 1990s, but the mechanisms of lag and measurement remain debated. It took time, massive organizational investments, and improved statistical tools. AI could experience diffusion and measurement lags analogous to those of IT, but its adoption and accounting modalities differ enough to prevent any certain extrapolation. The open question is whether institutions will be able to anticipate it, or whether they will wait, as usual, until transformation is already accomplished before beginning to measure it.

Sources

  1. NBER Working Paper 34984, https://www.nber.org/papers/w34984
  2. ADP People at Work 2026, ADP Research Institute (annual report, no stable URL guaranteed)
  3. Atlassian State of Teams 2026, Atlassian Corporation (annual report, no stable URL guaranteed)
  4. CoterieLabs analysis, synthesis analysis of Australian data on AI adoption in business (no stable URL guaranteed)