Strategy to Outcomes (S2O) thought-leadership article - Productivity and AI Predictive Maintenance Use Case

Australia’s productivity debate has renewed attention on business investment and capital deepening. For organisations, the harder question begins after the investment decision. How do we know that new technology actually improves the productive system?

Predictive maintenance provides a useful AI use case because the path from insight to operating outcome can be made visible, tested and measured.

Alan Kohler’s recent ABC commentary on Australian productivity highlights a familiar concern. Weak business investment and slow capital deepening have contributed to poor productivity performance. The broad economic logic is straightforward. Better capital, technology and operating systems can allow each worker to support more valuable output.

But capital expenditure and productivity are not the same thing. Labour productivity can rise because more capital is placed behind each worker, while the broader question is whether labour and capital together are being used more effectively.

That distinction is especially relevant to AI. AI can be an important form of capital deepening, but replacing labour with AI does not itself prove that productivity improved.

The organisational question is not simply whether we invested in AI. It is whether the investment changed the productive system.

Predictive maintenance is one of the more credible enterprise AI use cases because the underlying problem is genuinely information-intensive and probabilistic.

Maintenance teams may need to interpret vibration and temperature trends, equipment cycles, work-order histories, breakdown reports, strip-and-assess findings, manufacturer guidance, operator observations and condition-monitoring data. The information can be incomplete, inconsistently described and distributed across multiple systems.

AI can assist by finding patterns across those sources, identifying abnormal conditions, ranking likely causes and presenting an engineer with a more focused set of diagnostic candidates. That is useful. But it is still only a capability.

AI model deployed  ≠  productivity improved

The productive value only appears if the new capability changes operating behaviour and ultimately produces more valuable output, or the same output with fewer total productive resources.

The productivity claim sits behind the arrow. A business case can easily compress the proposition into a simple statement:

Predictive AI  →  Productivity

Almost all of the important reasoning is hidden inside that arrow.

AI-supported diagnosis  -> Earlier decision  -> Earlier intervention  -> Less emergent maintenance  -> Greater productive availability  -> Higher throughput  -> Greater economic output.

Each step is conditional. A correct diagnosis may still produce no benefit if maintenance capacity is unavailable, the required part is missing, production cannot release the asset, or the asset is not constraining throughput. That is why the productive outcome should not be attributed to the technology alone. It depends on the operating system around the technology.

Why the P&L is not enough

Most organisations will first see an AI investment through their financial reporting. Software expenditure rises. Cloud consumption appears. Consulting and integration costs may increase. Later, payroll may fall or labour may be redeployed. The P&L records these changes, but it does not retain the causal relationship between them.

Revenue  –  Cost  =  Profit

Productivity asks a different question.

Productivity  =  Valuable output  /  Resources consumed

Suppose AI reduces diagnostic labour by 25 per cent. That may increase measured labour productivity. But if the new operating model also adds AI infrastructure, specialist labour, verification effort, false-positive investigations and rework, the productive system may not have improved at all.

A labour saving is an input change. A productivity gain requires evidence that the overall system is producing more value from the resources consumed. The P&L tells management what happened financially. Productivity asks whether the productive system became better.

 Measurement must test the productivity hypothesis

A single KPI is rarely enough. If asset availability rises from 92 to 95 per cent, the result is useful but incomplete. Management still needs to know whether AI contributed to the improvement and whether the extra availability created productive value. A credible measurement approach needs evidence at different points in the causal chain.

At the front of the chain, the organisation needs to know whether the AI-generated insight changed a maintenance decision. In the middle, it needs to know whether maintenance behaviour and asset performance changed. At the end, it needs to know whether throughput, service output or another economically meaningful result improved.

The aim is not to create a larger dashboard. It is to retain enough evidence to challenge the original assumptions. The measurement question is not only 'did the KPI improve?' but 'did the operating mechanism behave as we expected, and did that change produce the outcome?'

The human remains part of the productive system

Predictive maintenance also clarifies the role of the human in AI-enabled work. The underlying information problem is probabilistic. A model may rank likely causes and compress the diagnostic search space, but maintenance interventions have physical, safety, production and financial consequences.

The relevant productivity question is not necessarily how many engineers can be removed. It is whether the human-plus-AI operating system makes better decisions, earlier, with less avoidable effort and a better productive result.

If the AI generates a large volume of low-quality recommendations that require extensive checking or create unnecessary work orders, apparent automation can consume productive capacity elsewhere.

Delayed feedback is part of the evidence

The effect of a maintenance intervention is rarely visible on the same clock as financial reporting. An abnormal condition may be detected today. An intervention may occur next week. A failure that would otherwise have happened may be avoided months later. Changes in emergent maintenance, asset availability and throughput may take longer again to become credible. That delay is not simply a measurement inconvenience. It is part of the evidence.

The organisation needs to retain what it expected to happen and compare that expectation with what actually happened over time. Otherwise, Finance retains the costs, IT retains adoption data, Maintenance retains work-order measures, Operations retains throughput, and the original productivity hypothesis disappears.

The counterfactual still matters

Even a well-measured improvement does not automatically prove that AI caused it. Other conditions may have changed at the same time. The organisation needs some form of counterfactual question. What would probably have happened without the intervention?

This may be approached through phased rollout, comparable assets, controlled pilots, historical baselines or matched operating periods. The purpose is not to claim perfect causal certainty, but to make the productivity explanation progressively stronger.

Where S2O fits

S2O addresses the organisational gap between an investment decision and the eventual outcome by making the causal business logic explicit before detailed technology configuration.

It establishes the relevant strategic outcome, the value-chain context, the operating changes believed necessary to produce that outcome, and the evidence needed to test whether those changes actually occurred.

In a predictive-maintenance use case, that means connecting the AI capability to the client-specific path from maintenance behaviour to productive availability, throughput and economic outcome.

The technology platform and implementation partner can then expand that design into the detailed solution. The role of S2O is not to replace the application or the partner. It is to make the business reasoning behind the implementation explicit and persistent.

S2O makes the productivity hypothesis explicit. The technology and implementation partner make it operational. The evidence determines whether it held.

From AI investment to productive AI

Australia’s productivity debate is a useful reminder that capital investment matters. But for an organisation, the critical question begins after the capital has been allocated.

Predictive maintenance demonstrates why. AI may improve diagnosis. Better diagnosis may enable earlier intervention. Earlier intervention may reduce emergent work. Reduced emergent work may improve productive availability. Additional productive availability may increase throughput. But every arrow is conditional.

That is why productivity cannot be inferred from technology adoption, labour displacement or a single operational KPI. It requires a persistent view of the operating logic, enough measurement to challenge that logic, and enough time for delayed evidence to accumulate.

The P&L tells us what happened financially. The causal design tells us what we believed would happen. The evidence tells us whether that belief held. Productivity tells us whether the system actually became better.

For organisations investing in AI, that may be the more important productivity question. Not how much work the technology can replace, but whether the combined human, technology and operating system produces more value from the resources consumed.

Selected sources

Alan Kohler / ABC News. The main culprit of Australia’s productivity crisis, 13 July 2026. Source link

Australian Bureau of Statistics. Estimates of Industry Multifactor Productivity, 2024–25. Source link

Productivity Commission. Annual productivity bulletin 2026. Source link

Productivity Commission. What is productivity? Labour productivity, capital deepening and multifactor productivity. Source link