Microsoft executive vice president for Copilot Charles Lamanna has put Australia’s weak productivity performance down partly to employers’ reluctance to take risks with artificial intelligence. The useful part of that diagnosis is not the slogan about “institutional inertia”; it is that Australian organisations have largely deployed AI as a personal assistant while avoiding the workflow, data and accountability changes required for it to alter measured output.
Human Resources Director reports that Lamanna made the comments to The Australian, arguing that companies are being too hesitant because AI makes mistakes and every technology carries risk. The quote has not been published in a Microsoft announcement or transcript, so the precise context and the event at which it was delivered remain unclear. But the broad proposition is supported by the Reserve Bank of Australia’s own survey work: AI is present in many firms, yet is usually shallow, piecemeal and employee-led.
Australia’s productivity problem is also real enough to make the debate more than an argument about Copilot adoption. The Productivity Commission said in June that labour productivity fell 0.6% in the March 2026 quarter and rose only 0.3% over the preceding year. More people are working more hours, but output is not keeping pace.
The uncomfortable implication for Microsoft 365 administrators and IT leaders is straightforward: buying licences, enabling Copilot and measuring weekly usage does not constitute a productivity program. It may provide useful individual time savings. It does not, on its own, change the way an organisation processes claims, responds to customers, produces reports, manages cases, develops software or makes decisions.
The Reserve Bank’s November 2025 review of technology investment found that roughly two-thirds of surveyed firms had adopted some form of AI. That sounds like a broad deployment story until the central detail is considered: almost 40% of those firms reported only minimal use, commonly off-the-shelf tools such as Microsoft Copilot or ChatGPT for summarising emails and researching information.
Only around 30% of the firms surveyed had moved into more substantive adoption, such as using AI to support forecasting, inventory management, fraud detection or multiple business lines. Even there, the RBA cautioned that the group it spoke to was skewed toward larger, more established businesses with greater resources. Small firms are less likely to have adopted AI at all.
The National AI Centre’s most recent SME tracking paints a similar picture from another angle. It found 43% of Australian small and medium businesses reported some AI adoption across the December 2025 to February 2026 period, with content generation and data analytics the most common uses among adopters. Those are sensible entry points: they are relatively easy to test, have a limited operational blast radius and often sit outside critical systems.
They also tend to produce benefits that are hard to convert into economy-wide productivity. A worker who saves 20 minutes summarising a meeting has more capacity, but the productivity statistic does not improve unless the organisation redesigns work so that capacity creates more output, better service, lower cost or faster revenue. If the recovered time is absorbed by more email, more review cycles or extra administrative work, the firm has bought convenience rather than transformation.
This is the gap in Lamanna’s argument. He is right that risk aversion can freeze projects. But the evidence says the bigger problem is not simply that employers fear AI errors. It is that many have stopped at tools that leave the underlying process untouched.
The danger begins when the pilot becomes the product strategy.
Australia’s government-run Microsoft 365 Copilot trial demonstrated why. The Digital Transformation Agency found that only one-third of trial participants used Copilot daily, with summarising information and rewriting content accounting for most use. Word and Teams were favoured, while access barriers hindered use in Outlook. The findings were not a verdict that Copilot cannot help; they were a warning that availability and broad entitlement are weak measures of success.
Treasury’s separate evaluation reached a similarly tempered conclusion. Almost all participants used Copilot at some point during the trial, but only 22% said they used it four or five times each week. Nearly two-thirds found it useful for basic administrative tasks and work processes. That is a credible outcome for a generic assistant deployed into a high-governance environment, but it is far short of evidence that an agency has redesigned a core function around AI.
The government’s own recommendations point toward the missing work: agencies were told to analyse workflows by job family, train users around specific tasks, configure systems and permissions for safe use, establish clear accountability, and identify where AI could improve a process rather than simply accelerate a document. Those recommendations should apply equally to corporate Microsoft tenants.
A pilot built around “give a thousand employees Copilot and see what happens” will find enthusiastic individual users, sporadic use and plenty of anecdotes. It will not reliably identify whether accounts-payable exceptions can be triaged faster, whether field-service notes can be turned into usable maintenance records, whether a sales-support process can remove duplicate manual entry, or whether a help desk can resolve a defined class of requests without creating a new quality-control bottleneck.
Those risks are concrete. A generative AI assistant can surface information a user was already technically permitted to access but had never been able to discover efficiently. It can write plausible but incorrect content. It can turn a weakly governed SharePoint library into a source of fast, confident misinformation. An agent connected to business systems can make a poor recommendation at machine speed, or take an authorised action in circumstances its designers did not adequately anticipate.
None of those hazards means organisations should prohibit meaningful AI use. They mean leaders should distinguish controlled risk-taking from simply enabling a powerful feature across a tenant and hoping good prompts will supply the controls.
The National AI Centre’s implementation guidance makes this distinction explicit. Higher-risk and more complex uses require governance, testing, monitoring, human oversight and clear accountability across the supply chain. That is not institutional inertia. It is the operational foundation that lets an organisation safely move from drafting emails to using AI within consequential processes.
For a Windows and Microsoft 365 environment, the first practical check is usually not model selection. It is identity and information hygiene: who can access which SharePoint sites, Teams channels, mailboxes and records; whether sensitivity labels and retention rules match real-world handling; whether old permissions are still justified; and which systems are safe to connect to an agent. Copilot frequently exposes data-governance debt that existed before the AI rollout.
Microsoft’s own Australia and New Zealand research has acknowledged that point. Its survey of larger organisations found that 96% of respondents had encountered data-security or access challenges during generative AI adoption. The study was commissioned by Microsoft and should be read as vendor research, but the finding aligns with the government trial’s configuration and permissions concerns. The prerequisite for broader use is often less glamorous than an AI strategy deck: fix the data estate.
That matters because productivity is an outcome of the whole operating model. A Copilot rollout owned solely by IT will naturally focus on licences, security, readiness, enablement and service health. Those are necessary responsibilities, but none answers the crucial business question: which process owner is accountable for turning AI-assisted work into a measurable result?
The better model is narrower and harder. Select a process with a known pain point, measurable baseline and accountable executive owner. Define the permitted data, the human review step, the exception path, success criteria and the moment when the old process will be retired rather than run alongside the new one. Then measure cycle time, error rate, rework, customer response time, output per employee or cost per transaction — not merely prompts submitted or monthly active users.
That approach also produces a defensible answer when leaders ask whether to scale. If the result is only faster first drafts, keep the deployment in the productivity-tool category. If it reduces handoffs, removes duplicated work or lets teams complete a process at greater volume and quality, it has earned expansion into a business-system change.
Australia does not need employers to pretend that AI is infallible. It needs them to stop treating cautious pilots and tenant-wide licences as the finish line. The Productivity Commission’s latest figures show there is little room for mistaking access to a tool for a lift in output: March 2026 delivered another quarter in which hours worked grew faster than production.
Australia’s productivity problem is also real enough to make the debate more than an argument about Copilot adoption. The Productivity Commission said in June that labour productivity fell 0.6% in the March 2026 quarter and rose only 0.3% over the preceding year. More people are working more hours, but output is not keeping pace.
The uncomfortable implication for Microsoft 365 administrators and IT leaders is straightforward: buying licences, enabling Copilot and measuring weekly usage does not constitute a productivity program. It may provide useful individual time savings. It does not, on its own, change the way an organisation processes claims, responds to customers, produces reports, manages cases, develops software or makes decisions.
Australia has adoption, but little depth
The Reserve Bank’s November 2025 review of technology investment found that roughly two-thirds of surveyed firms had adopted some form of AI. That sounds like a broad deployment story until the central detail is considered: almost 40% of those firms reported only minimal use, commonly off-the-shelf tools such as Microsoft Copilot or ChatGPT for summarising emails and researching information.Only around 30% of the firms surveyed had moved into more substantive adoption, such as using AI to support forecasting, inventory management, fraud detection or multiple business lines. Even there, the RBA cautioned that the group it spoke to was skewed toward larger, more established businesses with greater resources. Small firms are less likely to have adopted AI at all.
The National AI Centre’s most recent SME tracking paints a similar picture from another angle. It found 43% of Australian small and medium businesses reported some AI adoption across the December 2025 to February 2026 period, with content generation and data analytics the most common uses among adopters. Those are sensible entry points: they are relatively easy to test, have a limited operational blast radius and often sit outside critical systems.
They also tend to produce benefits that are hard to convert into economy-wide productivity. A worker who saves 20 minutes summarising a meeting has more capacity, but the productivity statistic does not improve unless the organisation redesigns work so that capacity creates more output, better service, lower cost or faster revenue. If the recovered time is absorbed by more email, more review cycles or extra administrative work, the firm has bought convenience rather than transformation.
This is the gap in Lamanna’s argument. He is right that risk aversion can freeze projects. But the evidence says the bigger problem is not simply that employers fear AI errors. It is that many have stopped at tools that leave the underlying process untouched.
“Safe” pilots can become a permanent holding pattern
For IT departments, a low-risk pilot is usually the rational first move. Deploy Microsoft 365 Copilot to a controlled group, validate tenant configuration, check data permissions, establish acceptable-use rules, train users, watch for oversharing and learn which applications receive genuine use. That is responsible practice, especially where SharePoint permissions and old collaboration sites have accumulated years of poorly governed content.The danger begins when the pilot becomes the product strategy.
Australia’s government-run Microsoft 365 Copilot trial demonstrated why. The Digital Transformation Agency found that only one-third of trial participants used Copilot daily, with summarising information and rewriting content accounting for most use. Word and Teams were favoured, while access barriers hindered use in Outlook. The findings were not a verdict that Copilot cannot help; they were a warning that availability and broad entitlement are weak measures of success.
Treasury’s separate evaluation reached a similarly tempered conclusion. Almost all participants used Copilot at some point during the trial, but only 22% said they used it four or five times each week. Nearly two-thirds found it useful for basic administrative tasks and work processes. That is a credible outcome for a generic assistant deployed into a high-governance environment, but it is far short of evidence that an agency has redesigned a core function around AI.
The government’s own recommendations point toward the missing work: agencies were told to analyse workflows by job family, train users around specific tasks, configure systems and permissions for safe use, establish clear accountability, and identify where AI could improve a process rather than simply accelerate a document. Those recommendations should apply equally to corporate Microsoft tenants.
A pilot built around “give a thousand employees Copilot and see what happens” will find enthusiastic individual users, sporadic use and plenty of anecdotes. It will not reliably identify whether accounts-payable exceptions can be triaged faster, whether field-service notes can be turned into usable maintenance records, whether a sales-support process can remove duplicate manual entry, or whether a help desk can resolve a defined class of requests without creating a new quality-control bottleneck.
The risk is operational, not abstract
Lamanna’s observation that AI is imperfect should not be waved away as a sales executive’s invitation to move faster. Microsoft is financially interested in enterprises expanding Copilot use, and “take more risk” is an unusually convenient message from a supplier selling AI subscriptions and agents. The company’s argument therefore needs to be tested against the risks that administrators and compliance teams actually carry.Those risks are concrete. A generative AI assistant can surface information a user was already technically permitted to access but had never been able to discover efficiently. It can write plausible but incorrect content. It can turn a weakly governed SharePoint library into a source of fast, confident misinformation. An agent connected to business systems can make a poor recommendation at machine speed, or take an authorised action in circumstances its designers did not adequately anticipate.
None of those hazards means organisations should prohibit meaningful AI use. They mean leaders should distinguish controlled risk-taking from simply enabling a powerful feature across a tenant and hoping good prompts will supply the controls.
The National AI Centre’s implementation guidance makes this distinction explicit. Higher-risk and more complex uses require governance, testing, monitoring, human oversight and clear accountability across the supply chain. That is not institutional inertia. It is the operational foundation that lets an organisation safely move from drafting emails to using AI within consequential processes.
For a Windows and Microsoft 365 environment, the first practical check is usually not model selection. It is identity and information hygiene: who can access which SharePoint sites, Teams channels, mailboxes and records; whether sensitivity labels and retention rules match real-world handling; whether old permissions are still justified; and which systems are safe to connect to an agent. Copilot frequently exposes data-governance debt that existed before the AI rollout.
Microsoft’s own Australia and New Zealand research has acknowledged that point. Its survey of larger organisations found that 96% of respondents had encountered data-security or access challenges during generative AI adoption. The study was commissioned by Microsoft and should be read as vendor research, but the finding aligns with the government trial’s configuration and permissions concerns. The prerequisite for broader use is often less glamorous than an AI strategy deck: fix the data estate.
Productivity requires a process owner
The RBA’s reporting also challenges the idea that technology alone explains Australia’s productivity slowdown. Firms identified skills shortages, uncertain use cases and return on investment, legacy-system integration, cost, digital readiness and regulatory uncertainty as barriers. It noted that technology was not the only near-term obstacle to better productivity.That matters because productivity is an outcome of the whole operating model. A Copilot rollout owned solely by IT will naturally focus on licences, security, readiness, enablement and service health. Those are necessary responsibilities, but none answers the crucial business question: which process owner is accountable for turning AI-assisted work into a measurable result?
The better model is narrower and harder. Select a process with a known pain point, measurable baseline and accountable executive owner. Define the permitted data, the human review step, the exception path, success criteria and the moment when the old process will be retired rather than run alongside the new one. Then measure cycle time, error rate, rework, customer response time, output per employee or cost per transaction — not merely prompts submitted or monthly active users.
That approach also produces a defensible answer when leaders ask whether to scale. If the result is only faster first drafts, keep the deployment in the productivity-tool category. If it reduces handoffs, removes duplicated work or lets teams complete a process at greater volume and quality, it has earned expansion into a business-system change.
Australia does not need employers to pretend that AI is infallible. It needs them to stop treating cautious pilots and tenant-wide licences as the finish line. The Productivity Commission’s latest figures show there is little room for mistaking access to a tool for a lift in output: March 2026 delivered another quarter in which hours worked grew faster than production.
References
- Primary source: hcamag.com
Published: 2026-08-03T05:14:44+00:00
Lack of AI risk-taking blamed for lagging Australian productivity | Human Resources Director
Microsoft executive points to 'institutional inertia' for lagging productivity despite AI adoption
www.hcamag.com
- Related coverage: rba.gov.au
In Depth – Drivers and Implications of Lower Productivity Growth | Statement on Monetary Policy – August 2025 | RBA
In Depth – Drivers and Implications of Lower Productivity Growth | Statement on Monetary Policy – August 2025www.rba.gov.au - Related coverage: pwc.com.au
Inaction may be the biggest AI risk
Why AI-driven innovation starts with bold action, steady momentum, and shifting habits: PwC Australia Assurance partner, David Mawww.pwc.com.au
- Related coverage: abc.net.au
Treasurer Jim Chalmers says AI central to raising productivity and lowering interest rates - ABC News
The Treasurer has given an insight into his April meeting with Anthropic, arguing that Australia embracing AI is crucial for turning around the nation's lacklustre productivity and economic performance over recent decades.
www.abc.net.au
- Related coverage: gartner.com
Gartner HR Survey Finds AI is Boosting Productivity but Widening Workforce Divisions in Australia
While AI is delivering significant productivity gains across Australian workplaces, it is also increasing workforce division and job uncertainty in 2026, according to Gartner.www.gartner.com - Related coverage: percapita.org.au
- Related coverage: pc.gov.au
Quarterly productivity bulletin - March 2026 | Productivity Commission
Australia’s labour productivity was flat in the December quarter and rose 1.0% over the year to December 2025, reversing the 1.0% fall over the year to December 2024.www.pc.gov.au - Related coverage: pc.gov.au
Bulletins - PC productivity insights | Productivity Commission
The Bulletins unpack the latest ABS productivity statistics and examines trends and new developments underlying Australia's productivity performance.www.pc.gov.au
