Speech: Moving from AI pilots to whole-of-government scale

NOTE: This speech was delivered on Tuesday 4 August 2026. Check against delivery.

Opening: beyond the pilot graveyard

Good morning, and thank you for the invitation to speak with you today.

It is a privilege to join Tech in Gov’s AI stream to talk about moving artificial intelligence from promising pilots to whole-of-government scale.

When I spoke at this event last year, much of the discussion was about accelerating AI adoption: knowing when to speed up, slow down, apply the brakes and stop.

A year later, the conversation has shifted—and the APS AI Plan has been released, giving us a shared direction for safe, responsible adoption across government.

As Matt Yannopoulos told Chief AI Officers last week, the challenge is no longer whether the technology is capable. It is whether our institutions can absorb change at this speed while preserving—and strengthening—the trust on which they rest. His message was that ambition and responsibility are not opposites: governing well and moving fast have to reinforce each other.

AI is now part of everyday public sector work: drafting, triage, analysis, service design, risk detection, coding, correspondence and the productivity tools people use every day.

The hard part begins when a convincing proof of concept meets real service volumes, imperfect data, legacy systems, procurement rules, cyber requirements and public expectations.

At scale, the agency must be able to sustain the cost, understand the architecture, manage the supplier, support the workforce and explain what happened when the system gets something wrong.

A pilot proves that something can work. Scale proves that we can operate it.

So the task is not simply to scale the technology. It is to scale the disciplines around it.

Government cannot do that alone. We need industry to design for transparency, interoperability, cost visibility and changing assurance needs—not bolt them on after deployment.

So today, I’ll make 4 points: scale the operating model, not just the tool; design for trust and the exception; build capability across the workforce; and give CIOs room to modernise, not just sustain.

1. Scale the operating model, not just the tool

Let me start with the first point: scale the operating model, not just the tool.  

Responsible AI has to follow the work—the data, authority, exception and record. It lands inside technology estates, supplier contracts, workforce pressures and live services.

Moving from pilots to whole-of-government scale is more marathon than sprint. Some agencies will move quickly; others need time to strengthen data, train staff or clarify ownership. The aim is not a single speed, but steady movement without exhausting the people it depends on.

Consider an AI tool that helps staff triage correspondence. Its access must be defined. Classification and escalation rules must be clear. Records must be created. Staff must know when to intervene.

The APS AI Plan, Responsible Use of AI Policy, AI Impact Assessment Tool and AI Technical Standard provide common scaffolding: expectations that help agencies ask the right questions earlier and explain decisions later.

The choice is not between moving quickly and governing well. It is to make governance useful enough to support movement: clear thresholds, simple escalation and room for local judgement.

Governance should make the safe path the easy path. If it is slow or disconnected from real work, people will find workarounds—and some will create risks the organisation cannot see.

DTA’s proof-of-concept-to-scale advice can help here. It doesn’t replace agency judgement. It helps structure that judgement: problem, value, cost, ownership, data, controls and the pathway into operations.

A valuable use case may still need better data, ownership, controls or process redesign before it scales. Finding that out early is not failure. It is stewardship.

2. Design for trust and the exception

My second point is about trust and the exception. When leaders talk about scaling AI, the language often turns to throughput: more transactions, lower latency and less manual effort. Those are legitimate goals. In service delivery, small improvements in triage, routing or document processing can reduce backlogs, help staff find information faster and make services feel less fragmented.

The measure is not only whether a task became faster. It is whether the work became better: simpler for the user, more consistent for staff and more defensible for the institution.

That is the difference between automation and improvement. Automation can make an existing process faster. Improvement asks whether the process should look the same once the technology is available.

The likely next phase combines process mining with specialist AI. Process mining shows where work actually waits, loops or duplicates; specialist AI can then be applied to a defined domain, grounded in its rules, data and professional context.

That matters for productivity. Durable gains come from redesigning the end-to-end workflow—removing bottlenecks and rework, and focusing specialist judgement where it adds most value. The unit of productivity is not the prompt. It is the service, process or outcome.

We can already see the direction in industry. Australia’s Cropify uses specialist vision to classify lentil samples in about 90 seconds, compared with a 24-minute manual process, with reported accuracy above 98 per cent.

CommBank, working with Apate.ai, uses specialist AI agents to engage scammers and turn those interactions into intelligence for scam detection and disruption.  

And Archistar applies domain-specific AI to planning and site feasibility, testing millions of design options at a scale impossible through manual analysis.

These are not generic chat tools placed beside the work. They are specialist systems embedded in the workflow.

And once those systems are embedded in the workflow, the design challenge becomes broader: at scale, they must handle the exception as deliberately as the standard case.

The better ambition is empathy at scale: removing friction without stranding a user behind an error message or leaving a public servant carrying unsupported risk.

That requires clear handovers, escalation, reviewable records and human judgement where needed.

Trust and pace are not competing objectives. Safeguards give agencies, staff and the public confidence that movement is deliberate, evidence-based and open to correction.

3. Build capability across the workforce

That brings me to the third point: capability across the workforce. Many pilots ask whether a machine can do a task faster or cheaper than a person.

The stronger question is whether AI expands capability—helping people find evidence, spot patterns and apply judgement where it matters.

That capability has to be broad. If only confident early adopters can use AI well, it will widen gaps rather than lift the workforce.

Production systems still need practical controls: evaluation, monitoring, access, records and feedback.

The best model brings technology, assurance and the people who understand the work together early—while the design can still change.

Those operational experts know where processes break, where discretion matters and where an elegant tool could cause practical harm.

IP Australia offers a useful example. Over about 8 years, it has embedded specialist AI in the patent process and learned through practical use. Independent scrutiny recognised the program’s strengths and identified opportunities to make controls more robust, which the agency accepted.  

That is a learning culture: trial in the real world, welcome scrutiny, act on evidence and improve quickly.

Leaders do not need every technical detail. They do need to know how performance is monitored, what happens when the system is wrong, who can override it and how the organisation is learning from use.

Leadership also has to create permission to learn. Staff need clear boundaries, practical help and confidence to raise concerns. Otherwise they will either avoid useful tools or use them quietly—neither is a sound foundation for scale.

4. Give CIOs room to modernise, not just sustain

The fourth and final point is the challenge facing CIOs: how to modernise, strengthen cyber resilience, manage AI adoption and maintain critical services in a tight fiscal environment.

They carry two cost curves: ageing systems that absorb funding and capability, and variable consumption through cloud, tokens, model calls, data movement and storage. Unit prices may fall while total expenditure rises.

The answer cannot be another layer of technology. We need to simplify the estate as we modernise it.

That means adopting before adapting: standard processes and platforms wherever possible, with customisation reserved for where difference creates public value.  

Otherwise, it raises cost, complicates upgrades and creates new legacy debt.

A simpler core also creates room to use AI more intelligently: interpreting context, supporting specialist judgement, identifying exceptions and connecting information across standard systems.

AI can also help with the legacy problem itself: mapping dependencies, documenting business logic, identifying duplication and helping agencies decide what to retain, refactor, replace or retire.

But sophistication needs financial discipline: clear ownership, real-time usage monitoring and measurement of cost against outcomes—not simply users or prompts.

The operating environment is changing too. Patching and model updates that once arrived over weeks can now occur within hours—or in real time—compressing the time to assess impacts.

We therefore need to work with technology partners differently.

Partnership must extend beyond procurement into continuous operational assurance: visible changes, timely evidence and confidence that controls still hold.

Moving faster cannot mean losing control. Agencies still need notification thresholds, audit trails, testing, rollback and the ability to pause when risk is unclear.

That is the CIO opportunity: create a smaller, simpler and more reusable estate, and shift investment from sustaining complexity to creating capability.

Closing: keeping judgement and accountability human

Let me finish by bringing those four points together. Moving from AI pilots to whole-of-government scale is not a software project. It is a test of whether we can build sustained capability without losing the fairness, clarity and accountability people expect from government.

Not every pilot should become permanent. The test is whether we have enough evidence and discipline to expand, pause or stop based on what we learn.

The technology will keep changing. Capabilities that feel novel this year will become ordinary soon. But the work of public administration remains the same: exercise judgement, act with integrity, explain decisions and hold responsibility where it belongs.

The next phase will not be won by the organisations with the most pilots.

It will be won by the systems we can operate, explain and improve together.

So my challenge to government and industry is this: stop counting pilots and start scaling what works—together. Prove the value, improve it in practice and share the learning.  

Our operating environment is changing quickly; our partnerships must change with it. That is how we lift productivity, move at pace and keep trust intact.  

Thank you.