At a glance
42% of Australian organisations already have a chief AI officer or equivalent, and two in three expect one by 2027.
Compagnone argues a workload's requirements should be defined while the use case is being scoped, so the infrastructure decision follows from them.
She says the role should own strategy, governance and outcomes at six months.
Sovereignty runs to three layers, Australian-located, Australian-controlled and Australian-owned, and no longer carries a price premium outside air-gapped environments.
She ranks Australia's shortages as skills first, data second, compute and power third.
Compagnone says AI pilots need production thinking from the start
Australian organisations are appointing chief AI officers faster than they are defining what those roles control. Few have settled how decisions about AI workloads should be made, from which use cases are prioritised to the requirements that shape where those workloads run. We put six questions on the infrastructure implications of AI leadership to Datacom's director of artificial intelligence, Lou Compagnone, who recently authored a whitepaper on the emerging CAIO role.
One of Compagnone's observations is that the infrastructure conversation often starts too late. She argues that requirements such as latency, data sensitivity, sovereignty, compliance and where AI processing occurs should be defined while the use case is being scoped, so the eventual infrastructure decision follows the workload rather than constraining it.
The whitepaper, The emerging role of the chief AI officer in Australia, was published in July 2026. Datacom commissioned the research house Omdia to survey 507 Australian IT and business leaders in April and May 2026, all at organisations of 200 people or more. Two thirds of the sample sit in the private sector, the rest across federal, state and local government, other public sector entities and not-for-profits. It found 42 per cent already have a chief AI officer or equivalent, a further 21 per cent expect to appoint one within a year, and 93 per cent expect the role to become a business-as-usual expectation.
Datacom is a New Zealand-founded technology services group, one of the largest in Australasia, with 5,935 staff across 32 locations and FY26 revenue of NZ$1.58 billion. It sells cloud, managed services and AI consulting to enterprise and government clients on both sides of the Tasman and runs five New Zealand data centres after acquiring Highbrook in East Auckland.
Compagnone has spent 20 years in the field and comes from service design and digital transformation rather than engineering, which is itself the argument she makes about the role. She hosts the AI Ready Women podcast.
Her written answers follow, published as submitted and edited only for house punctuation.
Q: What does the role actually own?
"At six months, a CAIO should own three things:
the AI strategy and roadmap,
the AI governance and risk framework, and
connecting AI initiatives and use cases to measurable business outcomes.
The value of the role isn't owning everything to do with AI, it's connecting strategy, governance, technology and people so the organisation can move together. It is an orchestration role, working closely with leaders across people and culture, group technology, finance and so on.
While the CAIO should be involved in the systematic identification and prioritisation of AI use cases, execution of those use cases should be handed off to the wider business. The whole point of a chief AI officer is helping an organisation to move from scattered AI activity to deliberate transformation. If they're still running pilots at six months, something has gone wrong!
And vendor management shouldn't be their main focus either. Only 9 per cent of respondents in our survey saw this as a CAIO priority, while 47 per cent cited clear executive mandate as the top success condition. Loading the role up with tool selection and vendor negotiations turns it from a transformation role into a procurement role. It just becomes about adopting tools rather than adapting the organisation."
Q: Who decides where an AI workload runs?
"Usually it's a negotiation between the CTO or CIO, the infrastructure team, and security and risk. But the Chief AI Officer should also be involved because they know which workloads are strategic and how deployment shapes the roadmap.
But the point I want to make is that it's important to decouple the AI use case from the infrastructure. The starting question shouldn't be 'where do we run this?' It should be 'what does this use case actually require and what's important?' For example, is it latency, data sensitivity, sovereignty, compliance and so on. If you can answer that clearly, then the infrastructure decision becomes a logical outcome. This needs to happen when use cases are being defined and prioritised, not after a proof of concept succeeds. The use case should drive the infrastructure, and not the other way around."
Q: What is the wall between a pilot and production?
"AI pilots often live in a forgiving sandbox with sample or synthetic data, manual workarounds, and minimal integration. Production changes everything! You need live, governed data pipelines; integration into core systems with proper security and identity management; resilience, monitoring and uptime; and a plan for long-term maintainability, such as how to manage factors like model drift, retraining, and token cost at scale.
Most organisations don't plan for this. Our survey found that only 25 per cent have a standalone AI strategy, so the infrastructure conversation happens too late. Organisations prove the pilot works, then discover that production is a very different conversation about compute, data architecture and security.
The first phase of AI transformation for organisations was all about experimentation, 'AI confetti' scattered across the business. The next phase is about operationalising AI: what scales, what stops, and what changes in the operating model. That's when infrastructure has to be a strategic conversation, not just a procurement one."
Q: When an Australian client asks for sovereignty, what are they asking for?
"Well first of all, sovereignty isn't one thing, it has a number of layers.
Australian-located: in other words, data stays on Australian soil. For government and regulated industries, this is simply expected now.
Australian-controlled: who can access the data, which laws govern the platform, and whether a foreign government could compel disclosure without Australian consent, for example what happened with the CLOUD Act. This is where defence, health, critical infrastructure and financial services are heading because where data sits and who controls the platform aren't the same thing.
Layer three, Australian-owned: the infrastructure provider itself is Australian-owned. This is narrower, typically defence and national security, where the sovereign supply chain itself matters.
But another important dimension when it comes to AI is the difference between where data is stored and where it is processed. You can store data in a sovereign facility in Sydney, but if you're using a frontier large language model via a hyperscaler's API, your prompts and documents are being processed offshore. Storage sovereignty means very little without inference sovereignty. That's what's driving growing demand for locally hosted AI processing, in other words models running on Australian-controlled infrastructure so both the data and the compute stay within the sovereign boundary.
In terms of pricing, two years ago sovereignty was a specialist offering with a cost uplift. Today we find that our clients expect it built in, not something that is considered a premium. Air-gapped environments still carry different economics, but for mainstream workloads the expectation is parity."
Q: What should Australian operators expect from government as a buyer?
"The mandate has had real impact. Our survey found that 73 per cent of organisations say the APS requirement is shaping how they think about AI leadership, and 93 per cent expect the CAIO role to become standard. People assume government is slow on technology, but on this one the public sector is leading.
But appointing a CAIO isn't enough in itself. What matters is whether agencies give these leaders real authority, the mandate to connect strategy, governance, technology, people and delivery.
As agencies move from pilots to running AI at scale, they'll have a growing say in where those workloads physically sit. And given government's focus on sovereignty and security, that's going to mean Australian-located, Australian-controlled environments.
For Australian operators, the next two years will bring three things: more demand for sovereign hosting, tougher procurement requirements, and clients who need much more than just compute (they will want data pipelines, model management, monitoring and retraining, all running locally).
And as I mentioned earlier, that includes inference, not just where the data sits, but where it's processed. Government will increasingly expect both to stay onshore."
Q: What is Australia short of?
"The other 93 per cent aren't short of just one thing. They're short of several, and they're all connected.
Skills are the most pressing. You can build a data centre, but you can't quickly train the people to design, deploy, govern and maintain AI systems.
Data is next. Too many organisations are trying to run AI on messy, fragmented data that hasn't been properly governed. And the tooling that fixes that, pipelines, integration, governance, is infrastructure too, it's the foundations. We should talk about it that way.
Compute and power is third. Australia needs more capacity across the board, and domestic sovereign capacity for sensitive workloads still has a way to go.
And lastly governance. Organisations aren't asking for more regulation, they're asking for clearer expectations so they can move with confidence."
What to watch
The commercially significant line is the pricing one: sovereignty at the storage layer has commoditised while the version buyers will pay for has moved to inference, where far fewer Australian facilities can evidence a claim. A buyer who needs pipelines, model management, monitoring and retraining onshore is not buying colocation. That is a managed sovereign AI platform, a different product at a different margin, and one most Australian operators do not currently sell. Operators are starting to staff for it, as NEXTDC has done with a dedicated sovereignty seat on its executive team.
Her sequencing point has the same edge. If the attribute a workload depends on gets specified while the use case is still being scoped, operators differentiated on latency, sovereignty or privacy need to be in the room earlier than their sales cycles assume. Those are the axes the GPU-first neocloud tier already competes on.
Her ranking of Australia's shortages is worth reading against the 9GW of data centre load in AEMO's connection queue. Putting skills and data ahead of compute and power makes the readiness gap a sequencing problem before it is a supply one, with capacity arriving ahead of the people to run workloads on it.
Her six-month clock is stricter than the market she is describing. Only 15 per cent of Datacom's respondents expect a chief AI officer to show impact that fast, and 46 per cent allow six to twelve months.
Every answer above rests on the same assumption, that the chief AI officer can act on it. Deciding infrastructure while a use case is still being scoped means overruling a technology team that has already chosen a platform. Refusing a use case on token cost means stopping something with a sponsor behind it. Both take authority that a title alone does not confer, which is why her research names a clear mandate from the CEO or executive as the top condition for the role delivering.
In part two we test that assumption, against Datacom's finding that 47 per cent name a clear mandate from the CEO or executive as what decides whether the role delivers, against the four different AI structures Australian governments have built in twelve months, and against the argument from CIOs and CTOs that this should have been a technical job all along. It runs as chief AI officers need authority, not just a title.