DEEB, Mr John, Vice President, Customer Success, and Field Chief Technology Officer, Workato Inc. KAHLER, Mr Cameron, Public Service Director, ANZ, Workato Inc. ROY, Mr Bhaskar, Acting Chief Executive Officer, Workato Inc. CHAIR: We will now resume our public hearing of our inquiry into artificial intelligence. Welcome. I understand information on parliamentary privilege and the protection of witnesses has been provided to you. I'd like to now invite you to make a short statement, as short as you possibly can, which will give us an opportunity to have the time for committee members to ask their questions. If there is anything you'd like to provide in writing to the secretariat to supplement your opening remarks, that would be greatly appreciated, but I'll leave it to you to make some opening remarks, and we'll get underway with questions. Mr Roy : Firstly, thank you, Chair and members of the committee, for the opportunity for us to be here. We truly appreciate it. In terms of introducing our organisation, Workato is an orchestration and automation company, and we operate as the control and execution platform for AI. What I mean by that is our customers and organisations use Workato to govern and scale AI capabilities inside their organisation, and they do that safely by ensuring that they connect to their enterprise systems in a safe manner, plus doing so in a vendor-neutral capability. In terms of the technology stack that they already own or AI models that are there, we provide that platform to be able to leverage those, agnostic of what kinds of models they use or what kinds of technologies they choose to use. As a company, we have 25,000 customers globally. We're so happy to share the learnings and observations from there as to how we see AI getting adopted but, more importantly, as to how our customers are scaling AI in order to drive, at the end of the day, what matters, which are the outcomes for their own business and how they do that throughout the business. In Australia, we have hundreds of customers. They range from small companies to large enterprises to local, state and federal agencies, and we are committed to Australia and growing that in a way so that it is, again, keeping our notion of being vendor neutral. We work very closely with a number of delivery partners here in Australia. We also work closely in terms of—you may have heard earlier the talk about the Dukes model. We are looking at how it does that. We can dive into it because early testing has been really great not just for Australia. We figure that, with a model like that, we can actually take it to our customers globally. There is another report that I read recently, where the Treasury found that close to two-thirds of organisations in Australia are using AI but that fewer than one-tenth of them are truly using it in some significant manner. This is not just in Australia; we see this globally, to the point where we have coined a phrase—'the intent-to-outcome gap'. The intent is always there. AI is a transformational technology. People want to use it. Not everything is ready for AI—so you choose. But where you can, you can drive a lot of outcomes for that particular organisation. That's the gap we see. The gap is predominantly because companies and organisations are not seeing eye to eye in order to drive that significant use. That's the gap we look to address. Our submission caters to that, to how we help drive that in Australia. There are three elements I'd like to mention. The first element is how the government itself can adopt AI and show the art of the possible to everyone—how can that be done? The second element is, from a procurement setting standpoint, how the government can favour local vendors, whether it be delivery partners or local model providers, so that you compete on an equal footing with the frontier models which are there. The third element is essentially looking at—and we encourage the government to look at—subsidies, especially when you look at small to medium businesses. The reality of the cost of AI is there, to the point where a term like 'token economics' has become a key thing that people talk about. How can you provide subsidies to small and medium businesses to upskill them, as well as leverage this technology, so they can reap the benefits of AI? With that, I welcome any questions. CHAIR: Thanks. We appreciate those opening remarks. I am keen to get further reflections on the broader adoption across Australia. We've been hearing a lot about the big frontier models and the capacity they have to potentially increase our sovereign capability locally. To the point you made in your opening remarks about adoption, what does that look like and how does that work across the economy, including in government? You describe model neutral, or vendor neutral. How quickly and at what cost can a customer currently move from a model, whether or not it's frontier, to another one? Do you see that getting easier or more difficult into the future? What's your sense of the ability to switch? Mr Roy : I'll tackle the first question first, regarding the frontier models. This may sound very cliche, but that reasoning is abundant, but that is not what drives business outcomes. Reasoning is abundant, but the reality is that results are not. Often you need to have specialised models in order to drive the right kinds of outcomes they're looking to drive. From an analogy standpoint, you won't take a Ferrari to do grocery shopping two blocks down the road; you just don't need to. You need those specialised models which come into play in order to drive the right kinds of outcomes which are there. We are seeing that globally. You are aware of this; everyone is talking about sovereign AI as a big thing. Every country is talking about it. Then there are various organisations that are building these domain specific models, SLMs and those kinds of things, which are catered to driving the outcome which is there with reduced cost. That's a movement we support. In terms of switching models, I cannot talk broadly, but within Workato it's done in seconds. This goes back to the intent-to-outcome gap; we understand the intent and select the right model for that right intent. The right model is looked at in three different ways—performance, cost and whether it is relevant for that particular use case. We select the right models. If, in certain cases, the model is not available, the switch is one toggle; you just say, 'I want to use this model for this,' and you just create this as a preference setting. Switching models or bringing their own model in order to drive the outcomes is not a big deal. CHAIR: Can you explain a bit further—this is Australia specific—where we have the advantage to drive different ways of innovation in this space. Where do you think, with the right environment, the right investment and the right framing, we're able to see Australian innovation here? Mr Deeb : We can talk a little bit about the experience we've had with local models. Previously, Trellis was discussed as being a local language model offered as sovereign capability, which we've been exploring through our partnership with Gulanga on Dukes. We've leveraged that model to achieve some pretty interesting outcomes around performance as well as relevance. If we look at it just in the context of a government agency that may want to use a LLM or AI to help draft content, whether it's a policy, an email or something relevant to their agency specific domain, when they ask a model like Trellis or Agentaus—which is the model name—the response is faster and more relevant than a frontier language model. That reflects on the applicability of that in government and in other areas that require that Australian knowledge. We've seen that performance and we've seen the outcome by plugging our agentic framework into the Agentaus model. Mr Roy : Why wouldn't you invest in something like this? AI is a transformational technology. If you look at Australia itself, you have the resources, whether it be land, power or people. You even have the trust as an organisation or a government globally. My question to you all—maybe I'm not seeing this right, but it seems like a no-brainer—is: why wouldn't you invest in something like that in order to make this a level playing field? With the technology, like John was talking about, there are promises that we see that such good technology providers, not just LLM—LLM is one part of the AI capability. There are such great technologies coming out from Australia. Why wouldn't you look at this and say, 'This is something that we need to build as sovereign capability'? CHAIR: I appreciate that. Deputy Chair. Senator DUNIAM: Thank you for making the time to join us here in person today. This is a historic opportunity for us as a country, and we're all optimistic about what we might be able to do here. To your last point, why wouldn't you invest? It's the question we're asking, but I suppose the canyon between us and where we want to be is, 'Well, how do we get it right?' Earlier you listed three points, which I want to get a little more detail on. In short, our job is to provide recommendations to government about how we get things right in order to get to the outcome you talked about, where we all want to arrive at. If we can just step through them and get a little more detail on how you think we might be able to do it well here—they may be experiences you've had elsewhere, given you're an international company. On that element of accelerating AI adoption across government agencies, how do you do that? We've had other witnesses talk about the slow or muted take-up, limited through particular models available to them. What do you think we need to do? Mr Kahler : Thank you very much for having us today; we appreciate it. The questions about scaling AI outcomes across government—what can we do? Our recommendations were made around term of reference (h)—'the adoption of AI by Commonwealth departments and agencies to improve services for Australians'. We see in the APS AI Plan, under the tools pillar, this big focus on reuse—so IP reuse and tools sharing, the idea being that we don't need to reinvent the wheel time and time again, agency by agency, and incur a cost over and over again. If we focus on that concept of reuse, the Australian federal government has huge advantages over corporations. There are 194 Commonwealth primary bodies, and, if we look at AGOR—the Australian Government Organisations Register—that extends out to over 1,300 separate entities when we go beyond those primary bodies. This is 1,300 individual entities who don't compete with each other. This presents a significant opportunity to broadly share AI learnings and proven solution templates and scale our AI outcomes and business impact across the whole of the government. Unlike the competitive organisations who don't want to share that IP that we see in industry, government agencies don't need to waste time, effort and money reinventing the same solutions over and over again. You have a force multiplier for AI outcomes and delivery in that context. Whilst we understand these entities are all going to be different in their core purpose or function, most are still going to have common use-case opportunities across core business processes. All are going to have HR functions managing onboarding, offboarding and the employee life cycle. They're going to have IT teams and IT service delivery teams helping serve employees, finance managing budgets, do accounting, do procurement and do things like that. Extending this further, many of these bodies will have similar case-management and service-delivery processes and functions when it comes to citizen delivery as well. When an agency creates a really great solution that's designed with reuse across government in mind, results should be very easily shared and reapplied across the government for maximum impact. I'll just talk about a couple of very basic examples that that we see. Take, for example, an agency who leverage AI within their procurement and IT stack to understand current software licences that they're contracted to, look at spend and align that with the actual usage of those vendors and are able to deprovision the unused licences and generate savings either immediately or at renewal. An agency who applies an AI pattern like that might save five to 10 per cent of their annual software licensing bill, and this would be considered a fantastic outcome for an individual agency. But, if we apply that across the federal government, Australia's highest single group of purchasers, that is a huge potential outcome, resulting in the reuse of just one pattern. That's one example. I think the data that we see is that the Australian federal government spends somewhere in the vicinity of $4 billion to $6 billion a year in software. That's quite significant straight out of the gate, with one use case being reused. Again, take an agency that leverages AI to assist with IT service help desk queries and is able to deflect or solve 50 per cent of those cases without human intervention or assist with onboarding and provisioning of employees across HR, payroll and IT systems. Again, depending on the size of the agency, this could represent a significant operational efficiency gain for that agency. But, across government, that ROI scales like crazy. This is before we even start talking about how AI can assist service delivery expansion and deliver greater citizen outcomes. I choose those particular examples because they are considered very safe. They've proven that they can deliver ROI very quickly. And, importantly, they build an internal capability within government and trust within the teams using AI and automation technology, who can then take those skills and target more meaningful use cases, be they service management, citizen outcomes or things like that. Again, these should also be very easily shared across government, creating a feedback loop and an innovation flywheel. This is the value of reuse across government and at scale. Hopefully, this is close to the answer to the question that you were asking. It is possible, if the tools and platforms are designed to encourage reuse and interoperability with the existing systems that are there, which are plug and play, as you mentioned earlier, and are made readily available to government. Senator DUNIAM: Okay. That does segue into this issue of procurement. You've talked there about government agencies accelerating, as per the first suggestion from Workato. Specifically, are there any other jurisdictions that do this well and that favour sovereign capacity? What entities or individuals are able to provide a service and models that might work? Mr Kahler : I might comment quickly on procurement before we talk about other jurisdictions, because that was one of the recommendations we made, and we've seen it come up a couple of times in the committee today. It goes back to that concept of reuse across government— Senator DUNIAM: To coin a phrase, a self-licking lollipop is something that's made its way into our APS. Mr Kahler : Yes, I think our partner, Gulanga, gets a royalty when that's used! If we look at whole-of-government agreements, for example, that are part of procurement, wider access to the tools which are proven for reuse is a current challenge. Whilst there are coordinated contracting arrangements managed by the DTA, these only cover six vendors. This was based on pre-existing procurement volumes, not technology outcomes and certainly not emerging AI technologies. Whilst it's relatively or comparatively straightforward for an agency or a department in government to consume services from these particular vendors, this doesn't really fit for the emerging AI technology vendors. We feel there's a gap there as well. While there's a focus on reuse in the APS AI plan, currently, it's difficult for that to be extended when a vendor solution that sits outside of those six providers is there, even if it's built for reuse. We understand the APS AI plan also aims at mitigating the risk of vendor lock-in, for example. And we do feel that our recommendations here are very well aligned with the government's own plan around making whole-of-government access for these types of tools for reuse. Senator DUNIAM: Perfect. Thank you. Mr Roy : If I could add one thing, having a pluggable platform is important in this particular case. Whether you're building tools or agents in any kind of environment, you're making sure that it is governed and run in a safe way. And tomorrow, if new innovations come in, you want to really leverage those innovations which are there without having to go through a full replatform or rebuild altogether. Getting something which enables you to work across all the technologies and leverage the innovations that are coming in while ensuring that everything is done in a safe way is something that we see our customers and various organisations that we work with leaning towards. We all know how fast AI is changing. We cannot really predict what it's going to look like even two or three years from now. Rather than have a locked-in capability and then, in two or three years, again figuring out how we really do this, because something has exploded, and we need to take advantage of this—that's a key part to remaining flexible in terms of how you leverage these technologies. Mr Kahler : Part of the question was on other jurisdictions that we've observed doing it well. I might pass to John. Mr Deeb : We wanted to share a little bit about our experience with the Singapore government and what they've done, like the establishment of a number of different programs and grants to support AI hubs to be created in Singapore. That's all about driving greater investment in the country. They provide a lot of incentives through grants to support the growth of organisations and hubs. We're happy to share more detail around that. It's all publicly available information. Obviously, there are tax concessions on elements of hiring and growing teams that meet the criteria. There are deductions and R&D helping bring global talent into the country, providing subsidies on hiring in Australia—or, in their case, in Singapore—and just co-funding of AI adoption. A lot of incentives drive the establishment of these AI hubs in Singapore to then drive that knowledge and retain that knowledge in country. That's something we felt has been effective. Certainly, we have established an AI hub in Singapore, leveraging some of those incentives. CHAIR: Senator Pocock. Senator DAVID POCOCK: Thank you very much for your submission and your time. I wanted to ask a bit more about procurement. In your submission you say that keeping the integration layer open ensures local AI model providers can plug into existing government systems alongside global frontier models, and you go on to say that whole-of-government licensing arrangements need to be more accessible to local and flexible vendors. We see huge contracts and spend with AWS, Microsoft and the like. Are Australian companies being given an opportunity as part of GovAI? What's happening there? Mr Kahler : When it comes to GovAI, the Department of Finance came out with a request for expressions of interest in June this year. It was open to all submissions. I believe that there's been a pretty high amount of submissions made. Us, alongside our platinum partner, Gulanga, who you heard from today—one of those who made that submission. I wouldn't be able to comment on the opportunities being given to local providers in that process, as it's ongoing. Certainly, prior to this, our position has been that we should be interoperable, agnostic and flexible to be able to pick whatever models—be they Australian or frontier fit for purpose—the choice should be the contract holder. Senator DAVID POCOCK: What are the risks to Australia if we have this lock in to some of the big multinationals and local providers miss out? What's at stake? Mr Roy : There is an opportunity cost. This was associated with it. We can discuss, John. Mr Deeb : If we look back at technologies that are less transformative than AI—nuclear, computing, space, wi-fi or solar—were all opportunities that Australia could have been more involved in earlier. Looking back, for example, at solar, University of New South Wales research set cell efficiency records. We had the talent, and we had the knowledge, but we didn't make the investment in supporting manufacturing. We see something similar. This is, again, probably why we're here to talk about the opportunity loss. We think skills are key. If we don't provide a path for AI research and applicability, and companies that deliver outcomes, then those people will go overseas and work. That's one thing we should be really concerned about, given what a great education system we have and that we do produce a lot of talent in this area. To the procurement example, government buying power to give Australian built, Australian governed AI a path to scale—I think we've heard that a couple of times today from other witnesses—is the opportunity. They're the things that can be controlled by government. At the application layer, integration and governance, these are the areas where we are emphasising our MOU around the integration of Australian LLMs in our products, giving customers both Australian and overseas the choice to select Dukes, which is powered by Trellis Data as an LLM for where they want to use it, whether that's in powering an Australian-specific process or choosing a cheaper alternative to the frontier models. Sovereign deployment and measurement are others that we would ask you to consider around the opportunity cost. We do think that there are things that could be done to improve that. Senator DAVID POCOCK: Are there specific things that you think should be changed or looked at in the Commonwealth Procurement Rules? I assume, when it comes down to some of these procurement decisions, you've got someone putting forward or making the decision who thinks: 'I'm not going to get in trouble for going with Microsoft, but this other company says they have the capability, but I don't know. I'm not an expert.' We've heard a lot about creating a level playing field. I would love to see us prioritising Australian sovereign companies, but what are the mechanisms to actually doing that in the rules? Mr Deeb : I think it was mentioned earlier—the idea of actually creating to policy to consume Australian sovereign capability in the LLM space. Again, you can run benchmarks, you can test the models that are out there that are Australian sovereign, and you'll find that they perform to the needs to certain governments agencies and certain uses. We've talked a lot about using the right model for the right thing at the right time. These things can be delivered, and setting goals and targets for the consumption of Australian sovereign AI is a good starting point. Mr Roy : There are some great technologies being built, including models. John was talking about Dukes. We found that performing so well compared to the others. Then the question becomes, as we test it and it becomes really good, why shouldn't that become the default model? When Australian organisations and the customers that we work with are in Australia, why shouldn't we look at making that the default model because it is really good? It works better in those cases. Those kinds of policies and looking at how we really promote those is a key part. It's not just in Australia. There is also an element of how you take that across to other regions as well. If that model performs the same way that it's doing in Australia and we tested it to see if it works great in the US, UK or other places, why wouldn't we lead with that? At the end of the day, we want our customers to drive outcomes and see the benefit from this. If that model drives that outcome and benefit in a much better, cheaper, faster way, then why shouldn't we provide that to them? It's a benefit to that particular customer. That's the angle that we look at not just in Australia but globally. Senator DAVID POCOCK: What is something that you'd like policymakers to understand in terms of models? There is so much focus on the frontier models. What's the next one coming out? I think everyone just assumes that whatever Anthropic and OpenAI put out next will be the best thing for everything, but how should policymakers be thinking about LLMs and their use cases? Mr Kahler : Great question. I might go back just to finish that last question that you asked around procurement settings. I think procurement settings haven't really changed that much, but AI as a technology is so rapidly evolving. If we look outside of government at our corporate customers, they are much more easily and freely able to trial different technologies, to try before they buy, without the overhead of the procurement processes that the Australian government faces—outside of the ones you don't get fired for buying, to your point. We would probably welcome a conversation and encourage a conversation with vendors and government around how those procurement policies can be better aligned to test, learn and actually understand whether it does what it says on the tin when it comes to these new emerging technologies. Mr Roy : To answer that, there are two ways to look at frontier models. Frontier models are learning from everything, all kinds of content which are out there in the world. They're really, really good in terms of individual productivity. But when you come down to organisations it does not translate directly. Individual productivity is not the same as driving business results. For that what matters is how are the processes getting orchestrated, because that's directly relating to the business KPIs. Sorry, I might get a little bit technical over here, but there is this entire thing around context, right? You provide documents and all that information. That's the context. In the business side of things, your context is there in your systems. It's not out there in the world, because it is private to you. It's in your financials, ERP, CRMs, whatever it is. It's in the policy documents and the SOPs that you write. But then there is a third element. It's in the processes that you do, which you have fine-tuned, whether it be serving the customer right or the finance process. All of those things are very specific to a business. Like Cam was talking about earlier, those are still horizontal. HR still has the hire to retire process. Finance has order to cash and procure to pay, those kinds of processes. But those are very specific to organisations. I think that's an area to look at right. Frontier models do a great job with world knowledge. Actual enterprise knowledge and being able to leverage that to drive business outcomes is a key thing where you start seeing business results. Senator DAVID POCOCK: And so you think the opportunity for Australia is to build up— Mr Roy : I actually think that's the starting opportunity for Australia as to how businesses, from small to medium businesses to large enterprises, can really reap the benefits of AI rather than going and saying—and this is a personal opinion—'Let's now build a frontier model.' They are ahead, in terms of that. But this is something to look at in terms of what can really drive the growth in Australia and drive businesses to leverage AI and to become powerhouses there. Mr Kahler : Another comment on that, which we talked about earlier as well, is that it's part of the maturity curve we see with our customers. They start safe internal business processes which builds an internal capability and a trust in the technology that they're using. They may even realise that actually—I know this is an AI committee—they can solve this without AI. Just because AI is a hammer doesn't mean everything is a nail. It's about the capability building as well, and there's a lot of low hanging fruit within organisations and across government for operational efficiency which will generate huge gains. The other gain will be that building of a skill set and capability and trust to then move more confidently into higher impact and potentially higher risk areas that will drive bigger rewards. Mr Roy : That's where Cam introduced this concept of reusability. Think of all these processes—things that AI is doing in order to really drive those outcomes. They become available to small and medium businesses throughout, so all of a sudden they are not having to put in the research—the R&D—to do it. They can just take those reuseable components, which have worked in other places, and apply them. Provide them with the mechanism to apply that throughout and all of a sudden you start seeing things getting built up, which I believe helps, overall, Australian companies. Senator DAVID POCOCK: Thank you very much. CHAIR: Senator Sheldon. Senator SHELDON: If an agent causes harm while acting on a government system, how does the responsibility get divided up among, in this case, Workato, the model provider, the agency and the officer who authorised the workflow? How do you equate that? How do you approach those responsibilities and accountabilities? Mr Deeb : Our capability as a platform is to monitor and manage the consumption of AI by the organisation. We have a gateway that can monitor what's being consumed and what actions are being taken by AI. You're creating rules and observability around identifying when an agent isn't doing what it's meant to be doing but extending beyond its guardrails. So we can identify when it's doing the wrong thing and help organisations capture that and perform actions relating to remediation. From that perspective, I think it's about bringing back the ability for organisations to monitor their use of AI and their AI agents. We hear the term 'going rogue', but monitoring and managing that is, I think, a responsibility of the organisation building the agent. What we heard yesterday was about testing those agents doing those things and that the organisation should be responsible for that. One thing that was also noted was notification. In privacy management—privacy of data—we make a commitment to notify within 48 hours of the discovery of any security incident. I think that's a global standard that has been adopted by a lot of companies around privacy and privacy of data. Similar things could be applied to organisations in terms of their need to communicate and notify of this notion of rogue agents. Mr Roy : There is this notion around a control plane. This is the place where you set up the governance as to what an agent should do. Then there is a run time. Is the agent doing what it's supposed to be doing, which is what has been set up at the governance layer? Typically, agents that go rogue are ignoring the governance that has been set up because, at execution time—I'm sorry if it is too technical—it has found new context and new things to do and then it goes, 'Let me go take a shot at that.' We believe it should come together: both the control and execution should be in one platform. Why is that important? Look at control, where you're defining what an agent should do. When an agent goes to execute it, they will stop it, saying, 'Hey, this says this and you're trying to do that; it doesn't allow that.' Whereas, if it is two different platforms, you don't know. It won't be in lockstep. Our belief is you need something which can control not just at the governance level, where you're defining it, but also at the execution level to make sure that it is sticking to those very specific guardrails and the authentications and permissions as to what an agent can or cannot do. CHAIR: Unless there are any further questions, we're right on time, so that's perfect. Oh! My apologies, Kate. Ms CHANEY: That's okay—just a quick one. CHAIR: Ms Chaney, yes. Please. Ms CHANEY: I did say I didn't have any but changed my mind. I want to test some assumptions when it comes to government use of AI. Tell me which of these statements is wrong. Firstly, the performance gap is significant between frontier models and open-source models when it comes to specific functions such as defence, so there are some things where we might actually want to have the very best, and there is a difference between the latest and the second latest. Secondly, that performance gap between frontier models and open-source models is likely to continue. I think those two things lead to the fact that, for a limited number of government functions, we're going to want to have those frontier models. I just want to check that and then also get some feedback on whether there are particular government functions where you think that frontier model difference is important. Mr Deeb : What you mentioned around national security—these frontier models are very capable of identifying vulnerabilities in technology. In terms of organisations and entities that control and want to identify exploitations in technologies, they would want to use a frontier model for that. So we absolutely agree with that statement that there are functions in government where you want the frontier models or the most advanced frontier models. If that gap continues, I think that's an acknowledgement of other models, like Agentaus—we'll use that as an example—which is a local sovereign LLM that has actually greater knowledge of Australian law and Australian policy because it's been trained on that data more specifically, and therefore it could be more applicable in other areas, like we mentioned earlier, where it could be a government agency worker who's looking to draft content, whether it's policy content, an email or a white paper, that is aware of Australian policies and laws. So I think you're right to say that it really cannot be a one size fits all when it comes to a model, and we just need to get better at identifying the pros and cons of each model and when to use it. Mr Roy : If you look at the frontier models, they are also really good at code development, right? On the coding side of things, the frontier models do a really good job in terms of putting those together. To try various kinds of operations and processes, you may not need something like a frontier model to do that, but that's another area where I would say leaning on frontier models for actual engineering work is something that we are seeing produce really good results. Ms CHANEY: This may be a very ignorant question, but as to Agentaus or other models, if we're talking about building them in Australia, is the assumption that they would be based on an open-source model and then post-trained and fine-tuned in Australia? I assume we're not building those from scratch. If they are post-trained and fine-tuned, what are the implications as those open models are upgraded? Do we get those upgrades, or do we end up five years behind? Mr Deeb : That's probably a better question for Agentaus to talk about their life-cycle management, but they would have that choice to leverage the latest models and train. I think the expectation would be— Ms CHANEY: But mostly they'd be based on an open source if we're building them in Australia— Mr Deeb : Yes, I think that's the case. Ms CHANEY: which is likely to be Chinese, is it? Most of the open-source models are Chinese, aren't they? So, if we said we want sovereign AI for a bunch of government functions, we may be using Chinese open-source models, post-training and fine-tuning them. Mr Deeb : I don't know that that's the default. But, again, I think we'd defer to the actual model providers to share more context there. Ms CHANEY: Alright, thank you. CHAIR: Thanks very much. I'll bring this session to a close. Thank you very much for making yourselves available to the committee and answering our questions..