STEEL, Mr Pete, Group Executive, Technology and Artificial Intelligence, National Australia Bank [by video link] [15:07] CHAIR: Welcome. I understand that information has been provided to you regarding parliamentary privilege and the protection of witnesses giving evidence to this inquiry. I now invite you to make some opening remarks before we get underway with questions. Mr Steel : Chair and members of the committee, I just want to thank you for the opportunity to appear today. Artificial intelligence is arguably the most significant technology shift of our time. Adopted thoughtfully and at scale, we believe it can definitely lift productivity and support higher living standards in Australia. As Australia's leading business bank, we share with our customers the goal of improving productivity. We believe Australia should approach this moment with careful ambition. History tells us that economies which adopt transformational technology with energetic aspiration ultimately outperform those that invent it. At NAB we're already using AI to help us respond to customers better and faster, reducing manual work so our bankers can focus on customer needs, and it's helping us run the bank more resiliently and efficiently, with examples such as protecting against fraud and cyberthreats. Realising AI's full potential means thinking differently about how we tackle the big challenges ahead of us. We're building new teams and careers. We're creating in-house AI capability and a culture of opportunity and accountability. These teams are bringing the people who understand a business problem together with the technologists who can solve it, but, critically, can solve it safely. The risks with AI are real, and the guardrails are absolutely critical. The most important safeguards we have are that we decide when and how we use AI and that we have accountable execs who oversight it. Human judgement and oversight remain absolutely essential. Getting these things right really matters for us at NAB and for the whole economy, which is why we're backing initiatives to support small businesses and help grow larger businesses. Our research, which we detailed in our submission, shows that 40 per cent of small and medium businesses already use AI today and another 13 per cent plan to, but that does vary widely by sector. In retail and transport, it's closer to one in five. Capability and confidence are the big barriers to wider adoption, and that's why we're working with the National AI Centre to deliver free AI training for small-business owners, built around the everyday problems that take up their time. We believe strongly that Australia should be more than a destination for AI built elsewhere. While we work closely with the frontier labs, we're also a founding partner in the government-backed Buy Australian AI Partnership Program, because we believe Australia and Australian businesses like ours have much to gain from establishing sovereign capability. That means helping Australian companies research, build and scale commercially viable AI here and nurturing an ecosystem that provides both resilience and opportunity so that this country and its people can reliably share in the full economic benefit. On regulation, we welcome the government's plans for a framework for Australian standards for AI. We support settings that encourage innovation and adoption while providing the right safeguards, transparency and accountability. Those settings should be practical, technology neutral and able to adapt as the technology changes. All of this rests on trust and transparency. We need to be very clear about the risks. We also need to be upfront about jobs changing. Some will go and some new ones will emerge. History suggests new and interesting jobs will be created over time, just like they were with the digital evolution. Our shared focus should be on helping people build the skills to take part in an AI augmented economy. Thank you. I look forward to your questions. CHAIR: Excellent. Thank you very much. I appreciate those opening remarks. Because I think this is something that is integral to how we, as a country, are shaping our own AI future, I'm going to start as I did with the previous witness, and that is around what AI sovereignty is. What does it actually mean for us, both for NAB within the financial sector and for the broader economy and our country? What do you think AI sovereignty should and does mean? Mr Steel : It's very topical. There are a number of ways you can look at it. Maybe I'll share a few thoughts. The first is that, for me, it means supply chain confidence, resilience and independence. We don't want a world where another country can determine whether we can use AI, particularly as it becomes more and more intrinsically linked into the Australian economy and into our businesses. We need to insulate ourselves somewhat from that switch. So it certainly means resilience and independence, and certainty of access to models and the ability to run them here are critical, regardless of undersea cables or any other geopolitical events. Also, for AI sovereignty, we obviously want to be doing as much of the building and training as we can. I know your previous questions have really highlighted that with other participants. To explain a vision, I'd also go to a couple of places where I see there are really productive ecosystems around previous generation technology: Silicon Valley and Tel Aviv. I've had the privilege of going to both, and there you see legs of the stool working together. You see universities educating talent. You see some of that talent leaving and starting careers in industry. You see them generating innovation and fostering startups and/or scale-ups that become commercially viable. So you see the monetising and commercialising research, then, potentially, investment coming back into further innovation and then, ultimately, more universities attracting graduates into what is a self-fulfilling ecosystem. If we can have that sort of opportunity in Australia, it also leads to a level of sovereignty and independence where we start to attract and generate real wealth for the nation by having all legs of that stool working together. CHAIR: Excellent. Noting those remarks and your opening remarks, what is your reflection on the idea of the connection between our sovereign capability—or the means by which we can develop that in terms of skills, access and broader capability—and our own AI ecosystem? How connected do you think that is, and what opportunity exists by enabling or leveraging off the fact hyperscalers at the frontier are keen to come and base their training here in Australia? Mr Steel : Yes, I think it's a very good opportunity for us to have the data centres and have some of the training done here. We should not bypass that. It's quite a unique opportunity to become relevant and a significant participant in AI globally. It shouldn't be at any cost though, obviously. I know we're probably going to talk about IP protection and some of the things that we are trading as part of that. Back to the sovereignty and independence point—being able to train and run AI here is really important, but not as an outlier. It needs to be almost able to run independently of whatever else is going on in the world. I think about attracting data centres, attracting training and models and us being able to create domain-specific AI innovation as well, whether it's by industry or some specialised subindustries that complement what's going on particularly in the US. That's a big opportunity for us as well. The ability to run data centres and train models here is giving us some critical mass to then enable sovereignty to become a reality and enable us to run things locally and sustainably. CHAIR: Drilling down to local level, what policy settings do you think we need to look at, improve or build upon to help increase or enable the uptake of AI across SMEs? Mr Steel : I think there's educational element as well. Again, we've been investing in education. I think the government has got a big role there as well. Some sort of incentive program to encourage education is critical, whether in universities or in businesses across the country. Those policies are very helpful. There's having an openness to encouraging a startup ecosystem. Again, we've been participating in that. But wherever the government of the day can encourage that, make that advantageous and attract that investment, it is really helpful. Having the appropriate safeguards and regulation is something uniquely government can do. We should obviously be self-regulating; we care about running the bank well, and we care about consumer interests and, for the long term, having a sustainable business. But it is certainly helpful for the government to encourage and mandate the minimum safeguarding and self-management that all businesses should be adhering to in Australia. They would be a few thoughts that we would welcome. CHAIR: Ms Chaney, I understand you've got some questions. Ms CHANEY: Thanks for your evidence and your submission. I hear you on the importance of sovereign capability and am really interested to understand more about the bank's current experience—t he tools that you currently use, how dependent you are on foreign providers at the moment and how susceptible you'd be to disconnection or access limitation risks. Are you seeing this need for sovereign capability in real time, in your current experience and use of AI tools? Mr Steel : It's a great question. We're conscious of it. We feel it's very early days in the AI journey. We have been using advanced frontier models internally for a variety of use cases, which I can touch on in a moment. Obviously, we are very dependent on US companies at the moment. While we have no immediate reason to have concern, as AI becomes more and more foundational to the way we run the bank and serve customers, certainly, it does introduce a level of risk of disconnect, in some form. We suffer that a little bit today—all the banks do—with the way we host services in the cloud. We typically run all our cloud services through US based providers. That's probably the more significant example. But, yes—absolutely, if for some reason we fell out with Western AI providers at the moment, we would be somewhat exposed to disconnect. Again, I feel that connection with sovereignty is really important, sustainably, to give us critical mass and to allow us to run things, which may be Western, US models trained and extended locally, to hedge that risk somewhat. Ms CHANEY: I am interested in the use cases and understanding to what extent we need to guarantee access to those frontier models, for which things like data centres might provide some leverage, and to what extent your AI needs, as you foresee them in the next decade, could be met by Australian AI providers on models that are built based on open source models. Mr Steel : On our use of AI—again, I can talk to some use cases—the general pattern is quite interesting. We typically start with the advanced frontier models—maybe not the very most advanced, military-grade models, but reasonably advanced frontier models. We solve some of the wicked problems that have presented themselves over the years, usually cognitive processing. We tend to repeat cognitive processing. That's where we apply the frontier models. And that's almost to crack the problem. Then—you're absolutely right—we don't necessarily need the most advanced models to run the agents after that. It can be substituted by much cheaper, potentially not leading-edge, open-weight models. That's something that, obviously, we would prefer to source from Australian providers. I think there's a real opportunity there. For the bleeding edge of innovation and problem solving, we probably would use, in the reasonable short-to-medium term, US based models. But we can run things, once we've solved the problems more efficiently, with our local model. I think that does present an opportunity for Australia. Ms CHANEY: You mentioned you would like to see appropriate safeguards and regulation and for government to set those minimum safeguards that companies should be complying with. Can you talk a little bit about what you see as being the scope of those minimum safeguards? Is it about testing? What types of things do you think should be covered in setting that minimum requirement across the board? Mr Steel : I've got a few themes there. First of all, I think the regulators have a role to play to help industry understand best practice as well. APRA gave a helpful response, where they wrote to banks and said, 'These are our expectations of what you should be doing.' And, look, we'd like to think we were compliant with most of that good guidance already, but there are gaps for us to fill as well. So that's a good role for regulation, anyway—to say, for certain industries or even enterprises, 'Here is the rising waterline of what good companies need to be doing,' whether it's governance or running AI. That's a first category. As a second category, we've all seen Hugging Face and some of the things that have happened recently, which you've been inquiring about. It's making sure best practice for how to run AI is understood, not just in government but safely inside companies. I could see a world where well-intentioned companies start and make the same sorts of mistakes that some of the frontier lab training has done. It's just making sure that guidance is absolutely clear. There's a third area that I think regulation or government generally could help with. I do worry about it. I know we're talking about cyber generally as well. We have access to reasonably advanced models to do our own self-testing, to check if our cyberdefences are strong. It's a privilege to have access to those, and NAB is a reasonably large company with well-funded cyber teams. Not all Australian businesses have that, and I worry that, where the open-weight models are lagging three to six months behind the frontier models, those methods of attack will become more commonplace as those models get into the hands of different actors. There are thousands and thousands of businesses in Australia that don't have the level of defence that we do and the level of access to resources. I think another way government can help, whether it's ASD or some other vehicle, is with the equivalent of what is called red-teaming—testing our defences on all the smaller digital companies in Australia and somehow providing them with a report or providing the boards with a report to say, 'You are exposed in the following areas; you need to patch up your vulnerabilities, defend yourself, before, in a couple of months, there are going to be bad actors who have access to capability.' That national service hasn't really been required before. I think something like that would be very welcome, not just for us but for the tens of thousands of businesses that are potentially going to be exposed in the years ahead. Ms CHANEY: Do you see that as a role of the AI Safety Institute or the Australian Signals Directorate? Or does it not really matter, as long as it gets done? Mr Steel : I think the operational use of those advanced models to achieve that is probably something more in the remit of ASD; that would be my guess. That's a somewhat secretive organisation. Probably the public engagement and communication is best done by the safety institute or some other part of government. Ms CHANEY: And on the cyber front, what types of AI-enabled fraud and scams are you seeing, including things like deepfakes and voice cloning? How is the race going between defence and offence? And what do you think AI developers or telcos should be required to do in that space? Mr Steel : The race is real, and it's ramping up all the time. We see directly fewer of the voice scams. We see a huge number of the written scams to customers, and social engineering scams as well. With what previously might have been a misspelt email, where there were ways to detect if it was quite fraudulent, AI is, unfortunately, allowing fraudsters to lift their game. That is a very real threat. An organisation like us does everything it can to communicate to customers about those risks, first and foremost. Our colleagues are educated and help customers to understand the risks and also, to the very extent that we can, intervene if we think a customer is in the process of being defrauded, which is really important. Then we have a whole range of automated mechanisms, some of which use AI, to help detect fraudulent activity. That's been happening for many years with traditional machine learning and fraud engines. But, per your example, where things can potentially be faked up with AI, we're in this arms race of using our own AI to try to spot the fakes and then apply our traditional techniques of alerts and triaging. So it's very real, and it's a cost of doing business that I think will be with us forever now. Ms CHANEY: Do you think the tech companies that are producing the AI models that are enabling that uplift should bear some of that liability or risk? Can you see opportunities there to hold them accountable for the bits they have control over? Mr Steel : I do. And I realised I didn't answer one part of your previous question, sorry. I think the upcoming fraud legislation very importantly goes upstream to telcos and to social sites like Facebook. That is critical. I worked in the UK for three years. Similarly, we see that many of the fraud scenarios start on social platforms, such as Facebook Marketplace. So having fraud regimes from the government extend upstream is extremely important, because it's almost not too late but quite difficult once the fraud has been monetised at a bank. That's really important. To your point, I see examples where frontier models such as Anthropic have learnt and now their models are trained to almost stop you doing what they think are nefarious things. With some of the models, if you try to initiate a cyberattack or something, it'll say, 'Sorry; I'm not allowed to do that.' That's good. That's one case that they're onto. You're exactly right. Fraud and other cases they're not onto so much, so I think the first request would be—our friendly providers from the US—all these models should go to the other cases that should be blocking at the root training for the model. However, that only goes so far. There are open weight models as well. We definitely think there should be some sort of liability regime as well, so that it becomes an enduring requirement for model providers to not participate in fraudulent activities. Ms CHANEY: Thanks, Chair. CHAIR: Senator Sheldon. Senator SHELDON: Thank you, Mr Steel, for joining us. In April, NAB and the Finance Sector Union settled a court case on excessive working hours. Part of that settlement was an agreement for an industry—and I describe it as leading—health, safety and wellbeing framework, which included a greater oversight of the risks posed from overworking. This is to the credit of all the parties. At the same time, NAB was targeting $450 million in productivity for financial year 2026. Your submission says that AI will help reduce manual tasks throughout the business. I just want to ask you: is your safety focused approach on workloads being extended to AI usage? And that includes the question of work intensification and burnout as well. Mr Steel : Can you repeat the last part of the question? Senator SHELDON: Is your safety focused approach on workloads being extended to AI usage, workload intensification and burnout or potential burnout? Mr Steel : Thank you for the question. Definitely as part of when we're considering—and as follow-up on notice, I can work with our workplace teams to give you a fuller answer. But whenever we— Senator SHELDON: That would be helpful too. Mr Steel : When we do workplace change activities, we always consider mental health implications as well and look at current workload. If I take a step back on the AI implications, again, I've used digital technologies for many years to help the group serve customers better, to be more efficient. AI is another tool like that. It's a different technology, but it serves the same purpose for us. Generally, the pattern is we will use technology where we can have a better customer experience—ideally time to serve or cost to serve. That will often result in colleagues' roles changing, and that usually takes them out of less rewarding roles. Our stance is we haven't lost roles to AI—sorry, we haven't lost FTE employees to AI. We have redirected people into customer-facing roles instead. And also we have created some new roles that didn't exist before to help manage AI as well. So far it has been, I think, a somewhat virtuous redirection of people using AI. That will only go so far. Eventually, I'm sure all workforces in Australia will start to see some change in size and shape. But, to the heart of your question, yes, mental health and workplace stress is very much part of our consideration set. Senator SHELDON: Just to tease that answer out a bit further on the size and shape—are you thinking overall size and shape? Or are you thinking that because of the additional services that AI will give assistance to, and because you're wanting to compete with other competitors, providing more services also means more human interaction and oversight into AI as the technology continues to evolve? What's your view of how that's going to roll out over the coming year and years? Mr Steel : Yes, my personal view is we as a bank want to grow our relationships with customers more than ever. It's a competitive market. As we get more efficient, there's definitely a movement to having more customer-facing colleagues, deepening relationships. That's genuine. We have thousands of lenders and bankers. We will have more in the future helping more customers rather than fewer. AI in our world is often more about training models, supervising them and continually evaluating the efficacy of those models. We are creating new jobs—that actually becomes an ongoing job. It's quite an interesting job, dare I say. It's writing test cases and saying, 'We want you to do all the following things, model, and this is what success and failure look like.' That's taking people who maybe didn't grow up in an AI world and redeploying them into that as an activity. Again, that's quite an interesting growth area. The bank also, like all organisations in Australia, has a backlog of problems to solve. We've got capacity demands that we can't meet at the moment. So we would also argue that, as we get more efficient, we're going to be applying colleagues to that backlog of problems that we've got. The bank has grown in FTE over the last several years. I can imagine, though, that we just want to become as efficient as possible, but at the moment our stance is that there is more than enough work, and the nature of the work is changing as well. Senator SHELDON: That's a great deal of help. Your submission calls for—and you touched on this already—safe and responsible AI use, with clear governance arrangements which provide transparency over how it's used. I know this is not atypical just for NAB, frankly. In 2025, when surveys were carried out on AI and technology, some of the results showed, across a variety of banks, that there was not enough transparency and consultation on the government's requirements for AI. What sorts of figures have you got from surveys you've carried out amongst staff? Have we seen a shift in the last 12 months, and is it a noticeable shift? Mr Steel : I'll share a few thoughts and figures with you. I think it's important to recognise there has been a lot of anxiety, and there will continue to be, about AI, whether it's uncertainty about 'my role changing in the future' or whether it's some of the doomsday scenarios. Things have moved quickly over 18 months. Things we were worried about have shifted from white-collar-worker concerns to other well-publicised concerns that are in the media at the moment. But I think there will be continued anxiety, no matter what we and other organisations do. It's just the nature of new, unprecedented technology that impacts society in ways we have never seen before. That would be my first overall statement. Having said that, we do a great deal to legitimise those concerns but also to encourage excitement about AI and individual learning and adoption. We had an AI learning week two weeks ago inside NAB. We had over 8,000 participants self-enrol to all sorts of workshops, with guest speakers internally and externally—Anthropic and others. It was extremely well received. We had an over 85 per cent satisfaction rate for those sessions, on average. So I think we've got an extremely willing workforce who want to understand AI more and learn. Many people want to get themselves future fit, using AI. Some people won't, and that's okay. They maybe aren't as excited about changing the way they work. But, generally, we're going to be a forward-looking organisation and embrace AI. We want it to make everybody more productive. I haven't got a colleague engagement stat at hand for you on their willingness or understanding and acceptance and excitement about AI, but I'm happy to take that on notice and see what we can get for you. Senator SHELDON: Thank you very much. CHAIR: Excellent. That brings us to the end of this session. Thank you, Mr Steel, for making yourself available to the committee and for the evidence you provided to us today. You've just taken a question on notice. If you could please provide a response through the secretariat within the next two weeks, that would be greatly appreciated. Thanks very much..