AI expert Alex Papley joins us to cut through the hype and explain how artificial intelligence can make real, practical difference to safety management — from analysing audit evidence to writing better reports in less time.
Artificial intelligence (AI) is moving fast, and aviation safety is no exception. This episode explores where AI can add real value to safety management and where it still needs a human in the loop.
Alex Papley, an AI expert with 25 years of IT experience and founder of Hypergen, joins hosts Grenville Hudson and Kathryn Harvey to break down what AI actually is and what it is not. Alex explains the difference between traditional AI and generative AI (GenAI), why GenAI has opened the door for smaller operators, and why treating AI as a summarisation tool rather than a thinking tool is the key to getting good results.
The conversation covers practical use cases across the four pillars of safety: safety policy, risk management, safety assurance, and safety promotion and tackles the big questions around job security, data myths, and the ethics of handing over tacit knowledge to a machine.
________________________________________
Upcoming Training: Making Your Safety Management System Operational
The SMS compliance deadline is 1 September 2026. Operators who are not ready risk serious regulatory consequences. Here are the course details:
Book your seat at aviationsafetycommunity.com.au
________________________________________
Alex Papley is an AI expert with 25 years of IT experience, including eight years specialising in AI. He has worked as an AI developer at Salesforce in the United States and as a lead data and AI consultant at Microsoft, working with large organisations across Australia. He is the founder of Hypergen, an AI consulting and development company.
Times are approximate, based on the conversation flow.
________________________________________
Don't forget: Making Your Safety Management System Operational
The SMS compliance deadline is 1 September 2026. Operators who are not ready risk serious regulatory consequences. If you have not booked your place yet, head to aviationsafetycommunity.com.au
________________________________________
Subscribe so you do not miss an episode, and join us at our next Safety Leadership Forum to keep the conversation going.
Hello, and welcome to the Aviation Safety Community podcast. I am Grenville Hudson, and my co-host is Kathryn Harvey. Today, we're talking with Alex Papley, who is an AI expert from the company Hypergen.
Alex has 25 years of IT experience, of which eight of those years has been in AI. He was a developer of AI for a company called Salesforce in the USA, which is a major US company, and after that had worked with Microsoft as a lead data and AI consultant to various large companies in Australia. He comes with an extraordinary amount of experience in AI, and we've got him here today to give us some insight into how AI might be used in safety management.
Hello, Alex. How are you? Hey, Grenville. Yeah, nice to meet you.
Thanks for... Yeah, thanks for the intro, and yeah, good to see you. Yeah, no, it's great. There's so many things we could talk about here, and I realise that aviation isn't specifically your area of expertise, but AI is something that's developing with great rapidity here in Australia and worldwide, of course, but aviation is, I think, keen to see how we might be able to use AI, especially in the areas of safety management.
Can you give us a bit of an overview of how you could see that AI may be used in safety? Yeah, no worries. I'm more than happy to. I mean, it's a pretty big field, so I'll do my best to cover some of the things off.
I think maybe taking a step back, so we think about AI. There's a lot of different bits of AI that actually make up AI, and so the thing that's changed really in the last few years is this concept called generative AI, which is what the AI that powers things like chat GPT. The reason it's suddenly taken off is because if we think about it, organisations up until maybe 2022, they were always stuck when they hit a document or a PDF.
So what we call unstructured content, which is things like it could be interview transcripts or it could be compliance documentation, things like that, if you ever were trying to work with those things, you literally were always forced to double click and open up the PDF and read it and try to understand it, or you might be creating a document and doing those things. And the fundamental thing that changed was that when OpenAI launched chat GPT, and there's a thing called a large language model, which is the thing that powers that software, it suddenly opened up pretty much every industry, particularly ones that are paper heavy, where they're full of what we call unstructured data and made it easier for them to actually make sense of it. So when I think about use cases in aviation, for years you've been able to do actually the mathematical types of AI and what we call traditional machine learning, and nothing's really changed in that regard.
That's been able to be done and airlines have been doing that for probably quite some while in terms of predicting the right price to sell a ticket for, for example, on a plane. And that's been traditional AI and it's been done. But the new stuff is the stuff that's actually really interesting.
And the reason I say that is that it's made it a lot more accessible for smaller firms as well. So when you think about it, if you're doing safety reports and you're doing interviews, being able to ask a natural language question over a transcript and get a particular detail that otherwise would take a long time to get, that's one example. Being able to look at, you know, changing regulations and make a good sense of, you know, what might have changed between one document and another.
Helping do deeper research into areas that were very, very rich in text, but again would take a long time to analyse. Those are all the areas that AI is now able to help. So in a nutshell, it's about being able to make sense of what we call unstructured content, which is primarily text, but it also could be images and video and audio.
And being able to make sense of it when we just couldn't do it in any automated fashion up until probably the beginning of 2023. So, Alex, with regard to safety management, let's take a little bit more of a deeper dive into, I guess, some of the core areas that safety managers and safety professionals look at. So, you know, you've sort of sparked a few ideas with you saying that it can transcribe interviews or it can analyse interview text, it can look at compliance documents.
So some of the areas, ICAO, which is the International Civil Aviation Organisation, talks about four pillars of safety. So they talk about safety policy, risk management, safety assurance and safety promotion. So I guess we could sort of break things into those pieces.
Maybe looking at risk management. You know, one of the things that we see a lot of is that risk assessments can be quite cumbersome, quite document heavy, quite administrative. People, you know, don't really like doing the admin bit.
The risk management pieces can be interesting, but it's about making workshops interesting. But it's a lot of sort of, you know, these spreadsheets often or similar, like there might be software out there that they've got, but it's like this field and this field and this field. Could you see AI supporting the risk assessment process? Yeah, look, it's really interesting.
Like, I think there's different touch points across the process that we can look and see where it is most useful. I think being able to help make sense of a large set of data is one of the areas that works well. And I generally talk about particularly generative AI, which is, as I said, sort of the new part, but really the part that I think has got the most, the broadest applicability to companies.
You've got to treat it as a summarisation technology, not a thinking technology. And so what I mean by that is that where you actually say we're going to take the cognitive load and, for example, we're going to run an interview, we need to get certain bits of information, we need to create a report, where it's going to be quite helpful is helping you synthesise and get the right information together to then repurpose it in the way that you want. Where I would generally, and particularly in aviation with the safety culture as well, when you hear about things around AI makes stuff up and it's not always accurate, it's almost always because you're not treating it as a summarisation technology, you're asking it for answers without giving it information to work off.
And so that concept of hallucinations happens generally because we're not giving it enough to start with. And so when I think about, you know, your question, I'd be saying certainly for being able to synthesise key bits of information out of, say, infears is a big one. Software is an interesting one as well.
We as a company, we help companies both understand where the opportunity is for AI, but then also help build solutions. And so we're sort of, we're in the thick of it when it comes to actually writing code and making things that work, but also looking at where you should invest and where you shouldn't. And one of the interesting trends that we're seeing is the cost of writing software is not zero, but it's rapidly going to zero.
And it's still notable, but it's different. And so one of the things I think we're going to see is people, not in every single case, but organisations saying, look, actually the software that we need to go and buy is expensive, it's cumbersome, it's not really doing what we want. And, you know, we're sort of buying a shoe and then we're trying to fit our foot into a shoe that's not really well suited.
We're finding now that once you hit a certain threshold in cost, you're actually better off just saying, look, let's just roll a solution that meets exactly what we need because, you know, and again, I don't want to exaggerate it, but if you go in with a mindset of saying the cost of writing software is a lot less than what it was, you can still spend all the time doing the discovery and the consulting and making sure the applications are what you need, which has to happen anyway in any project. And obviously the testing as well, but that potentially 50% of the time is spent building something, which is now significantly reduced. And so one of the byproducts I think we're going to see is probably better suited systems for people that actually help them do what they need to do, just because organisations will say, actually, let's make the boot fit our foot.
Yeah, pretty much every single company I've worked in, in safety, they've been through a tech change with their risk management system. And they have always said off the shelf product with minor configuration updates. And the frustration that that's caused to the end user because it doesn't quite fit and to get changes and the costs involved in, you know, making tiny changes to fields or backend configuration is huge.
So that's a real, yeah, that's a great evolution in that space. Yeah, we're definitely seeing, it's funny, it's not, we didn't deliberately go out and build Hypergen around that, but we're finding more and more companies are saying, look, you know, it's funny, we're doing one for a consulting firm in a different industry. But, you know, they've got a very spreadsheet based process that sort of use that to document things, but maintaining it and keeping it current is just a real issue.
And, you know, you can think about like all the different types of compliance requirements that organisations have and smaller ones might be doing something very similar as well. You know, that process is saying, actually, let's just turn that into a small app that's secure, deployed, backed up, got all the governance around it, the things that we need, the reporting. I think it's going to help a lot of organisations actually just get stuff done because if you go, if we can, I mean, systems are always going to be there and I'm not one to say that, you know, AI is going to remove everything, we're just going to talk to a computer and you don't need them.
I don't think that's the case at all. But I think we are going to find, you know, in the space of safety and in general in business, people being a lot more deciding, you know, what is the right thing for us? Should we buy or should we build? Because maybe it's going to be more cost effective in building out some IP that's actually adding value to the business and going down the second path. But, yeah, so when I think about safety and compliance, so much unstructured information means that you've got this massive opportunity to say how can we selectively make more sense of this and generate the information we need in the format we want with less fuss on the creation so we spend more time on the, you know, what do we do with this and actually add, you know, improving safety and improving culture and all those things that are actually probably the core function people are trying to solve for.
Just with regard then to say assurance activity, so audit is a big one and I'm thinking about the time it takes to audit. You know, you've got external, internal audits. Sometimes these are one-, two-day audits.
You're collecting a lot of evidence. So synthesisation of this evidence could really cut your time down. Yeah, it could and I think the thing that we find that the biggest challenge is you've got to think about going, well, there's a very deliberate process and what evidence do I need to gather? And there's almost those questions that, you know, you're in someone's mind thinking I need to find this and this and this and then look at it and see if it's actually complete and if it's actually meeting a requirement.
And then once I've got sufficient information, then it's about saying pulling that together and turning it into another report or, you know, that audit. I think there is a lot of opportunity for it. The thing I'd sort of caution and say is that there's effectively this concept of tacit knowledge that you've probably heard of, which is effectively the knowledge that you have in your head that's not necessarily, you know, Google, you know, and so it's not about just finding it online.
It's about this is the way we do things. This is how we think about it. So for those processes to work really well, you need to start capturing your tacit knowledge.
So when you're thinking around, you know, when I'm looking for certain questions, if I need to generate a report, I'm actually looking for certain answers within all of this information. And so it becomes more and more important we start really thinking about what are the questions that we're actually asking and why? Because AI, it's going to summarise, right? But it's going to summarise in a way that if we give it a certain set of criteria to summarise against, it's going to do a very good job of doing that. But if we give it a loose, hey, was this looking compliant to you? It's obviously very unchanged.
And we're not giving it enough information to actually do a very good job summarising. So the thing is that I think there is a massive opportunity and the way to think about it in safety is saying what are the questions that we're actually asking? Why? What are the things that we're looking for that are effectively in that experienced mind that we sort of know, you know, almost the sniff test and as much as possible capturing that? Because then if you can apply that, and it's almost like it could be a really long document that's actually a brief that is actually a prompt. But you save it as a document so you can reuse it.
So you can say, right, here's a transcript. Here's the document with all the things you need to think about when you're looking at this transcript. Give me a report of that.
You'll find that consistently you'll get a better result because you're actually capturing all that tacit knowledge in something that's reusable. And so that's the way I think about it. And the reason I say that is that I don't want to sort of mislead people and think that AI is just going to give the report without that middle piece.
And the really critical thinking and that, you know, the cognitive load that we take on with those tasks is still massively important. And so it's almost documenting that in a way that we can actually start thinking about reusing that across different... ..yeah, across different domains and areas. And that's in every organisation, not just, you know, this industry as well.
Alex, I was just thinking about some of the comments you made about developing and buying off the shelf and all that. Can you give us a bit of an idea of what's actually out there software-wise and how the development side might link in with that or if it's a separate activity in itself? Where do you start your approach? If you're going to go down the line of off-the-shelf type software, where do you go to for that? And then if you're going to do the development side of it, does that link in with the off-the-shelf or is it separate? That's a good question. I mean, so if we take a step back, so generative AI, particularly the latest models, are very, very, very good at writing code, right? So they can write code very well, they can test against the code and make sure it works well.
So really I think in any industry we would say that there's always going to be a handful of really good, you know, preferred software vendors out there and often a long tail of smaller ones that also do some good work in niche areas. So I think in general we say, look, go along and investigate those because it makes sense to if the problem's been solved and they offer support and they're doing a good job, then, you know, it's a logical place to go. The thing is, and maybe to Catherine's point before, the question comes up, though, when we implement the software, are we spending a lot of time then trying to get the software to work for us? And the more time it takes to customise it, the more it becomes often cost prohibitive versus actually saying let's get AI to write the code because if you think about it, in any project there's a discovery piece and we'll call it wireframing, which is when you're agreeing on what's the UI, you know, what's the user experience going to look like, how are we going to think about interacting with this? Those things will always stay the same, whether it's, you know, off-the-shelf purchase software or whether it's software that you might choose to write yourself.
So that piece of work, that discovery, what's the actual process, what are we trying to solve for is always going to remain and that needs to be done, right? The piece, though, is the so what afterwards. So it's at that point where you say, right, we know what we need and then you make an assessment and you say, right, is there a product that meets that perfectly? And if there is, then we'd probably say go buy it because that's going to be really easy to turn on, it just works. From there then the next piece, though, is if it's not, then that's where we can rapidly turn all that discovery information into code and that's really where we find it goes a little bit smoother.
So we generally say there's like the more integrations, then often it's easier to write your own rather than build, but there's a few other rules that we have on our blog that we've sort of set up. There's about four or five criteria. And that's an important point, I think, just to emphasise the more integration.
So you mean the more software and tools and apps that a business has, the better it is for them to go down to develop their own software? It can be because you're spending so much time often paying for integration software to actually integrate and then you're paying people to integrate that and do all the work. It starts becoming more, you're just sort of adding cost and cost and cost. And particularly with SaaS or software as a service, you've got that monthly recurring.
So you're actually adding a fixed cost to your operations bottom line and that's the thing that's a real challenge. Sometimes companies would prefer to cap exit and just say, look, let's just get it done, have a lower ongoing cost so that we can deal with ebbs and flows and cyclical issues and things like that. So I think, again, each one's a financial decision, but we would generally say that, you know, I've got them here.
The four factors that we look at, one is high cost relative to value. So if you're looking at it going, what's the system that we're looking at, you know, and again, that cost relative to the value that it's going to provide back. The feature utilisation, if you're paying for all of the features that you're not using and there's 5% that you really need, you know, those 5% are the ones you want to be paying for.
From an integration perspective, it's twofold. If it looks like it's actually going to be a very standalone system and doesn't need to integrate, it also simplifies the build effort significantly. So the replacement effort is quite low when you're going, we don't need to integrate.
And that's also particularly the case where you're trying to connect to data sources that are very tightly controlled. So, for example, you know, a lot of people use Xero, right? Like Xero itself is in arguments, sorry, in some perspectives, a very simple application. It's just a forms-based application.
But the things that make it difficult are the fact they've got bank feeds and their integration to, you know, government for super handling, you know. And so those things are, you can't just go along and say, hey, Commonwealth Bank, can I have an integration? You know, you need to do a whole lot of work to actually get to that point. And so those are the things that make it more difficult.
And so that's where you say, we're going to buy software that's already solved for that and have done it well because it makes sense. But then if you don't have those dependencies or the integrations of very simple services on the internet that you can just connect to, then that's where we look at it and say sometimes it's not there. And the last one is the workflow logic.
It's not a complicated process when you look at it. It doesn't have a huge amount of domain expertise that has very, very complicated logic that's baked into a product. And, again, Xero is probably a good example where they've built in a whole lot of accounting rules over the years that mean that to replicate that is quite a lot of effort because you've got to rethink that and you've got to reinvent the wheel.
But where it's actually not, it's sort of a stage one, stage two, stage three, report created, off we go, have a couple of approval gates. Those things are usually simpler and that's where we say often the economics, you know, change quite a lot. And so we also offer advice to SaaS providers to say this is how you can think about structuring your application to provide a level of defensive motion so that you've actually got, you know, a really strong value proposition moving forward because a lot of them still do, right? But they need to focus on the areas that are going to be valuable.
What about scalability of AI within businesses and, you know, different business sizes and all that sort of thing? Is the use of AI, you know, limited by the business size? It is a bit. It's twofold. So when we think about traditional AI, which we talk about machine learning and you may have heard the comment you need to get your data in order to do AI, that's very traditional AI thinking and it's to do with analytics.
And so typically, yeah, you need to be a big company to buy and build an analytics estate, which is when you get all your different systems and they're all saving in a data, they call it a data warehouse or a data lake or there's a few other terms going around the industry nowadays too, or a lake house. But those things where you need to do traditional AI, which is to do with predictive analytics, so looking at here's a, you know, again, what price should we sell a ticket for? You absolutely, there is a barrier to entry for smaller firms, typically because they don't have the data, enough vast amounts of it to actually get decent predictions and also the cost to sell those platforms. With generative AI though, which is, again, the AI that powers like a copilot or a chat GPT, that barrier is not there.
So everyone's got a Word document, PDFs, and these AI models you can use quite effectively to actually help understand those. So we generally look at it from a size of business where we say, right, if you're a very small business, we'd actually recommend just get the AI tools off the shelf. So something like a copilot, you know, a chat GPT where you tick the right boxes to make sure your data is kept private.
Claude as well have a very, very good service. Those ones we typically say for smaller businesses, start with those. And then, you know, if you need to do something more sophisticated, then we can come in and help them as well.
But I think the question really depends on what type of AI. And as I said, the new stuff, the document heavy dealing with unstructured text is a lot more accessible to more businesses. And I think that's why there's so much more interest in AI than what there was a few years ago, just because now it's much more accessible.
A question that I think a lot of listeners will have in their minds. So I'm just going to ask it. It's in terms of, you know, is AI taking jobs? What about job security for safety professionals? And, you know, you mentioned that tacit learning.
You're handing over your knowledge. You're saying, this is how I want you to think. Start thinking like me.
Are we handing over our power? Are we handing over our knowledge? In a sense, you know, I guess it's an ethical question as well. It's such a good one. And it's really interesting because you hear all the hype.
And again, I just saw the news today, Atlassian is saying they're getting nearly 1,600 people for AI. And it does cause a lot of concern and understandably for people. What I can say hand in heart is I very rarely see jobs go as a result of AI.
It's usually a company that's lost its strategy, hasn't got a growth plan, and is using AI as a scapegoat to cover up the fact that they probably haven't got a really solid growth. I'm not saying that specifically about Atlassian. I'm just saying that in general.
What I'm saying is that it's often used for other reasons. The biggest changes are in software. And so when it comes to software development, as I said, the time it takes to code, you know, we've got software developers and we call them AI engineers deliberately because they're not writing code.
They're assessing code. They're looking at it. They're deciding what code to write, but the job has changed.
But we're also seeing as a result that there's going to be a lot more demand because, as I said before, like the build versus buy equation is changing. And so we actually think people will be doing a lot more work than what they were previously in terms of building applications because now if we can build something that fits our needs much better, there's going to be more demand there. So I think it's a shifting space as opposed to just simply jobs going.
In the case of safety, there's obviously one of these things where we're not going to I don't think anyone would go along and say let's just let AI make the decision based on my tacit knowledge without looking at it. And so I think that we'll find that as individuals we'll be shifting our time to the things that take more cognitive load where we can be more thoughtful. How you actually conduct a quality interview, you know, rather than just saying I've got five interviews to smash through in an hour or maybe in a day or whatever it might be.
Being able to maybe have the time to go let's make sure we really think these things through. And I'm not suggesting that that's not already done, but maybe it's a very simple example of saying, right, we can fine tune the art of what we actually do to get that as good as possible because that's where the value is that we're providing. And then there's the critical thinking and looking at things subjectively.
So I do think that although we're not seeing it yet, I do feel there will be a bit of a renaissance for wanting a better word in some of the humanities and the arts because I think that we will it won't be replacing and saying it suddenly stems no longer relevant at all. But I think that we are going to find that we will put a lot more time into thinking critically and looking for shades of grey, which I think would be very familiar in this industry, you know, when there's not a clear cut answer. And so I think people that can see the world in shades of grey are able to leverage AI to help do the stuff that maybe is something they wouldn't probably be wanting to do but can spend better or more effective time on the things that actually matter, you know, rather than writing the report, you know, communicating the report and making sure people understand exactly why it's been written the way it's been done and what the findings are.
I feel that and even just making the report better. Like if you had one hour to write a report, for argument's sake, you're not going to get much done beyond just putting the words on paper. But if you're speaking to the computer, which is the way we work now, and I'm seeing it generate the report and I'm doing maybe four or five iterations of the report over that one hour, I'm getting a much, much better quality outcome in the same amount of time.
So I don't see that there's going to be a mass displacement of work. I wouldn't be in this business if I thought that was the case because ethics drives me as much as anything else in that regard. You know, we're onshore, like for example, we're very much focused on, like we're fully onshore for even software development and all that work.
That's a big part of what we do. I think that there's a lot of noise, but then when you peel under the covers, I think that it's probably, it's just a good news story. You know, there'll be certainly some displacement, but I think it'll be displacing as opposed to replacing.
So we'll find that people will move to other things as much as anything else. That's sort of a bit of a myth buster there, but are there any other myths that are out there in AI world that need to be busted, do you think? Look, the biggest one I'd say is, I mean, that's certainly one that your time will tell how accurate I am there, but I feel that I'm probably, you know, at least half right around the jobs. The other one would be around getting your data in order.
It's not to say that it's not worth going down and getting your data in order. And what I mean by that is for reporting and analytics and understanding your, you know, business and trends. That's always something we'd recommend people do, but it's incorrect.
And it's a very big myth to suggest that you need to get your data in order to do AI, because it suggests that AI is one thing and it's not. Generative AI is a new technology that does not need anywhere near the sorts of volumes of data. It needs tacit knowledge.
But in the world of IT, typically we talk about data and we mean databases and very structured data that's in nice rows and columns and very well thought out and very tidy and very clean. That's not a dependency for generative AI. And so the biggest myth I've found is that a lot of organisations are not doing AI because they've had someone tell them that they need to get their data in order first and they get the quote for doing the build and the licensing for building this big data platform and they go, we can't afford it.
And it's like, it's a missed opportunity for them because they're looking in the wrong spot. They should be looking at all the unstructured data and saying, you know, these documents we're trying to make sense of or create, what can we be doing with that? Because every process, you know, we're stopping and we're having to double click that PDF or open that word and edit. Those are the opportunities where you can get some wins.
And so that's the myth I'd probably say I'd want to bust. And again, it's a valuable thing. Don't not get your data in order, but don't do it because someone said you had to do it to get AI.
That's not a correct reason for doing it. It's a nice to have in that regard. Alex, it's been an absolute pleasure to have you on our podcast and talking about AI.
And I want to thank you very, very much for taking the time to do so. I guess just as we close, I was wondering, is there one key takeaway that you feel we should take out of all of this? Yeah, I think just think about AI as a summarisation technology. And although that's not strictly correct, it's the best way to think about it.
So when you're using, you know, a copilot or a chat GPT or a Claude and you're finding it's not working well for you, it's probably because you haven't given enough information. And so if you think about it, you know, give it, you know, potentially pages of information and ask it to summarise it or repurpose it in a way and then give it clear instructions on how you want it to behave and work and examples of how it should respond, you'll find that it'll work a lot better for you. And so that would be my key takeaway is treat AI as a summarisation technology, not as a thinking technology, and it'll help you go a lot further.
Even if someone's not in the safety profession, this podcast is very relevant to every person in the workplace. So I can't see how this wouldn't be useful to everyone. So thank you.
Yeah, pleasure. Yeah, thanks for the opportunity to have a chat. And yeah, look, if you want to find me online, I'm on LinkedIn if people want to find me.
But yeah, look, I'm always happy to have a chat and help people understand it. So yeah, thanks for your time this morning. Thanks for listening.
If you haven't subscribed, please do so and join us at our monthly Aviation Safety Community online forums. Let's keep in comms.