From crawl to run: The reality of AI implementation in pharma with Arthur Alston

Future pharma host, Romain Bonjean, CEO of RoseRx, sits down with Arthur Alston, Head of Solid Tumours and Radioligand Therapy at Novartis Australia and New Zealand, who brings over 25 years of life sciences experience as both a trained physician and commercial leader. 

They explore Arthur's "Crawl, Walk, Run" framework for AI adoption, why most pharma companies are still crawling despite buying licenses and the brutal realities of moving from enthusiasm to actual implementation.

Arthur shares practical insights on building AI agents across multiple functions, the limitations of RAG architecture in pharma, navigating regulatory compliance in pharmacovigilance and where executives are actually seeing ROI.

He also addresses the biggest mistakes companies make, why company policy might be your biggest barrier, and where AI can genuinely reduce administrative burden to help healthcare professionals focus on patient care.

Transcript

SPEAKER_03 0:00

If you give me an hour of your time, you will know more about AI than anyone else in the room. Joined by Arthur Alston. He has over 25 years in life science industry and he's one of those rare types with experience in the clinical and commercial side of healthcare. In the past couple of years, Arthur has become an authority on AI in pharma, speaking at conferences and engaging deeply with the technology to understand how it can be best leveraged.

SPEAKER_00 0:23

I realize that in AI, it's useful to think about crawling and walking and running. It's not a linear curve, it's an exponential J-shaped curve in what you can do with the technology. And most companies are somewhere between crawling and taking their first steps.

SPEAKER_03 0:40

I agree.

SPEAKER_00 0:40

But then when it comes to running, you suddenly are replicating vast portions of someone's job within seconds.

SPEAKER_03 0:53

Hello, I'm Roman Bongin, CEO of Rosarex, and this is Future Farmer. On this podcast series, I explore the fast-moving evolution of life sciences, technology, and the evolving patient and healthcare professional experience. I'll be talking to industry leaders who are looking forward toward the future and understand the challenges and opportunities that lie ahead. This series is brought to you by Rosarex. The future of healthcare lies in truly personalized patient and healthcare professional journeys. And conversations like the one we're about to have today will help us all get there faster. I'm very excited for today's conversations as I'm joined by Arthur Alston. Arthur is the medical head at Novartis Australia in New Zealand. He has over 25 years in life science industry and is one of those rare types with experience in the clinical and commercial side of healthcare. Previously, the head of medical affair at Beijing and medical director at Merck, his expertise is deeply rooted in oncology, hematology, immunology, infectious and rare diseases. In the past couple of years, Arthur has become an authority on AI in Pharma, speaking at conferences and engaging deeply with the technology to understand how it can be best leveraged. Arthur, thanks for joining us on Fisher Pharma.

SPEAKER_00 2:05

Thank you, Romain. Great to be here.

SPEAKER_03 2:07

Now you're a trained physician with an MBAE, and you're now the callface of AI in Pharma. That's an unusual parkour. Could you um tell us more about how you got there?

SPEAKER_00 2:21

Um Romaine's it's been 20 years for me in Pharma. So it's that's pretty much my career. But um before that, I worked for six years as a physician. So I decided to do an MBA to give me a launch pad into the world of business, you know, that uh naively I thought existed. Um and that led me to pharma. Um, I first joined Roche, which is one of the big pharma companies, and with them I eventually had seven different roles in three countries, and that sort of diversity is one of the things that I absolutely love about the pharma industries, the ability to do different jobs, to move around, um, and to experience new things. And AI was one of those new things. So three years ago, um, when that AI moment happened, the Chat GPT moment, I'm sure we'll come back to it, because for for for at least the last 15 or 20 years, I've been one of those tech technology-interested people. When when with one of my very first paychecks as a young doctor, I bought myself a tangerine-flavored iMac. Yes. And so I became a big fan of Steve Jobs, a big fan of what Apple was doing, and that led almost to a lifelong interest in technology. So I've weaved in tech into my career at pretty much every single stage of my career. So, you know, it was quite a natural step to become aware of it. And once I became aware of it, I could not let it go. And so you mentioned now being at the coalface, and we'll get there in a minute. But a little bit like Roche, which is the first company I worked for, Swiss farmer. I now work for one of the other big Swiss pharma companies. I then started introducing AI at this company where I'm working, and that's led me to um do a couple of speaking engagements on the topic, which I love doing, and that's how we met eventually, which was at the next farmer summit. So it's really a a combination of my work, uh, diversity, finding new things, being in love with technology for almost all my life. And then also, I really like using tools that make my life easy.

SPEAKER_03 4:44

You're very good at um highlighting what works and doesn't work in AI. Could you expand a little bit more on this? On your journey and what you've seen being hype and and and what you've seen being real.

SPEAKER_00 4:56

I uh you know, there's all these expressions, the graveyard, there are graveyards full of unused ideas, and farmers the same. We try new things, we become obsessed with them, and then we cast them aside when they don't work. So um with AI, what I realized was that I needed to always think about the use case for AI when I when I think about it and when I talk about it, um, and also at the company I work at. So if I can tell a little story. So I started there roughly two years ago. I was already at that point deeply embedded in the world of Chat GPT in my personal life. So when I joined the company, I thought, okay, this is one of the big prestigious pharma companies. Everybody will already be using AI. And to my surprise, I realized that it was not the case. One of the sort of the tricks that I used with this was to understand how it can be used, do it myself, and then show other people the same thing. Rather than just being almost blindly um enthusiastic about AI, I would explain to people you can use it to read your emails and give you a summary of all of the outstanding action items, and this is how you do it. Or you can use Teams to record your meeting, and then it will give you a summary and it will give you all your action items. So despite the fact that the company had licenses at that point for copilot, so copilot is the big Microsoft AI implementation. I think many big companies will use copilot. I was in meetings, I would see people's slide decks, I would see their PowerPoint open. Sometimes I would see their Outlook open, you know, in a meeting, and there would be this copilot icon, but no one actually clicked on the icon. So it was really interesting. So what I would do is I would give people very simple examples of how to do it, but always do it in a way that was non-confrontational, you know, which for a South African and a Frenchman is quite difficult because sometimes we can sound quite pushing, you know. So I I had to be quite diplomatic about this. Because at the same time, in within companies, and I think when you implement anything new, is you don't want to show people up either. So that's what I did. So I just slowed, I literally, it was baby steps in the beginning, and then there was one specific moment where I learned that the general manager at the time was going to go to Europe for a big general manager's meeting, and I took a sort of a chance, I said, because I heard that AI was on that agenda of that meeting. I can't even remember how I knew that. And so I thought if I were a general manager and I go to a meeting with all my peers and AI is on the agenda, I would like to know something about it. So I I sent so I thought about myself as someone in sales. So I sent him a note saying, I know you're talking about AI next week. If you give me an hour of your time, you will know more about AI than anyone else in the room. So within five minutes, I had a meeting with him, and usually his calendar was booked up for six months at a time. To that general manager's credit. So I then spent some time, I thought carefully about what to say to him, and then presented him and got him so excited about it, he went to that meeting, and at that meeting, he he spoke to the right people and then got a copilot license for every single person in Australia. So we were the only company, uh, the only um Novartis affiliate in the world where everybody had a license for co-pilot. It was very much a baby steps, baby steps, baby steps approach. And even now, today, um, when I talk about it, for example, at the next pharma summit, I tend to always imagine my grandmother sitting in front of me in the audience. And one of the things we do in my line of business, we talk a lot to audiences. So whether it's a small team, whether it's in a leadership team, or the whole company, or customers, or or at external meetings. So we do a lot of presentations, and I always think about presenting something in a way that my grandmother in the front row can understand. And I think one of the mistakes that people tend to make when it comes to implementation of technology is to get too stuck in the on the technical side of things and too stuck on using language that actually excludes the audience from understanding what they're talking about. You know, so so making it very, very simple. I remember even in the early days, I actually, when I spoke about chat GPT, I would actually say to people, generative pre-trained transformer, and I would explain what all three of those things mean, and then you could actually see them understanding. Ah, so now I understand what this chatbot does. It generates new information based on old based on old information, and it transforms that into something new, you know. So even something as simple as that, I think works very, very well. And at that audience that day, I did the same thing. I I kind of thought about AI in the language that the audience would understand. And it's a farmer audience, so I thought about what are the key, what are some key concepts that the farmer audience knows, and how can I explain AI in those in those terms?

SPEAKER_03 10:37

It was actually quite fascinating in the conference. It's the level of maturity that was from Australian executives around the urgency of understanding AI and deploying it in an organization. That took us by surprise, because obviously we've been in the market for a couple of years now, and I'm trying to reconcile that soft and steady approach that you're talking with the feeling of there's an existential speed by which it's coming. Like the even in the last few weeks, we've seen an explosion of new user interface, and you can't escape the feeling that hey, are we in a bubble or is it really the adoption just exploding? And then you look at the data by which it's being adopted, it is just phenomenal. Like the in history, we have yet to see a faster adoption of user interface. We've yet to see a faster um in from a business perspective, um, uh growth in in ARR and revenue. And it's going it's just going gangbuster. So you're trying to reconcile that slow and steady approach that you want to see in in traditional pharma and accompany the team with this urge and the the the French and South Africa going, wake up guys, you really have to uh you really have to move because it's um it's exploding.

SPEAKER_00 11:55

I think that alone is a topic we can spend hours talking about. So my my thoughts on that is that when it comes to implementation of new things in companies, it's typically driven from head office. So if we think about let's say something like um, you know, dashboards, right? Dashboards and the implementation of dashboards came from the head office. There's one example of the introduction of a technology that came from consumers in the last 10 or 15 years, and that was the iPhone. So when the iPhone was introduced, we had Blackberries, and Blackberries was a typical corporate tool, and only really corporate people use Blackberries. And then suddenly what happened is people, consumers like you and I started using iPhones, and they brought their iPhones into the work, and they then started demanding access to their emails on their iPhones. And so for the first time, we added example of where a technology was introduced into corporate coming from almost the ground level and from a consumer technology coming into a corporate environment.

SPEAKER_03 13:05

And I think we're seeing something similar to that now with with AI, yeah, and at an incredible speed and and some serious security issue as well, like a uh, for example, uh cloud work we can install on your uh browser essentially, so completely bypassing security, uh corporate security, and it literally will do tasks and work using your internal system for you. Um so we we've seen this onslaught from a user interface uh perspective uh taking over work um and the adoption is is is very phenomenal. And we we use that as a as a sales tool essentially when we talk to leadership and saying if if you don't implement it, people will implement it for you. If you don't get that productivity increase, your competitor will have it and and ultimately it'll suffer from your market penetration um perspective. Um going back to um Novoti specifically, so you you've managed to have this huge adoption of Copilot, which is a very helpful tool. Um, what other AI initiative um have you managed to push through?

SPEAKER_00 14:12

So within my function, which I which is medical affairs, this technology improves so quickly that it a lot it gives us opportunities to think long term when it comes to uh implementing new ideas and new new things. So, what I mean by that is that in using an analogy which I realized that in AI it's useful to think about crawling and walking and running because this is not a single tool. If you think about ChatGPT right now, we started, I think the first public version was 2.5. Then, of course, there was three and then 3.5 and then 4 and then 4.0, and now we're in 5.2. But with every new model, it's not like a car, which is you know a slight improvement, but it's still the same car. They basically introduced, I would say, if I think about the car analogy, they introduced a bicycle, then the next time they introduced an SUV, and the next time they introduced an aeroplane. So we as users suddenly had access to incredibly powerful tools. However, what I found would work was to think about people along that continuum, but never underestimate the importance of crawling before you can walk and run. So what I did often was to just have, we we call them lunch and learns, or we had team meetings, and just go over the basics of AI and how to use it, how to open it up on your on your computer or your phone, what a prompt is, what a good prompt is, what a bad prompt is, why it hallucinates, how you can import documents, and then let people use that. And then at a next stage introduce reasoning models when they were introduced. And reasoning model, of course, allows you to do much more complex tasks than a single shot prompt. And then finally, I think at this point in time, I would say the run in my analogy is agentic AI. Yep, crawling is about increasing personal productivity, it's helping you with your email, it's simple things like capturing notes for you, it captures your minutes in your meetings. I mean, that on its own is fantastically useful. It's already worth the price of admission. Then when we start thinking about walking, is when you start using it for project-related work. And a very good example here is deep research. So deep research allows you to generate a report that would have taken an agency or market research or a person weeks, if not months, to do properly. And I have an example of that I can share in a minute. But then when it comes to running, when you start introducing agentic AI, you suddenly are replicating vast portions of someone's job. So you go from simple productivity gains to being able to vastly improve the speed with which projects are done, like research reports, to actually being able to replicate someone's big portions of someone's work within seconds. So you can see almost how the bicycle, the walk-crawl, run, you know, it's not a linear curve, it's uh it's an exponential J-shaped curve in the in what you can do with the technology. And the the amazing thing is, we don't even have to do anything to get the new technology. As OpenAI releases the new models, we just get that automatically. You open your phone and suddenly there's a new model there. The I think though, what this touches on, and we can come back to that analogy in a minute, is I think this can be overwhelming for a lot of people. I agree. And this brings us to again a whole new topic, which I think can be some sort of AI fatigue, which might set in at some point, or something along the lines of, you know what, I already use AI for my meeting minutes. That's enough. Yeah.

SPEAKER_03 18:28

AI light, I call it.

SPEAKER_00 18:30

Yeah, so so so it opens up so many questions. So, what is the role now of management and leaders or people managers like me? I think I think part of it is actually for you to stay on top of this to some extent so that you understand the capabilities of this technology, so that you can have use cases so that the people that are in your sphere of influence can use the technology. Yeah, so but but what I've what I've learned, Romaine, and I think you'll we've seen this in all walks of life, all spheres of life, is there are early adopters, yeah, but there are laggards, and there are the sort of the middle ground, right? So I think companies want to find the early adopters, they want to give them freedom and flexibility to do things, and then they want them to influence the the sort of the pack, so to speak. We see that, for example, when we think about introducing new drugs within the market. You know, we think about who the early adopters are, we think about the Lagos, and we think how we can influence those. And I think it's the same thing with this technology. Not everybody like you and I are going to read about the new Chat GPD 5.2 and understand the difference between 5.2 and 5.1, or 5 and 4, which are massive differences. But most people don't care enough about it, I think. So so when it comes to to driving implementation or uptake or usage, this world around us, this AI world is changing so quickly, I think we can't ignore that, right? So again, I think the use cases therefore are are really important. So, you know, in your work, you define clearly what problem you can solve.

SPEAKER_03 20:19

Yeah, absolutely. And I mean, in our perspective, we we're in the commercialization space, and we we have a we we see an existential threat to pharma that ultimately if you don't communicate with your consumer, be it an HCP or a patient, someone else will, and it'll just make your task harder. Um, and we think uh AI, and to be more precise, uh generative AI is the best tool out there to offer personalized content to your consumer when they need it. And so that's what dictates our existence. That's what we decided to do with with Rosarex and and and quite successfully. But if you're not talking to your end user, someone else will. Ultimately, there's a uh an avalanche tsunami of uh information coming out, some of it's very synthetic. Um, we've seen Chat GPD, I think the numbers are insane. It's just like a third or 40% of Chat GPT is health-related search. Um, we've seen that uh same with Claude, very similarly, um, and that ultimately the choice will have to be made. Like if the life science industry are gonna become a service company and service the end users, the consumer, which is an HCP or a patient, or not, in which case they'll fall into the commoditization bucket um and and and cross fingers on approval and access and and negotiation in the back end. But I'm seeing it very starkly in saying you either become and own your end users, or someone else will probably big tech or probably ChatGPT or Cloud or you know it. Um But you have a chance to fight that because the magic of AI is its hypercommoditization. The speed by which you can use it within your proprietary data, considering that you are um the owner of the RD, the IP owner. You've got an edge there and you can you can push out. It's it's very striking when you when you talk to the US, people start to really grasp that, saying, we either fight or someone else will take that user for it. One of the things I've always had a problem is AI is a very wide word. And so it suffers from a definition issue. Like if you don't get an eye roll, you get people thinking immediately, oh, that's strategy PD, or it's ML and genomics, or it's you know, um AGI, or you name it, like, which is this idea that you're gonna have a super intelligent being out of out of system. Going back on your your crawl, walk, run analogy, where are most pharma at the moment? Are they crawling or are we starting to see some running?

SPEAKER_00 22:59

I I can only talk from my very personal perspective on this. So I I I imagine that if you read articles and you go to different conferences and you know talk to different executives like what you do, everybody will probably have a slightly different answer on this one. My view is that as an industry, we are waking up, like I think all industries are. Um we we understand that if we don't get onto this train, it is going to leave without us, like all industries. Um we we also are starting to see that this is not a flash in the pan, like maybe um some of the technologies have been in the past. It feels like this thing is here to stay, and it feels like it is transforming everything around us. You spoke earlier about this existential threat kind of feel. I think if you ask a question about what's keeping very senior executives awake at night, I believe that will be in the top three things, if not maybe the top two things. What do we do about this? Yeah. So then to come back to answer specifically your question. So a lot of the part of the answer that I can give to you now also comes from my observation from talk from talking to my peers in many other companies. Um and my sense is that most people are are crawling right now. And the reason I say that is because everybody I talk to seems to have a corporate AI um implementation in the form of something like co-pilot or maybe Gemini for some companies. But then the interesting thing, Romania, is how how how do how do companies measure this? And the truth is they can only see when you open up your your copilot instance. They can't actually see what you use it for. So at the moment, companies have got a very crude understanding of what people use it for. So so, for example, companies will tell you we have 90% of our employees have a license, and in the last month, 50% of our employees use Copilot.

SPEAKER_03 25:18

Yeah, that's that's definitely crawling.

SPEAKER_00 25:19

So even companies don't yet, I think, understand how their employees are using it beyond very, very crude measurements like this. So, like anything else, I think there are early adopters. Um, and the early adopters, I believe, are certainly jumping on board this. And beyond the sort of beyond the what I what we can think of as personal productivity gains, which is your walk crawling, I really think, using Copilot, like we said, for emails and meetings and all that stuff. I I don't think people have yet gotten onto how powerful the reasoning agents are, the reasoning models, you know, for using things like deep research, for example. Um, I I and then certainly I know, but when it comes to running, which in our analogy refers to using agentic AI, almost nobody's using that yet. Yet. If I had to give you a specific answer, I would say it's just a personal view, most companies are somewhere between crawling and taking their first steps.

SPEAKER_03 26:25

The first steps, nice.

SPEAKER_00 26:27

That's kind of where I would put most companies. But I do know at the same time, from what I've read, there are some companies who are like always, like anything in life, who are more progressive. Moderna is a company that springs to mind. So Moderna had a uh a collaborative agreement with OpenAI from very early stage, and we know that they benefited from the dramatically, that it's public knowledge, they talk about it openly. So there we know that, for example, they already have hundreds of agents running in Moderna.

SPEAKER_03 26:56

So again, I think the distribution there's Stefan, uh Stefan Bensol always had a very good understanding of existential threat from very early on. Yes.

SPEAKER_00 27:06

So so so there are companies who I believe who are way ahead of the pack. And I I would dare say, I would suggest based on just what I've read, I think that Moderna is one of those companies.

SPEAKER_03 27:19

It's fascinating because we see software company or a technology company where AI and agent specifically have taken 70 to 80 percent of the productivity in the last 12 months. Certainly for us, um, we went from having 20 engineers to literally working with four engineers, but now the pyramid is inverse. We have a lot more product designer uh versus whereas you used to have one product designer for five devs, you know, turning that around and and and so could you explain that a bit for the audience why why specifically the group of in a software in a software design world, you you have to productize. So you you you think of a product manager and a product designer to think about what the software is going to do and what's gonna be the function, and you have to define it. And you have a series of tickets that you have to define. So if this happened, that happened. Then you give those tickets to devs, which will code those functions. The speed by which the code is happening with modern tooling, such as cloud code, or even you know, you name it, is such that your bottleneck is not the devs anymore. It is the definition of your cards, is the definition of your product. So whereas you used to have one product manager and then you'll have a series of devs taking out tickets and finishing them up over a matter of days, sometimes weeks, you know, have I mean, we jokingly the other day, we I think we we swallowed 343 tickets in one day, just gone how long would that have taken just in the past? Oh, that was months, months, months and months, month of work. And um, and you you make them into two, three days workload. Yeah, but the speed by which and the cleanliness of the code, we've replaced if from a technology standpoint, in in the process, we've replaced every single human interaction and made them superior by agent. It's still humanly supervised, the team still owns the code, so it's not like a uh a complete job replacement, it's an enormous productivity increase. Um, and it's fascinating because now the good old designer, the prompt engineer essentially, becomes extremely valuable. Of like, let's define the problem really well. Um, things like uh a tester, we used to have two types of testing automatic brute force testing, and used to have manual testing. People would click through and pretend to be a grandmother and see if it breaks. And both of them have been completely virtualized through agent and do it in seconds. Like you immediately have a report, this broke this that there's a weakness there, and it then self-correct itself. And that's progress we've seen. When I'm using the word existential, I'm not trying to be provocative. In the last three months, we've seen like an enormous jump in productivity. In the last year, it's like you think it was another century. Um, and you either decide to adopt it, and then we're gonna use it as a competitive advantage. And the earlier you adopt your stack, the faster you move, or you have a degree of cynicism, but I guarantee you you're gonna be a legacy player very quickly. Like it used to take. I mean, I actually had uh a Salesforce presentation talking about it. It used to take 10-15 years to become a legacy player, where you had to reconsider your entire software stack and rewrite it to use. I think now you're probably looking at six to twelve months. That's how quickly your reinvention and your thinking. And so, meanwhile, you're talking to Pharma that gives you a 12 or 24 months horizon to do a pilot out, and and you just can't help but think you really have to move faster. I I I keep on going back on this existential fear.

SPEAKER_00 31:10

So, Roman, so the so so what I think when I hear that, and I I I read and I know this is happening. I know that the productivity gains when it comes to software and software engineering is just simply astonishing, like you just said. So the question then does become, and not just some from someone in pharma, but every industry, what what what part of my industry is going to suffer from or benefit from that phenomenon that you just described?

SPEAKER_03 31:39

So I actually I I'm an ethanol enthusiast, I think we will all benefit from it. I don't see it as a job destroyer or anything, because fundamentally you'll still need to be authentic and you still need to have a human amplification but a human touch in all your interaction. Um, but I do think a lot of the minute task, a lot of the um will will will get done by machine. And and if you look at the pharma process, uh which you're well expert in, there is an enormous amount of administration, back and forth, MLR reviews, checking, viva file update, like a whole series of workflows, all the way from RD to commercialization and patient support, which would benefit enormously from having an agentic mindset from A to Z. And all of this is designed for one thing and one thing only, is reallocating resources to better treat patients. And so I don't I don't look at it as a destroyer of job, I look at it as a massive enhancer of what you do, which is making people have better health outcomes. But but going back on Novartis, you implemented agent across multiple functions, which is when you think about it, extremely novel for Pharma. Could you give a little bit more details about how those works?

SPEAKER_00 33:01

So I can't remember exactly where my first introduction to agents began. But what I realized at one point was that within the world that I work in, which is a highly regulated world, we have very well-defined data sets, and we need to refer to those data sets for answers. So, what I mean by that, so an example of that is in Australia, we've got something called the Medicines Australia Code of Conduct. That tells us how we conduct ourselves in our industry. Everything from what we can say and can't say, how big the font must be on an ad, you know, um, whether a salesperson can go with a medical person to a meeting, what their different roles must be. And all of this, the the sort of the colloquial term for this is compliance. Uh is what you are doing today remain, is it compliant? So it becomes a part of the vernacular within pharma companies. And I if I had a dollar for every time that I heard that question, is this compliant? I wouldn't have to work anymore. So it's really at the center of everything that we did. I'm just talking about Australia now, but I know all other countries except the US have similar. So what we do in our pharma world is what we think of it as a highly regulated workflows, right? And so maybe I should say first, maybe we should just first define what what are the three requirements for an agent. So basically, an agent you you can think of as an agent as a standalone tool that can do work that is trained on something. So it is usually built for a specific purpose, whereas Chat GPT is just a very broad thing, it can do anything that you ask it for. So in this instance, the three elements, I had a very defined data set, which is the reference materials, which is all the code of conduct information. Then I needed what is called a set of instructions. And a set of instructions basically, as the word says, it tells that agent what it is and what it must do, and what sort of information it's going to receive, and where it can go and look for the answers, and what sort of answers it must provide. So luckily, um my brother, who's like you, he's in the AI world, he's a long-term tech guy, said, Because I started then writing the instructions, and I realized that my bottleneck then was to write the instructions because I had no idea. I mean, I've I've never used an agent before. And he said to me, Why don't you just ask Chat GPT to write the instructions? Oh my goodness. So this is like using AI to build AI, and that's a little bit like what your tech, what your what your um what your coders are doing. They basically are using AI now to code. To code, right? Right? It's doing the hard work for them. So in this case, my my simple analogy was I used AI to write the instructions so that AI could become an agent. The agent is just as good as what your instructions are. So to come back to then to code buddy, so code buddy then and I built this on my own basically. I think even before not I think before I joined Novata. So it was something I built. It basically now, and I joke and I say, I finally have an answer for is this compliant? Yeah. So any question I get that's related to compliance, I use code body for. The beautiful thing about it is it's always right. And this now comes back to something we need to talk about, also, which is slightly technical, which is called retrieval augmented generation. So what so how it so we know that there are all these famous examples of you ask ChatGPT a question and it gives you a beautiful answer, but the answer is wrong, and that's called hallucination because it guesses what the answer must be. But now to overcome hallucinations, you can do a couple of things. One is you can become very good at prompting, because the better your prompts are, the the clearer your instructions to the co-pilot, the better the answer usually is. You can become better at recognizing hallucinations, or you can use rag, retrieval augmented general. In other words, you ask AI, the AI, to only go and look in a certain place for the answer. So it can't hallucinate because that's essentially what we do.

SPEAKER_03 37:34

Yeah, we we we we we uh we literally ingest the relevant document. But but you said you use Quote Body within Novartis. Uh how much has it changed Novartis's workflow? Is it being used regularly?

SPEAKER_00 37:49

Um it's hard to to quantify that, but what I will tell you is when people don't have it, they miss it. Of course. So so and the reason I can say that is because it comes to as the organization has evolved its understanding of AI, its policies have changed. And I think most companies probably went through this. It's I think it's just completely normal. So it's not a criticism. I think that is just the response of the organization to this rapidly changing environment. And so one of the things that happened, for example, was um they changed the policy when it came to agents. So at one point, I could share my agents freely within anyone within the organization, and then a policy change came, and then I could only share it, I had to know, I had to put someone's email address in, and then they also had to have access to my reference documents. And that happened without me knowing about it. So overnight I suddenly got all these questions what happened, what's happened to Code Buddy, what's happened to Code Buddy? And the reason that they lost access was because of the policy change. So so that for me meant people were using it. Then then, but what I can say to you also is the be one of the beautiful things about it is, and that's part of the instructions, is I've asked it to always the answers, must always be informative and educational. So it doesn't just blurt out the answer and that's it. It will always refer to the relevant section of the code, and it will always give answers in an educational way. So the experience of using that agent is also a very pleasant experience. And as you know, ChatGPT tends to want to please, right? We talk about personalities of agents. So I've also you know included in the instructions that it must ask open-ended questions at the end of so that you you continue the discussion with, you know. So um then the other thing about it is is the speed with it gives with with with which answers come, of course. And then I think what's really, really important is that it does not choose sides. So the reason I say why that's important is within pharma, you often find that the commercial guys want to push more than what compliance or metafares or regulatory fares wants to go. In other words, there's this tension that exists. And the let's say, for example, marketing wants to, you know, paint the paint, they want to have a green ad, and we say no, it has to be blue, and because green is not compliant. The marketing guys will make half of it green and half of it blue, you know, this sort of thing. But the beautiful thing about code buddy, it give it it does not choose sides, so the answer is completely neutral, and therefore it immediately becomes valuable, more valuable than an opinion from a person, if that makes sense. I agree. So, so so that I found this was interesting because I started hearing from colleagues that it was in fact speeding up cross-functional team meetings dramatically because they would have one answer, and everyone would everybody the answer would simply be so good that everybody would agree and everybody would move forward without answer. And it also settled arguments. I love this.

SPEAKER_03 41:15

And so you you you could nearly quantify it if you had access to the data in terms of of productivity gain and and and ROI on it.

SPEAKER_00 41:22

Look, you could, but that comes back to how the company measures the stuff and stuff, and and you know, the uptake of it. And and I I simply don't know that information. But what I can say to you, Romain, is it's one of these tools that once you use it, you can't imagine not having it. So it immediately becomes indispensable. And I think that's a sign of a really good tool. Yeah. And then what happened, if I can spend a little bit more time on this, I then realized that the sort of the architecture of Code Buddy allows me to think about other problems in a not dissimilar way. So, what I mean by that was because our the Medicines Australia Code of Conduct is embedded across everything that we do, and the regulations are embedded across everything, that there are different applications of this code. And so one of them you mentioned earlier, which is medical legal review. So, what we need to do, all of our materials, our promotional, our ads, they need to be approved for use with our doctors or or patients. This is called MLR, medical legal review. It's a process that everybody does. And this is a process that can take weeks and sometimes even months to do because um of the intricacy sometimes and the tension, as I said, between what some people want to do and what some people don't want to do. And what I realized was that MLR is simply an application of Code Buddy. Correct. So what I did was I then sat down, went back to my other AI and described what I want to do, and then asked it to write a set of instructions, and then that became my code reviewer agent. So so the code reviewer now, and that would be something good to quantify, because a a big part, a major part of what we do in head office farmer roles, talking about productivity and and and work that might be eroded, is actually MLR reviews. Some companies even have committees who do this stuff, and there's a lot of toing and throwing. There's platforms involved like promomats and all of this. What I do now, I just run it through my my promo reviewer, and within within seconds, it'll flag with an answer. Um, and again, the same thing I've trained it so that it it it will give it will refer back to certain sections of the code. And I've also trained it to give to make suggestions for what would be appropriate. The third agent, actually, it was the fourth agent I created was we we've got a function called Medica. Information. So within pharma, we all need to be experts on our products. That's a basic part of our job. And we have a dedicated function so that if you use a Novartis product and you have a side effect, you can call someone or email them, you can go to your doctor, and the doctor can do the same. Every company must have a medical information function because that's part of our therapeutic goods agency agreement that we have that in place. So to come back to the medical info, so with so I moved from one function to another function and I started working in a new or a therapeutic area I knew nothing about. And that would usually take me months, maybe even sometimes years, to learn everything. So what I did was I got all the reference materials of the new therapeutic area and put it into a SharePoint folder, and then I wrote a set of instructions for an agent so that that agent is then a medical information agent. So what I can do then is I I simply, if I have any question about the therapeutic area, any question about a drug, even our competitors I put in there. So I can ask any question that I could possibly have, and within seconds I've got a perfect answer. So think about the applications for this. So as I said, we all need to be experts, but suddenly you can become an expert within seconds instead of months. Or you if you're a salesperson, they go through a long period of training. It can be six weeks to eight weeks that they get trained on the information. They do tests and then they get certified and they can go out in the field and talk about it. Suddenly they can become an expert. They can just be an expert instantaneously if they've got access to this, to this, to this agent, to an agent like that. Within head office, I constantly have questions myself. How does a product work? What about this? What about that? That's just normal. I just pull it up and I've got an immediate answer. Then you can flip it around and you can say, I'm going to go and see Dr. Romain this afternoon. I want to talk to him about that product and that indication. Give me three talking points that I can talk to him about. So now to pack all this all of this together. So recently, um, someone in my team who was a deep expert in a subject matter expert went on parental leave for a year. And all of a sudden, I had this massive gap in my team. I didn't know that stuff. Yeah. We also realized we could we didn't we couldn't hire someone, and there was no one that we could promote into this role in quick quickly, and they would also have to be trained up. So what we what I realized was I could actually use these three agents or four agents. So I could bring someone in from outside the country and give them code buddy, yeah, and they would immediately understand the Australian complexities of our MA code. Then I could give them access to a promo reviewer, and they could just simply run all of the MLR work, which would be a core part of their job through promo reviewer and immediately have an objective and correct answer. And then I could give them access to the medical information agent, even in a therapeutic area that they did not understand, and immediately they could be an expert in it. So, Romain, this means that you can bring, so in this case, it's someone from Korea who's already at a medical lead level, in other words, sort of mid-senior level. They're already an expert, they know how pharma works, yeah, but I could offer them a th a two-day secondment in Australia without the burden, and it is a burden, of onboarding them over several weeks or months. That's an enormous cost saving for the enterprise. So, this is something I'm gonna test in the next month or two is to see whether I can bring someone in from another part of the world and embed them, and within a day they can become a highly performing team member instead of it taking months. So it so that's just by combining these agents together, and again, there's a use case that I don't I don't think anybody has ever thought about. I just it just sort of came to me that why don't I try it?

SPEAKER_03 48:21

So essentially, you you're a your suite of agent allow everyone that's competent in pharma to become immediate subject matter expert. That that's fascinating, and surely that that that's got to be quantifiable.

SPEAKER_00 48:34

I can see you really want to quantify the. I do, I do I do. So if you want to quantify something, hiring someone, training them up, um helping them in the first three or four, five, six months, let's say it's a$200,000 salary, yeah, you know, you can quantify that and the and the hiring and the time it takes, you could you could probably put a number of$200,000 on something like that. That if you use these agents in an onboarding capacity, you could you could that person would be up and running within a day. Fascinating. And the other the other aspect of this that I think is important, um I think it's obvious, but still worth stating, is I believe AI makes experts deep experts. And building on that, I actually think that um my observation is that people who are already experts benefit incrementally more from AI. So, what do I mean by that? So, so in my case, I'm I I I understand pharma, I understand, let's say, the cardio, I understand how things work, but then I got this AI agent, I called it Ask Emma for the medical information in in memory of someone who was actually called Emma, who was a deep expert in that subject matter. But it it gave me immediate capabilities of being a cardiovascular subject matter expert, which would have taken me maybe a year or two. But because I already understand the world and the the industry and and what I do, I know what questions to ask and not ask. I know what the information means and I can make sense of it, and I can use it to like that. So I think the analogy in the sort of the real world is if you give a very good carpenter access to a brand new saw that can now saw wood that in the past was not you was not able to saw, he can make things with that new saw that no one else could make. But if you gave that same saw to someone in the street, they probably would not know the difference and they would not be able to use it like that. So I think there's an element there, and this goes back to this sort of cliche that we that that that we that we have heard now for years, that AI is coming for our jobs, but it's not AI that we need to be afraid of, it's people who who are better at AI than what you are. They're gonna come for your job. So so recently we were asked to do a collaborative project with um with a company of researchers. We had very senior people in the room. I was the only one who thought about using deep research to explain the problem, upload a couple of different documents, and ask it for an idea that we can use as a collaboration, as a starting point. It produced an incredibly good proposal. The point is, in that meeting, we had five or six very, very senior people. Everybody has access to Copart, everybody's got access to deep research, but no one thought of using it. And then someone asked me, Arthur, did you do it or did AI do it? And then I thought about it for a second, and so tongue in cheek, I said I said to him, Does it make a difference? Of course. Because I think I think I think there's there's a little bit of a kind of a negative view. Oh, he used AI to write that report. But the point is, you still have to do that. I I I had I I had to I had to know that I could what prompt to use. I I had to I thought very deeply about it, and I think I had a couple of different versions, and also what some what reference material to use. So it's not just you don't just open up Outlook and go to Copilot and write give me a research report. So it's a little bit of a dismissive approach, actually, that oh, he used AI to do that work. So I think we're gonna get to a point where hopefully we move beyond that.

SPEAKER_03 52:43

It it's fascinating because um my daughters are in public school and they embrace AI for the exact same reason. Uh, instead of saying do not use AI, use your own research, they're actually starting to train at school in New South Wales how to use AI intelligently. And still I I walked in thinking for sure they're gonna ban AI completely at school, and they're already changing the program. So it becomes a skill, it becomes a skill and how to use it, it becomes a tool, same as computers or mobile phones or social network when they came out.

SPEAKER_00 53:15

So to go back to the walk, crawl, walk, run, yeah, and the sort of the approach I took. I I I I I spent a lot of time, I think, teaching people the skill of using this tool, and and um that I think is critical, and I think that will continue to be critical, especially because the technology is evolving so much. So every time it evolves, there's a new skill that needs to be taught, right? So if you think about single shot prompting and as a skill where you must be good at on your prompts and understand prompt engineering and all that sort of stuff, to using a reasoning model, it that's a different skill, and then to using an agent. That is again a different skill. So this is probably also one of those things that companies must think about. It's not just one single time that you teach someone how to use AI.

SPEAKER_03 54:08

No, it's a control.

SPEAKER_00 54:09

Yes, it's never going to end, I I suspect.

SPEAKER_03 54:12

Um this this productivity gain should be incentivized and should be rewarded to accelerate transformation. I um it it it's a big debate. Now, uh going back to your earlier analogy, um, you you've mentioned crawl, walk, run. What's after run?

SPEAKER_00 54:30

You tell me, Romain. You know, that so you said earlier things have changed so much, even in the last three months. So what so let's think about some of the things that we've seen in the last three months. Suddenly now robotics is a thing. Um Tesla, in fact, yesterday announced that they were going to stop making the model S and X's, and they're just gonna focus on making the Optimus robot. And they're and these are not just robots, these are robots with AI. But to back to the analogy, as we said, you know, we think, I think most companies are still between walking and crawling. Yep. So I so so what comes after crawl walk run is walk. I think I think I think I think a lot of companies and people still have to walk. But if you ask me, like in the like a like the future pharma thing, I believe that a pharma company will exist with maybe the same number of people, but that the people will become managers of agents. And we we might have hundreds of agents, if not thousands of agents, doing a lot of the work for us, the menial work, all this pharmacovigilance, the medical information we spoke about, onboarding, training, maybe even customer interactions, all that sort of stuff. It will all be done using agents, I I believe. So, you know, we will we we might even see small pharma companies pop up that have got the reach of a major pharma company because they use agents and use. So maybe crawl, walk, run, fly. Yeah. Who knows? Because it because they as we said earlier, it's it's exponential, right? So our brains are actually not very good at understanding exponentials. So who knows what comes after?

SPEAKER_03 56:18

That that is the thing that that was quite fascinating, and and and you just you see a very exciting future farm out there. Yeah, so uh before we wrap up, I normally ask the guest five critical questions. Uh are you ready? I am so biggest industry prediction for what's to come in 26.

SPEAKER_00 56:35

Uh agentic AI.

SPEAKER_03 56:37

Love it. What industry challenge keeps you up at night?

SPEAKER_00 56:41

The impact of AI on our jobs and how it's going to affect our jobs.

SPEAKER_03 56:47

And uh, where is AI being underutilized in the pharma business?

SPEAKER_00 56:52

I'm going to say everywhere. Love it. Um, so I look, I say tongue-in-cheek everywhere. So if we think about pharma, as you know, it is a massively complex beast of an organization because we've got so many different aspects. A pharma company simplistically uh must research and then develop drugs and then sell them. But they also need to manufacture them, they also need marketing, they need sales, they need HR people, they need finance, they need operations, they need supply chain, they need logistics. I there's no doubt in my mind that every single one of those functions is going to be affected by AI in one way or another. Supply chain is an obvious one. We've spoken about that. Marketing will be, medical affairs will be, medical information, all the regulatory functions. I think sales also definitely got some ideas about that, but then there, but then I think where we will see and where pharma companies are investing is in the research and development. Because if you look at the sort of the expenses, if I can use it simplistically like that, of a pharma company, where does it spend most of its money? Yes. So drugs cost billions of dollars to research and can take years and years to research. So if we can speed that up even by 50%, and you know that with when it comes to AI, we are not talking about 1x or 2x improvements. We're talking about 10x improvements. So imagine if you could 10x a$10 billion development into$100,000, what that would mean for the number of new products that can come to the market, the number of new products that can be fixing ailments, the speed with which things can change. But there are aspects of that also, and I'll stop here, is pharma does not live in isolation. Our I I think our biggest one of the one of the problems that will need to be solved will actually be the relationship between pharma and the regulatory authorities. Because the regulatory authorities are the guys who give us our licenses. The licenses come from a summary of all of our efficacy and safety data that come from all the clinical studies that we do. But it can take up to a year for the regulatory authorities to review all that data and approve the drug. So even if we can 10x our development, the bottleneck.

SPEAKER_03 59:24

You would want to make sure that they 10x their analysis as well.

SPEAKER_00 59:27

Exactly.

SPEAKER_03 59:28

Going back on a fire question, um, if you could give pharma executive one piece of advice about AI, what would it be?

SPEAKER_00 59:36

I would I would encourage everybody to play with it and not to ignore it and not to be afraid of it. And as I said uh before, you get to a point where you look at everything through the lens of AI. And once we get to that point, I I think then we will then we will all be flying.

SPEAKER_03 59:57

I love that. And uh what's the biggest mistake companies make when implementing AI?

SPEAKER_00 1:00:03

Ignoring the old truth that culture eats strategy for breakfast.

SPEAKER_03 1:00:08

I love that. So true. And um, how can AI be used to help the wallabies beat the spring broke? Impossible. Come on.

SPEAKER_00 1:00:20

AI has limits. Very annoying. I think it can always really good to have you on the show.

SPEAKER_03 1:00:28

Thank you so much. And uh yeah, looking forward to uh keeping touch now. If if um if someone on uh one of our listeners wants to get in touch with you, what's the most effective way?

SPEAKER_00 1:00:39

LinkedIn is probably the easiest place to get hold of me. Terrific.

SPEAKER_03 1:00:43

Martha, thank you so much for coming today. And uh looking forward to see you again. Good chat.