When AI Replaces Regulatory Professionals

Start. Hey, good morning, everybody. Welcome to Friday on Focus. As you know, I'm Sean Smith. Publisher of MedTech Leading Voice. A couple of housekeeping announcements. This webinar will be recorded. Hey, Marcus, are you going to share the slides?

Yeah.

Yeah, okay, the slides and the recording, will be available in 24 hours to everyone who registered. You'll get an email from Zoom, and we'll have the link to the slides and the replay recording. So, that's a nice thing. I'd like you to use the chat function vigorously throughout this. Is there anybody who's a first-time caller joining the call for the first time? You know, drop in the, In the chat where you're, where you're calling in from, you know, like. Detroit, or whatever. And so… today, we, have the privilege… Okay, oh, so, upcoming.

announcements. So, next week, We are going to have 3 people speaking from Phmti.org, which is a new association. It's about 4 years old. It's called pharma… pharma and medical device intelligence, and it's comprised of former information professionals from pharma and med tech companies, aka librarians, your best friend in the world. And, you know, they kinda had to… get new identities, and recast themselves from librarians, because librarians went the way of all flesh.

At most of these companies. So, it's a really interesting group. I've known them since, since their previous incarnation at the Special Libraries Association and the speakers are coming from Springer. And there's gonna be Diana Gao from J&J, and a couple other leading manufacturers. So. worth joining, worth understanding, all that. And then the following week, we're gonna have, David Fillmore is gonna be joining from Market Pathways talking about MDUFA. Everybody's favorite subject, MDUFA 6. I think it's 6 or it might be 7. At any rate, MDUFA 6 is here and y'all got to get ready for it. So we're going to have a fun discussion about that. And then

The following week, is… Oh, Lord, never mind. I'll put that in the chat. So, without any further ado, I… wow, 152 people online. So, today's topic, Marcus, when AI replaces regulatory professionals. Oh, by the way, if you want to be promoted to panelists, raise your hand, and Anna will promote you to panelists. Show… right. The prologue. So, a few weeks ago, there have been a number of business headlines about companies saving millions and millions of dollars with their implementation of AI. Santander reported a savings of 430 million pounds. Centrica reported 10 million in direct cost savings by expanding its AI in-house capabilities.

We've seen reports from IBM, Microsoft, Klarna, all of them saying their AI initiatives are saving them tons of money. So that seems clear, right? AI is driving productivity, reducing costs, transforming how work gets done, blah, blah, blah, right? Then, at the same time, there's a different set of headlines. Walmart quietly introduced strict usage caps to prevent AI costs from spiraling out of control. Reports suggest that Uber had exhausted its entire 2026 AI budget by April, forcing executives to reassess whether the promised productivity gains justified the rapidly Growing costs.

So… what's happening. Is AI the greatest productivity revolution in decades, or is it an expensive experiment that many organizations have yet to fully understand? Or are we simply witnessing one of the largest workplace transformations in modern history? Tbd, so AI is no longer a future discussion here. And it's already changing how organizations operate, how people communicate. And so forth. The question is, what's it mean for regulatory affairs? Really. And, how is it going to change it? Which tasks will change first? Which skills will become more valuable? Which

will become less valuable. which roles will evolve. Evolve is a gentle word there. Evolve is… happens… evolution happens slowly, and I don't think that anything's happening real slowly right now. So, to talk about this, and to help us, help us walk through this, we have Marcus Emne from, Huden, CEO, and I've known him for several years. He's, been talking with regulatory professionals, hundreds of regulatory professionals about this exact topic for a long time now. And so his insight into this comes from actual conversations, actual feedback with

people working in the field and what they're concerned about. So I'm going to let him introduce himself. And, yeah. Thanks, Marcus, for joining us today.

Well, thank you, Sean, for a splendid introduction of the domain and where we want to drive this. Presentation and dialogue today because I think one thing that strikes me when you mentioned the initiatives from, was it the Abbott? No, it was Santander and then we had Uber. So there's so many aspects of this and all of it is taking place at the same time everywhere and so forth. And it could be. It could be a little bit, hard to grasp what's going on, and as you said, it doesn't really happen slowly, it happens very fast at the moment. So, yeah, thanks for having me. And you, you mentioned the title of the webinar, regulatory when, when AI.

replaces regulatory affairs professionals, what happens next? That is a little bit of… a little bit provocative, of course, and we will get to that, but there's a… there's a… There's a good point in making that statement, because It is, it is true. But we'll get back to that. It is provocative and it is true, but it's there's a twist. So, before we start, I just want to make one thing clear, and I'm not here to predict the future. I'm here to share what I've observed over nearly 5 years working exclusively with AI for regulatory affairs within life sciences. And after more than 200 conversations with regulatory professionals around the world.

So, and some of these observations really, really surprised me, and some changed my own thinking completely, and today I would like to share a few of them, and also make an announcement that I hope will suit you all very well. So… what are we doing? I'm not here to promote Houdin and the Compliance Studio. I will, however, use it quickly as a reference to, Get us gathered around what is what. Because when we talk about AI. We all have different perceptions of what that is and what it can do for us. I would say if we would run a…

a poll right now. The majority would probably say that you use a chatbot for daily conversations in your professional life and in your private life, but it's chat-based. Or you use an LLM that has a reasoning model on top of it to help you solve problems, explain something maybe, or you just need a second opinion, or you need it to have You know, check a document, maybe there's a new guidance, and you don't have time to read those 80 pages, and you toss it into ChatGPT or what else, and ask for an executive summary. That's how most people use AI today.

And there are pros and cons with that. I'll get to that, because that's one of my key findings. There's… A thing… in an AI chatbot's nature that makes us miss something very important. I'll get back to that. That's a cliffhanger. I'm going to I'm going to reply on that at the end of this presentation. So now you cannot leave. And I, I, it might be in the end, it might be in the, in the middle, so you cannot leave ever, and you cannot stop watching. All right, so… AI, the definition of AI, it's mixed, and something that Sean said, the way I heard it, is a little bit that the future is here already, but it's unevenly distributed amongst us.

Right? And we… our perception of the future is different. And I'm not here to predict it. But let me quickly show you, because a phrase that I will use a lot during today is vertical AI. Some of you might have… an idea about what it is. Some of you might be experienced already using vertical AI tools, for workflows, for certain tasks, in your professional life. Some of you might never ever heard about it. However, since I'm gonna go back to that phrase and use that as an example many times during today, I will quickly, quickly show you, and I'm sorry that this might look like some kind of marketing? It is not. I will, of course, use our platform, Compliance Studio, as an example, because that's the one I know the best.

But it will be very, very quickly. Where do I share from? Is it in the bottom? You're muted.

It's on the bottom.

Yeah, it's very in the centric bottom. Okay, so I have to open my… Systems here, and allow… Google to do so. Allow. It might be so that I drop out for a second, then I will be back in, because I had to change the setting in Chrome, okay?

Okay.

And I'm sorry about that.

So, in the meantime, I remembered who is speaking on the 31st, and that's going to be our very own Andre Dimino, who I believe is on the call today, and he's going to be talking about… electronic medical device engineering and design for manufacturing. So they do engineering and contract manufacturing of electronic medical devices. And many of you may remember that, we did the voting for the coolest thing in New Jersey, and he had that really cool veterinary device that helps old pets Feel younger with, electronic, you know.

Do hickeys going into their body and I have to tell you a personal story that my own 14 year old Labrador Retriever Emily that I had before Frankie had arthritis really badly and we would take her in Washington to get Acupuncture with electrodes attached. to the end of it, and she would run out of there like a puppy every week. That's what convinced me to start doing acupuncture, because, if it works for a Labrador Retriever, it can work for me. Thanks. Welcome back.

Thank you. So now we have the answer to why you look as 45, but you're a little bit older. It's acupuncture.

That's it. That's the reason.

Yeah. Okay. Let's not go there. I know you dislike to have that discussion. Sorry for leaving. And thank you, Sean, for being the entertainment while I was gone changing. As again, I'm just going to show you very quickly. So vertical AI is when you take. I… trained LLM AI model, and put that in the context, and fence that into context, so it can only reason about that context, right? So in this case. The entire idea is to help a company establish a regulatory position. It could be so that you already have your regulations registered for a certain market and the product. Now, you want to run a due diligence on that, making sure that you have what you need.

And that's not that easy. to do it manually. And the trick here is that you give the, the, AI a context, alright? So you define the domain. In this case, I'm gonna take Omnipod Go. And in Sulet, Corporation. It's a medical device. So now I'm telling… I'm here, I only want to deal with this product. I could have chosen my company as well. That's a different… slightly different setup. But here's the thing. Rather than jumping into dialogue with an AI immediately. You dictate and design what you would like to have done, and that's when vertical AI,

evolves. So in this case, I could choose any country on earth. Let's do European Union. Add that. And then I have to add Metadata. that the Studio Agent, and this is the Studio Agent, it's a context-aware regulatory support, and now it knows that we're talking about Omnipod Go on the European market already, and I can add attributes, intended purpose, I'm just going to go fast here and say it's an insulin. I don't know Pump. or I, fetus. One and two. It's a legacy product, The medium risk class, for.

Nothing here. I'm just gonna go through this very, very quickly, but you see what's happening here now. I'm… Providing static information. As a starting point to the. AI model, so that it knows about me. And what what is going on here? Okay, that's it. Now, what happens is that probably you're familiar with having conversations with with Copilot or ChatGPT or Claude. And over time that. That dialogue and the reasoning from the AI starts to drift. from the original topic. Sometimes that is good, because you didn't know where you wanted to take it.

But when you have a clear direction of where you want to go. You need vertical AI because drifting. on a topic is completely dangerous, because then the AI can start to make misinterpretation about what you actually want from the dialogue, and the reasoning model goes blurry, because you went here and there in your conversation, and all of a sudden, the AI model doesn't really get the picture. So with this in mind, what happens now is that I established a regulatory position. I'm not going to talk very much about this. I think you get the idea. And now the studio agent has the vital information regarding this product in order to figure out what regulations should be applicable.

For this product on the European market, and will present it to me as what ones are applicable, what ones are non-applicable for defensive reasons, and also. Which one that requires further review? Because the set of metadata that we just provided for the AI will never ever really be enough. You know that there are so many nuances, in things that having this as a self-playing piano at once, that would be not trustworthy, really. So that's what happens here. And it's reading through actually all the potentially applicable regulations on the European markets for a life science product. I think that's near 150 regulations identified, so we have all those already in our backbone, and that's also important that you have.

I shouldn't use the word curated, but probably stack of information that is designed to make use of for the AI. So we, you know, we tell them not to go out and search on the internet generally, but look in the set of data that is assigned to their knowledge base. All right, so for some reason, a little bit slow here now, so I'm going to skip this. We can go back and look here later on if we want to. But this was just, my little example, just to state what is what. It takes approximately 2 to 3 minutes, sometimes a little bit longer, to get that first regulatory position. Now, what happens later on, is, of course, again, context. So, maybe I get, GDPR as an applicable regulation.

And then I break that down to see what clauses within GDPR is applicable on my product, given all the information the AI knows about it. Then I can have a dialogue about that clause, and that rational. And when is there an update? And now comes the cool thing. When there comes an update, let's say it's an amendment, or it could be… it could be other kind of signals, a case law, regarding that, that has impact on the clause, that it's applicable to my product. That's when the real value starts to happen. We're digging down deep in the relations between regulations and my product and how the AI can help.

me to understand what's going on, what potential implications should I consider, and so on. So that's vertical AI. And what we usually think about would be general AI. I will get back to this a little bit later on. Yeah, that was… a short introduction, just to settle the scene. Now, let's get back to the real story here. where I will share with you what I learned, during the last 5 years. And, Sean, please, go ahead and interrupt or, ask questions. Anyone is, of course, allowed to, to, raise their hand and, and interrupt me for. questions or if you would like to, argue something or wish me to slow down and deep dive a little bit upon something, feel free.

I'm going to go ahead and share my screen.

Hey, Marcus, before you go ahead.

My sister.

Bob, you got a question? I'm finding it hard to… To ground what you're saying. So I understand what you mean by vertical AI. I put some questions over there. Could you just help us understand, is this a tool you're offering? Is this something you're working with? How do we use it in the US? Where's the data coming from? There's a lot of AI. Vertical AI is just — you could call it a RAG. You could call it an SLM. It's focused. But the question is — Where's the data? AI is spelled with one word, data. Without data, you don't have AI. So I'm just having a hard time grounding.

Okay.

say, practically.

quickly, because we could spend a lot of time, walking through that, because there's a lot of different kind of data. First of all, it's… the important data would be the regulations per se, right? The actual documents that are, verified and gathered manually. kept as a database for this. So that's the true food, of course. And then, on top of that, you have the train part of the AI model, the reasoning model, which, of course, is not ours. That's different kind of AIs that we use for different purposes. So we're not an AI company in that context.

Okay, so if you take that example, because we're all over here, we're all, you know, consultants in the space, right? So you're saying if one of my clients could use your… is this a platform you're offering? Where if I was a diabetes product, like you showed the example. Would I upload all my content and then do reference only related to my specific space?

Yes, s.

Or is this a general service I'm purchasing? And then I have to click, I only want the diabetes-related regulation. That's where I'm lost. I don't know what you're talking about.

Yeah, so let's… if we take this example that I just did, it's… every product and every company is kind of unique, right? It has some… if we boil down the… The technical file. to the true essence of it. We get some answers from the uniqueness of the product, right? What kind of components it contains that dictates whether it should have the reach or roach. Or maybe it contains some radioactive parts, and it's being shipped with radioactive parts, and so on. And that gives us the clues and the relations to what regulations that should be applicable for that specific product in that market. All right. So we have the data set on those regulations. The user provides the, the, the,

The metadata, the important metadata related to the product, and using that. as a combination, the AI has what it needs in order to provide a tailored, regulations register for the, for the user on a specific product, on a specific market, or a company on a specific market.

So when you say we, you're representing this company and it's a service you're selling?

Yeah, it is. It's a.

Exceptions.

So, yeah. So that's… yeah, sorry about that. So I just used our own platform as a reference for… for, to explain what vertical AI is. And as I said, you, Paul, you seem to be very familiar with the terminology, and we're 150 people here today, so I just wanted to make sure that all of us are on the same page when it comes to the Perception of what is what. Yeah, cool, very good question, thank you. Did we have any more questions on this?

It was a similar question that we had from Karen. Who checks the knowledge base is accurate? You said it's done manually. So I assume that's the answer. Then, so is it fair to say you're offering observability around regulations related to a product?

Observe regulation. Oh.

The…

Did you say observe or obscure?

Observability, meaning vigilance, meaning.

Yeah, you know, here's the thing. When you have that panned out, your applicable list of regulations or regulations register, as we call it, that cuts away the noise, right? So the AI, in this case, we use a combination of OpenAI and Perplexity and our own monitoring system as well. Just to make sure that whenever there's something related to that specific regulation, in the context of how it is applicable for you. Which means that the AI can actually tell you, yeah, this is important, because I read it, and I also read the rationale on why you have this applicable regulation in your, this regulation that's applicable in your register.

Right? So that's… that's the true meaning of vertical AI. It's… the AI is trained to do something very, very specific. Right, so…

So if I can summarize for people, I think the thing that you need to understand is that you put your product information in here, you check a couple boxes and it's going to tell you which regulations are applicable to your product and what you need to know and what you need to monitor. Is that fair?

Yeah, and actually you don't need to do anything because the AI does it for you or the platform. What the user needs to do, and this is very, very important, is to verify.

Right.

Making sure that you perceive the information and that you agree. If you don't, then you rely upon AI as a decision maker. And I would strongly advise everyone to never, ever do that.

Right. So we have a question here. What type of ML is being used for compliance studio, i.e. unsupervised learning or reinforcement learning? Well, all of these tools are here to help you. like with general AI itself, if you take what it says at face value, you might get it wrong. You know, we've seen cases where, you know, lawyers wrote briefs entirely with AI, and it just manufactured things that can happen. That can happen even with a structured machine learning system. It's not entirely out of the question. These things drift. I don't really know how they drift, but they drift.

No.

They, they, they, they, so, you know, this is not. replacing human beings. Your job as regulatory affairs professionals is to use the tool, but to use tools wisely to help you make decisions Faster. Easier. You know, with not as much manual lookup, cutting and pasting, blah blah blah, you know, the things that we…

Nope.

That's one of yours doing.

Sean, to your to your point, if I can jump in, right, If you take our space alone, regulatory, right? I mean, you take somebody like Carrie, who can go through the entire value chain of every stage, every inflection point in the product development life cycle, there are different kinds of regulatory guidelines you need. And if you're saying, the statement you just made, we have to be careful to validate. If we're going to validate now, say, at the 510K level versus the safety level where Leo would come in, right? Is your system, I'm just asking because this is a very important Approached. process to take care of. Is your system going to say, when doing a 510K, all this research, please validate, ta-ta-ta-ta-ta? And would you be offering, at every inflection point of product development, a recommendation on what to check? Because if that doesn't happen, we might as well do the research ourselves.

I'm just curious about that because that's a very powerful, it's an important statement. The human has to validate, but validate what? What can we be a thousand percent sure that the AI is done correctly? And what is the AI asking us? Is there a threshold, a gateway for risk management?

Yeah, that's a really good question. But also, if you think about it, what you're saying is that when do we know what to validate? How can we? How can we ever know that, really, honestly?

Sure, but then we have to revert back to what we know for sure, right? We know that looking through whatever our current systems are working, if I'm going to consider a different tool, I need to be able to trust it. What can I trust and what will the AI tell me? Demonstratively prove, trust me here, but don't trust me there.

Yeah, there's a, there's one, there's one good thing here where you can program the AI to have to reason with itself regarding the, the percentage of How sure it is on the statement it makes. So, that's… that's a very good technology to use, which we do use often. So, when you get… as I said, sometimes the AI will present an applicable regulation for further review. And that's when it's less than 90% sure whether the. Regulation should be placed or proposed as applicable or non-applicable. So that's one of the examples. It's a really, really big question. I think that question, per se, is… is a Friday call focus, on its own. It's… I really love the question, and it's huge, and I don't think we have time to… to address it.

Yeah, 90% is not acceptable, right? I mean, that sucks. Lives matter here. People die. 90% accuracy sucks in our space.

Okay. Well, you know, I think that we can get. sidetracked on this discussion for quite a while, and I want to keep things moving forward, because I think that's important. You know, to acknowledge your point, Paul, lives matter. Yes, of course they do. when… you know. We're trying to validate whether or not something is written in a regulation or not. I think that we can all agree that AI can do that better and faster than human beings can. And I think that's what we're talking about when we're talking about doing that at scale. I think the question about applicability.

of those regulations to a specific use case are where human judgment come into play. So it's not the… AI is 90% sure that this application that that. that this… I don't think human lives are at risk is what I'm trying to say. I think that what is at risk is probably more your professional reputation. And so what you want to do is like with everything that we've done for our entire careers. Is use our best judgment and use our experience and use our knowledge. to, validate things, you know, it's just like reviewing the work of a junior employee.

That… That that's how I.

And read very fast. Yeah, something like that. But let's save that for now, because I really want to, make sure we have time enough for me to, to share with you all my findings, because I think that in itself will, will, answer some of the questions that arised here, and thoughts you guys have. I… I, I hope so, at least. So, when AI replaces regulatory professionals, what happens next? What I learned from working with AI and interviewing PLOS 200. Over the last 5 years. So… Will AI replace regulatory professionals? Bold statement. And I say, yeah, but of course, not the way most people think.

So I know this is a little bit corny, but it's also important, because AI will replace regulatory professionals as we know the profession today. So I think we can bypass the entire question, whether it will, whether it's not, what parts and so on. We need to. we need to think about a little bit differently. It will replace regulatory professionals, as we know the professions today, but it's not unique. AI will reshape almost every Knowledge profession over the coming decade, you know, think about farmers. 20 years ago, a significant part of their work was maintaining machinery. But today, many spend their days analyzing data, configuring software, interpreting sensor information, and optimizing automated equipment. They're still farmers.

They're doing the same job. And the exact same thing will happen in regulatory affairs. And the question is whether our profession disappears. The question is how it evolves. And how we can prepare ourselves for that, for the better. So… With that said, what is regulatory affairs, really? And this, this is, maybe, a little bit boring, because you know this from before, but we have to agree. So, I'd say that regulatory affairs is about knowledge, research, then monitoring, interpretation, decision-making, communication, and making sure that we have accountability, where Paul's question comes in, because that's the top of the ice… the top of the mountain, more or less. It's super important.

And how do we make sure that happen? If AI is such a great tool, and we can't reach accountability, and we think that the process to get there is even more, cumbersome. then it's no use, but there is. So… Many people might think about regulatory affairs as, you know, documents, technical files, submissions, SIRS, SOAPs, but the documents, they are only output. So what I'm stating with this picture or slide is that regulatory affairs is About acquiring, knowledge and monitoring and exchanges for that, understanding what implies interpreting the requirements, and assessing the impact, and making the decisions, and communicating that, those decisions across the, the organization, right?

So in my opinion, that would be the profession per se. And looking at it in that way, leads us into. How to think about this just as the farmer evolved around from moving from maintenance of machinery and and you know knowing everything about soils and crops. Yeah. So again, general AI, vertical AI. I didn't talk so much about general versus vertical when I showcased Compliance Studio quickly and briefly in the beginning. But again, this is really important. They're not… it's not so that there's a beef going on between general AI and vertical AI, or that they are great for everything. Not at all. There are pros and cons, there are risks, and there are rewards.

And just a quick idea or or you know, showing this, how one can think about it, that general AI, such as ChatGPT or Claude and so on, their strength is creativity, drafting, brainstorming, translation, and so on. Limitations would be, obviously, that it doesn't have very much of an organizational memory. No regulatory context. It can provide hallucinations and you're prompt dependent. So, whatever you put in. dictate what's come out. And if you're not extremely disciplined in your prompting, having the same style of prompting, using the same phrasing, words, definitions, and so on, you will drift.

So it's great for individual productivity versus vertical AI. Can support an organization's regulatory decision support. And. So again, just a quick reminder of what is what. And also, I think this is also important to touch upon, what AI is already better at. I will soon get to my findings. I just wanted to settle the scene a little bit. What is it better at? And when I say better at, it's better than you and me and any one of us in this room. It is better at finding information. It searches thousands of pages in a second and reads it for us. It is way better at comparing requirements and finding differences across regulations and document versions.

If we tame it and use it correctly. And it is better at us in monitoring changes, continuously monitoring thousands of regulatory sources. We can't do it in that speed. And it does… have pattern recognition, so identifying relationships humans are likely to miss. Because our brains cannot process that huge amount of data. So this is what it's better at, at the moment, and also important to remember. But these superpowers comes with a risk, right? It's easy to see if you think about it. AI is already transforming how regulatory information is processed. It happens already now.

So the question we have to address now, I think. And that this taps in a little bit to what Paul said, and in his questions, and I think that. This is a busy slide, but I'm just going to take a few of these questions. I hope we all share that. These are very important to address. What work should AI perform, and what should always remain human? How do we validate AI-supported regulatory decisions? What level of transparency should AI provide? How do we maintain trust in AI-assisted regulatory work, and what competences will future regulatory professionals need? I mean, if we are to use this very, very powerful tool.

We obviously need new competences. How should organizations govern the use of AI? How will regulators and auditors expect it in coming years? How do we ensure consistency and responsibilities adoption across the profession? This, and probably 100 more questions, needs to be addressed now. Not now, as in this webinar, we don't have time for that. I can speak fast, but not that fast. And it will bore you. So… Those are the questions. I just want us to have them in the back of our heads when we jump into this. So my biggest observations after 200 conversations, probably almost 300 now. Of course, many of those happened five years ago, and they are more or less obsolete because it was in the beginning of the beginning. But.

The path! the… The way I had the opportunity to follow people's reactions, behaviors, perceptions of AI and regulatory affairs along the way has been really valuable. It changes. Of course, a dialogue with someone today is completely different to what it was five years ago. But nevertheless, this is what I heard. So… so the… the biggest… the biggest finding, really, it's… it's still… it was, and it still is, inconsistency. So, some regulatory professionals, they're generally afraid of AI. Will it take my job away?

Will it change my job description and my profession into something, you know, less valuable? Will it make me stressed? Many, many various things that creates fear. And simply a statement, we should not use it. Then. Some… Dismiss it as unreliable and overhyped. And I cope with that. That's, you know, we're different humans. Some of us are early adopters, some are late adopters, and some are in the middle. That's the way it goes with changes. It's nothing new. And this is our… This is our built in system on how to react to news. And we're different there.

And then we have… the experimentation. We're testing a few use cases, and they just jump in and test various things, and have iterations and all of a sudden dies out and then comes back. That's a big chunk as well. And then we have the ones that goes into dependency. I use it for almost everything. Sean. And And then we have the final ones. I grouped them, there are probably more groups than this, but then we have the ones that performs misuse. They use it without a sufficient context, a review, or control. They… As I say in the bottom of this slide, very few organizations… well, I shouldn't say very few, but a lot of organizations appear to use AI in a deliberate, efficient, and mature way. And most believe they do. They think they do.

Maybe that's my perception because. I have more experience and I know what is truly efficient and what is not. And this is a learning curve, of course. Okay? But what surprised me most was how few organizations seem to use AI in a deliberate, efficient, and immature way, which shouldn't be a surprise. Maybe I had high expectations. And, and, But it kept coming back over these 5 years. That's the one thing that has been constant over 5 years, really. Okay. It's just…

Interesting that you bring that up, you know, the 5-year trajectory. You know, I started using it about 3 years ago. I'd like to get a sense from the people on the call, you know, if they're using AI, if, you know, they feel that they're… that they're… that they're using it, what is deliberate, effective, and mature way. Or… because, you know, I think there are a number of people on the call who probably just started using it in the last 6 to 12 months, you know, and… I think there needs to be an experiment, experimentation stage, and, you know.

Good evening.

to become convinced that it can help and understand how it can help. Yep. But where are the people on the call? We have 200 people on the call. You know, please, you know, put in the chat where you are. You know, you can write the word dubious. You can write.

That's a cool.

You can write, you can, you can write the word dependence, experimentation, dismissal, fear. You know, I don't care. You know, we're not here.

Or misuse. That would be really mature to state that. Yeah, yeah. You know, I'm sorry, I think part of the reason, just as you said, it's a learning curve and the future is a little bit uneven distributed, but a lot of people may not fully understand what AI is, what it's good at, and where its limitations begins. I think there is an other reason, and this is really important. Most of us started to use AI through a chatbot. And what happens is that a chatbot provides instant gratification. Especially now, it didn't do so in the beginning, when we tried out ChatGPT 4.1 and so on, it was just…

crap, really, but we were surprised, like, how on earth can it reply so well in no time? This is magic. That instant gratification, you ask a question, you receive an answer immediately, and that makes it very easy to focus on the next prompt, the next document, or the next shortcut. And very easy to lose sight of the bigger question. What should AI change? What should it not change? How should regulatory affairs evolve around it? Because you got dragged in. due to this instant gratification. And the dopamine that we get in our brains from using AI is probably equal, if not worse, compared to social media.

So that's the danger of using chatbots and AI, really. I think that is a bigger thing than we may be thinking about. And it's a part of my experience as well, what I see happen with people, especially the ones that goes into experimentation and dependency. Yep. Sorry, you're m.

Maria E. Donawa, MD: John, you're muted.

Yeah, we have a question from Maria.

Maria E. Donawa, MD: Okay, it's a quick comment. You mentioned, Marcus, that there was no organizational memory. Are you Maria E. Donawa, MD: But there is a way to personalize AI, and I mean, I tend to work with ChatGPT, but you can certainly personalize it, and if you use it over time, if you're using the business model, it certainly remembers the context of what you're talking about. Maria E. Donawa, MD: So that's there. So I wasn't… I didn't understand exactly what you said when you said there was no organizational memory.

Yeah, you're completely correct on that. You can create agents, and you can create projects, and you can create a lot.

Maria E. Donawa, MD: I don't mean agents and projects. Sorry, I gotta stop you there.

Okay.

Maria E. Donawa, MD: I'm just talking about the basic use of ChatGPT.

Yep.

Maria E. Donawa, MD: where you personalize it so it knows the user. That's one thing, but the other thing, I'll just make a quick comment. I think those of us who use ChatGPT to check things in our work, if you use it enough, you learn pretty quickly Maria E. Donawa, MD: to check certain things. And if you use it to any extent, and of course there are times when you don't, there are times when you do, and in between, and knowing when to check it is really, really critical. So it doesn't take… Maria E. Donawa, MD: Too long before you, if it refers you to a reference, first of all, it's a good idea to make sure if it's a regulatory issue that it is providing references.

Maria E. Donawa, MD: And the trick is, of course, is to check them, because every now and then it'll give you a reference, and it's just not… either it's not correct… I'd have to say it's generally correct, but there are things that you have to check depending upon the complexity of the task. Maria E. Donawa, MD: So, as I think, Sean, anyway, I think the statement was made, it can do things and go through tremendous volumes of work. Maria E. Donawa, MD: I'm sorry, of data and papers and that sort of thing. But then when you get down to a regulatory issue, I think, again, those of us who work with it know that you can't just send it on its way and you're done, that you're going to have to end up checking.

Maria E. Donawa, MD: what it's saying. And what I like about it is the back and forth. And some of us may have gone into this, it's like, why are you telling me this? This is not correct. Or, this is not precise. And of course, it's frustrating, because it'll say, oh, you're right, because it's trying to please you. So it… Maria E. Donawa, MD: you cannot use… you cannot learn how to use this well without using it, and being careful. I think that's what it comes down to.

Thank you.

That's a very wise thing to say. Let me just reply on the difference between organizational memory and personal settings. So with personal settings and ChatGPT and well, pretty much all of them now is amazing. Remembering from your previous chats, remembering your personal preferences on I would like to have this in British English and I would like to have this in this and that format. So yeah, you can do a lot of settings. Now… The tool per se is just a Unstructured, huge folder of memory. And that's the… difference between an organizational memory. So you organize your memory saying that this is product X. When I'm here in this view, we will only talk about this product on this market for this purpose. So you get a more stringent workflow

Set up with that. then you can't… well, you could probably create that one way or another with ChatGPT, but it comes with a high risk. And I wouldn't really recommend it. But yeah, thanks for sharing that. I think that what you just said, that you need to tame the beast and understand how it works, no matter if it's vertical AI, because the reasoning models, they are remaining the same. No matter if it's an AI model that's within a vertical AI tool or if it's a general AI. Sorry. Yeah, thank you. Sorry, you said something.

Maria E. Donawa, MD: Just a very quick question. When you talk about what you presented earlier, are you talking about, are you using AI in an agentic manner? Maria E. Donawa, MD: Is that how you're using it.

Yes.

Maria E. Donawa, MD: Agent?

It's a combination. We use various. Depending on the… it's task dependent, really. Okay.

Maria E. Donawa, MD: I don'.

You know.

Maria E. Donawa, MD: I'll need to… I'm sorry, go ahead.

Yeah, but this is important. So in the vertical AI, so whenever there is, so it's actually presented in our case and then also in EQMSs that I've seen where they have started to build it as vertical AI, they have various models or not models, but various instructions for the AI, depending on the task, right? So… if it's a certain kind of document, it should know exactly how that document works, but that AI version of that AI agent will not be able to mix that up with another kind of document. So, it's like… Hundreds and hundreds of versions of the same AI with different instruction and different knowledge sets.

Okay, I'm.

Bing.

Yeah, I'm gonna say that we got 5 min left, and you have an important announcement to make, and I don't think that you got. Through your slides.

Where do we go from now? So this is important. How do we separate hype from real value? Where should AI support decision and where shouldn't it? What competences will future regulatory professionals and so on? this… the most important thing to say here now is that no single person, company, or organization has all the answers. But as we all recognize here. With all the fantastic, Questions and the shared stuff that I've seen in the chat. We're sitting on pieces of information and we're sitting on pieces of experiences, all of us together. And that combination of that can't provide answers. So, the big announcement that I will come to is something that me, Sean, and a good fellow of us, Leon, I think he's here somewhere?

We've been thinking about this for a long time, and we felt that today would be a great opportunity to announce the birth of our dear Sarah. So Society for AI and Regulatory Affairs. We think it has been long needed and it's, it's finally alive. Now bear with us. We're just getting started. So what it is at the moment, it's not wraps. Not at all. It's not 100% governed. There are flaws. It's, it's a beta version, but the intention is true and honest. It has Whatsoever. No commercial parts in it where where we would like to promote things or say that something is better than something else.

We just want to share. So… we kindly ask you to join this growing international community, learn from experts, share your own experiences and perspectives, and help share the future direction of SARA itself, because all of us together should be the one that define what SARA is needed for. We have a fairly good idea. There is a codex that you can read on the SARA website. It's unfortunately hosted under a corporate domain at the moment. We will move that, but don't be scared by that. Maybe I shouldn't even mention it, because now you will read in the URL. But it will be moved.

So, it's not the finished version.

Yeah, very much.

You know.

And we fixed the typo on this slide, but then we… Oh, yeah.

Then we forgot to change the slide deck.

So, yes, thank you for pointing that out.

Hide it somehow.

And that's okay. But we were aware of it.

Yeah, we ran out of time, just as we're about to do now.

So there was one question, someone, who was it? Hmm. No.

I've got a question, comment, if I could. Yeah, I would just say, you know, great presentation, Marcus. I think a lot of what you're saying rings true. I think my observation on this has been that, or I'll give you my position, I think AI is at the point where it can be like extremely, extremely, extremely Powerful, and can do almost everything you need it to do. And… to the extent that a person doubts that it's largely tends to be a skill issue. In other words, like the power that it's got right now is related to the skills that you have on how to properly use it. Like, you know, the people who think, like, ChatGPT is, like, the… like, a primary means of doing things is, like, okay, you're probably, like, at a level 3 out of 10.

And so I think I applaud this effort that you're doing with Sarah because, I mean, I think that's the challenge for all of us in the industry is like, how do we continue to climb that learning curve? Andrej Karpathy is like a major influencer in the AI space. he's… I think he's now at, Anthropic, but he was essentially a co-founder of OpenAI. He has a line that I like, which is, like, he's come to that if there's something that you can't do properly, it's probably a skill issue. Like, it's not that the model can't do it, it's probably just that you haven't, like, built the skills up to do the thing that you're trying to do. You know, being in the software space, I mean, AI has, like, completely revolutionized the development of software.

You know, being in cybersecurity, like AI is completely revolutionizing cybersecurity. I mean, I think we all got a taste of that with what happened with Fable and GPT 5.6, where the models are so powerful that literally the government had to step in. And so I think, again, I just applaud this effort because to the extent that as a group, we can continue to climb that learning curve, it'll just allow us to do better work. And I don't think it's going to replace jobs as much. I think, you know, in the software industry, where it's been so powerful, there's actually more job openings now for software developers than before, even though a single software developer can do 10 times as much. I think similar things are going to happen in medical devices.

I think these tools are making the cost of developing new medical devices drop, which is going to make it possible for more and more devices and more products to be launched, which means that more people with this expertise are going to be needed.

oh, you just stole my ending phrase about the volume of devices, but yeah, and I was about to miss that one, and that's a really… interesting conclusion, because we see, as you say, it's… this has happened in software, it has happened in… even in software testing, where everyone thought that, yeah, that… that job is gonna go away. But. The, the, the total size of the industry will grow.

And we have done here.

Yeah, sorry.

I can't believe it.

Way longer. We will come back and let you know more about Sarah. There will be webinars, there will be roundtable discussions, there will be… Discussions and everything. We have a lot to do and a lot of responsibility together. With all of you.

Is there a link or something?

There will be in the email that you will send.

OK, so the one that I'm sending tomorrow is going to have a link in it, is that correct.

Yes, yes.

Okay. Well, thank you everyone for your time as always. Special thanks to everyone who contributed live. And I would But to remind you that. Next week, we have, a number of speakers joining us from The org.org, phmti.org, so take a peek at pm.org. Phmti.org and their mission. It is a nonprofit. We like nonprofits and. Then the following week, Andre is going to be speaking. Right, Andre? Say wave hello.

contract manufacturing and engineering of electrical devices. So.

Cool.

Yeah. If any of you need a manufacturer, he's your man. So… Here we go. I know I'm forgetting something, but all right. Thank you, Marcus.

Thank you so much. I will now go out in the Swedish heat. of extremely hot 26 Celsius degrees, as you laughed about.

Huh. Yeah. Babies.

Yeah, I know. Vikings. We're Vikings.

Biking.

Maybe White Case. Maybe White Case.

Thank you so.

Thank you so much. Have a lovely weekend, all of you, and thank you, Sean, again for for arranging this and also for the great contribution to Create ZRA.

Thank you.

Thank you. Take care and say hello to the zucchini flowers.

Thank you, bye-bye.

Bye.