Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
Service Provider Executive Series #18: Rainer Schwarz on Automating Regulatory Review with AI
58 Plays1 month ago
In today’s episode, Steve Gens sits down with Rainer Schwarz, CEO of KlearDoc and Executive Advisor, to talk about two decades of regulatory software innovation and the rise of AI in the space. In this episode they explore early NLP‑driven text extraction, today’s LLM‑powered review automation, and Rainer’s staged vision for AI adoption across regulatory.
Listen now on your favorite podcast platform or on our website: https://gens-associates.com/gens-associates-podcast/
Want to hear more from Rainer? Connect with him on LinkedIn: https://www.linkedin.com/in/rainerschwarz010371/
Recommended
Transcript
Introduction to the Podcast Series
00:00:06
Speaker
Welcome to the Gens and Associates Regulatory Executive Podcast Series, where we explore innovation with key leaders in the regulatory space. This is Steve Gens, Managing Partner.
Introducing Rainer Swartz
00:00:17
Speaker
I'm very happy to have Rainer Swartz, who is a founder of several companies.
00:00:22
Speaker
Rainer, I think we first got to know you pretty well back in the Kunsoft days, where you started Kunsoft and recently cleared Doc, and you've always been on the cutting edge of innovation and i have a lot to say about AI too, which is the big topic of the day.
00:00:37
Speaker
So before we get started, for our listeners, why don't you introduce yourself and your company?
Rainer's Career Journey in Regulatory Software
00:00:43
Speaker
Yeah, thanks very much, Steve. And thanks for having me as part of your podcast series. I'm very excited about this.
00:00:50
Speaker
I have been working in the area of regulatory software development for more than 20 years now, and time is passing very fast, I have to admit. And I would describe myself as an entrepreneur who is very passionate about spotting technology trends quite early.
00:01:09
Speaker
And then turning them into real business value. That's what I think is a key strength and has been somehow the the theme of my career. So you mentioned innovation, and this is part of the innovation podcast series. And innovation is a very important aspect when you are an entrepreneur. And so how to achieve innovation in a very simple way. And I've i've looked at back at how I was doing that in the past, and I think The thing that I've learned is you start innovation with listening.
Innovation through Customer-Centric Approach
00:01:39
Speaker
So most of the time, these problems are clearly defined and you just have to listen and understand the market pain points. Customers will speak to you about their problems and you just need to listen.
00:01:52
Speaker
And then as a entrepreneur, as a software developer, as part of a software developing company, you just need to understand what those problems are. And very often, you know what new technology is available to to solve those problems.
00:02:08
Speaker
And basically innovation is really connecting those thoughts, right? So you're mapping the real world inefficiencies with the emerging technology, and then you deliver some meaningful impact. And that's what I've tried to do along my career. And now with AI, it's very exciting to have this kind of conversation more geared towards the topic of AI.
00:02:35
Speaker
Yeah, so thanks so much for that intro. And I know when we were discussing doing the podcast, you mentioned something to me that really stuck with me. we were saying, I've been doing AI since 2017 and all that, because a lot of times when we think of AI,
00:02:50
Speaker
It's kind of 2022, 23, where it emerged. And you know I'm a big supporter of the ah Gartner hype cycle, which is actually a Mars law. In the the short term, things are very overestimated. In the longer term, they're underestimated. But you know you started working on
AI and Cloud Technology in Regulatory Work
00:03:06
Speaker
this around 2017. And I do recall very clearly, Greg and i we were presenting to one of our clients. you were at Koonsoft. And um This is back with the IDMP and IDMP compliance for a large multinational. If you did it manually, it was just impossible, right?
00:03:23
Speaker
And that's, I think, some of the background with IDMP, some of the automation you brought to the forefront. So maybe for our listeners, before we really start jumping into the juicy topic of AI, a little bit more about the history, kind of as you started innovating as an entrepreneur in Koonsoft,
00:03:40
Speaker
and a little bit about ClearDocs today, give a little bit background, and then then we'll just jump into the AI conversation. Yeah, absolutely. And as you said, the over the past decade, I think there were two major career milestones that stand out and one is QnSoft and the other one is now ClearDog.
00:03:58
Speaker
And QnSoft, we founded QnSoft back in 2013. And by then the goal was to move regulatory software into the cloud. That was the trend by then and back in 2013.
00:04:10
Speaker
and And really automating certain areas. And the industry was still very manual and it was cloud averse. It was very interesting by then. We were a smaller organization and talking to large pharma companies.
00:04:24
Speaker
By then, they we we were told that regulatory processes will never be handled in the cloud. They are far too risky. And i even heard from one representative of a top 10 pharma.
00:04:37
Speaker
IT department said, well, as long as I work with that company, and we will never do anything in the cloud. So that was, if you imagine that today, right, just just about 12 years later, how much the industry has changed. And probably that's also the case with AI. is also
Pioneering Text Extraction Technology
00:04:54
Speaker
changing a lot. And there there are a lot of...
00:04:57
Speaker
pros and discussions pro AI, a lot of discussions con AI, a lot of people that have very high expectations, other people very low expectations. So it's all emerging. and We have to observe how it is going to emerge. Yeah. and And back in 2017, as you said, with Quesoft, we started working with AI. And again, that was a problem that was being brought to us by the industry was to how to prepare for IDMP.
00:05:25
Speaker
IDMP is the identification of medicinal product. That's a requirement for the European applications and say, IMA health authority requirements that requires pharmaceutical company to to submit a lot of data instead of just documents. And lot of that data was trapped in documents.
00:05:45
Speaker
So how to get from unstructured information into structured information. So it's how to extract data from documents. And by then the large companies that had a lot of burden with this IDMP preparations and to try extract data, they were like outsourcing that to services providers.
00:06:05
Speaker
And then our approach was to say, this must, it must be able to automate that. And so what we started to do is we developed a text extraction engine and we called it by then Q-Distiller because it distills data out of documents.
00:06:22
Speaker
And that wasn't, that was really a breakthrough solution by then because it wasn't possible to do that with technology so far. we had to use natural language processing and deep learning mechanisms. We had to find algorithms and train algorithms. it was very difficult process because text by then was not really an area where I was focusing on.
00:06:44
Speaker
So we're somehow pioneering that area. But ultimately, we were successful and many large pharma organizations started to adopt that technology. And it's still in use with some, i I'm hearing.
00:06:57
Speaker
And I think that says a lot. Yeah, so that was the AI start of the AI journey. And at some point, Kunesoft got acquired and and after that I took
Role of Large Language Models in AI
00:07:08
Speaker
a break. and And then once the large language models came out, I got again very intrigued to continue that AI path. And only recently, about two years ago, founded ClearDoc.
00:07:22
Speaker
And again, the idea is to take that original AI vision further. And the large language models that were emerging or have been emerging during the last three to four years now, they're really a game changer.
00:07:40
Speaker
And I can say that because we have been developing 2017. And we by then or do we did by then was to take an algorithm and to identify training data, to create a training model, run a lot of test data through that training model, and then train the AI. And that took very long for one simple task.
00:08:02
Speaker
And the game changer now is with Gen AI and with large language model, You can say that somebody else trained the algorithms already, and you can directly start with the automation process without the need to, of this complex training cycles. And I think, I do think that this opens the door to real scalable intelligence and what we see right now, it's much faster. It's, it's adaptable and it's very easy to access it.
00:08:34
Speaker
Yeah. And what that therefore, so what ClearDoc does and what we do is we do provide a flexible AI engine, call it the AI core engine, and that is supposed to be used within or embedded in another software.
00:08:50
Speaker
And the first partner is Lorenz, our first development partner. And what we're doing is we're co-developing a very innovative solution that is brought to market under the brand name Verify, Lorenz Verify.
Collaboration on Regulatory Document Automation
00:09:05
Speaker
So what does Verify do? It basically automates the review of regulatory submission documents. And when you think about submission documents and how these are being reviewed at the moment, has a lot of manual process, especially when you look at it from a regulatory content compliance perspective.
00:09:25
Speaker
It's all done manually right now. And what we do right now, we are piloting Verify with some health authorities, but also with the industry. And with very remarkable early results. And it's it's very exciting to see what AI can do today.
00:09:43
Speaker
Yeah, so that's um that's actually a perfect segue, Rainer, into kind of AI for today.
AI's Supportive Role in Regulatory Processes
00:09:49
Speaker
And again, on a recent conversation, as we are really discussing AI in different use cases, you brought up the thing about AI in phases, which actually made a lot of sense because we're in the early days when people are talking about eugenic ai And then kind of like the old school, as you had mentioned a decade ago, so much data was trapped. I really like that, trapped in documents. And, you know, could we even imagine 10 years ago that we could have tech that could read all the dossiers, all the approval letters, you know, all the correspondence and starting making meaning of it Because that's why we always say here at Gens and Associates, AI is just different expressions of ah an assistant, right? writing assistant, an authoring assistant, a research assistant. The one thing that did strike since you brought up Lorenz when we were you know both in Denver and had that sidebar conversation, what really impressed me is there's so much focus
00:10:44
Speaker
and rightfully so on content generation, right? Right. So if you think about the generation, um but more time is spent, if you take it from a process standpoint on the review process, you have technical review, you have content review, up and down the hierarchy, but a lot of times to reduce say last patient visit to database lock to the you know first filing, a lot of has to do with you not so much the generation, but the review process.
00:11:12
Speaker
That's where you go out. So I was really impressed with the verify and envision in the future. Maybe it's word plugins or whatnot. How much more could that be done with AI at the source with the author on the generation, which would actually probably eliminate a lot of downstream tasks? Because as you know very well, one of the big KPI metrics from like a reg ops standpoint, because they're the receivers of all these hopefully submission ready documents.
00:11:38
Speaker
is the percentage of documents that are truly submission ready, you know, and kind of past that where most people don't realize, as you pointed out, it's a very manual, very burdensome process. So that could be reduced and it's not fun. People don't go to work. Let me review, do technical review on documents.
00:11:57
Speaker
So kind of the automation of that. So that's a really good first use case that you expressed about what's going on today and maybe in the longer term. So what are other areas of focus that you have or observations you have in regulatory or maybe the broader life sciences about the application of AI and genic AI?
Potential and Staged Adoption of AI
00:12:16
Speaker
Yeah, so thinking about stages of AI, I think that's a very good approach. And the stages, the the potential obviously is enormous when you when you think about AI, but there's also a lot of hype out there.
00:12:32
Speaker
And I do think that AI adoption has to be seen with taking stages in mind. And that's what we're doing with Verify with pilot projects, with customer projects, with looking at it in stages. But I also think that overall the AI emergence has to be looked at from a stages perspective. Yeah.
00:12:55
Speaker
And probably each stage has different timing. So you can look at it from a timing perspective. What is stage right now and what's happening in the next one to two years. and know two to five years and after that.
00:13:08
Speaker
But also what are the values for each stage? I think that's also very important to look at. and I do think when we consider the current stage, it's for moment, say that stage one and this stage takes for maybe one to two years, maybe this year, next year and a bit of the year after.
00:13:28
Speaker
Then i do think where where we are is that in this phase, we look at areas where plenty of workflows exist.
00:13:39
Speaker
And the risks have to be manageable and the value then is clear. And what does that mean in regulatory? So that means looking at administrative tasks that you just mentioned, Steve, the reviews that are completely done manual. And there are various areas of of those review steps. and The scientific review is a very complex one, but the one that's happening before, which is the the regulatory review of whether the content the documents have been provided to the health authority as per the guidelines, that's a more administrative tasks in some aspects. Some are more complex of those guidelines, some are more simpler.
00:14:23
Speaker
And the simpler ones, and you would imagine the simpler ones are to check whether the correct boxes have been ticked in a form, in the, say an application form or a form word that you add to a submission. and And what we heard and in the pilot is that the review starts with checking those administrative reviews.
00:14:44
Speaker
So, for example, the reviewer who is supposed to review whether the drug is really safe starts with reviewing, is that really a variation or a submission is is the submission type that has been ticked in the box really the one that that has been submitted?
00:15:00
Speaker
Have the correct dates been put into the document? And there are some logical checks that take... quite significant time. And looking at those, I do think in this phase where we are right now, you can start automating those. So um i I definitely think that's possible with the current stage of where the AI is today.
00:15:21
Speaker
And you can gain efficiencies there immediately. And when we talk about efficiencies, I do think that's a reduction of the regulatory review time of those simple administrative tasks. And they are huge and they're lot. 50% is absolutely achievable from a conservative perspective. That's what we think. And yeah, so it's ideal for automation in this phase and to start implementing it. And would you agree with that? Or to what extent do you agree with that, Steve?
00:15:50
Speaker
Yeah, I think as I spoke in Denver, and I was very careful with the other words, and because I did a lot of thinking and talking with different colleagues over the summer that we're really starting a new era, right?
00:16:01
Speaker
So like that 50% gain, and I just did a little math based on what you said here. so I'm going to go to my cheat sheet that, you know, we had something 10 years ago about the time it takes regulatory colleagues to verify information, you know? Yeah.
00:16:16
Speaker
So we had a sampling and it turned out to be like 6% to 9% on a given week. And if you think about 6% to 9%, it doesn't seem like it's that much. ah Certainly like from a cycle time, reducing by 50% of what you just mentioned, that that's very, very significant. But even from a productivity, that 6% to 9%, if it's um like a 50-person regulatory shop, so that's on like the small mid-tier,
00:16:41
Speaker
you know You're starting to save 6,000 to 9,000 hours a year, so it's very substantial. But I think kind of what you're saying is, okay, from a process gain, maybe we can reduce that process by 50%, and hey, and we're saving you know thousands of hours with that that are reallocated.
00:16:59
Speaker
to more interesting tasks, working on product strategy, health authority interactions, et cetera, et cetera, maybe fine tuning the label. But the thing we're really interested about the research that's going to come out in January, one of the questions was, what is the hypothesis? We had like 10 or 15 use cases, half of them in the authoring space, you know with some CMC docs, a label docs, clinical you know safety. Those are the four areas we picked on.
00:17:26
Speaker
about what's the hypothesis for efficiency. And for those in pilot or in production, what are you getting? So I think, and i I completely agree with you, and our data suggests too, that we're going to see kind of the bulk of maybe the first generation, or I i like to term stages you're using.
00:17:43
Speaker
The first stages of AI really is 26 2027. twenty six and two thousand and twenty seven Because people think about, oh, that's the implementation tipping point that we're looking at, but maybe we have to think about it as the implementation tipping point for stage one.
00:17:58
Speaker
Let's get through some of the basics, right? Now, when I know we started talking about doing this podcast together, you said, well, I can really start looking further out in the crystal ball about maybe what could be in seven to 10 years.
Vision of AI in Personalized Drug Production
00:18:11
Speaker
So maybe I'll put you on the spot there, Rainer, and like,
00:18:15
Speaker
um I mean, some of these very near-term things everybody's working on, but when you look at your crystal ball, you're an entrepreneur, an innovator, you tend to see things before others. So I'm really setting the stage here for you, Raider. So what's what's in your crystal ball? and What could be a reality of regulatory or you know, when we think, I think it's more about R&D, the clinical safety, regulatory R&D quality, that community from an AI standpoint. I mean, anything...
00:18:42
Speaker
that you think could be a reality in seven, 10 years with different stages of AI. Yeah, that's a very good question. And it's something where you can be very visionary, and but you still need to be realistic. And what you always should keep in mind as an entrepreneur is to how can I create a moonshot? I learned that from, who was like i think Google was that who invented that name, the moonshot. The moonshot means minimum a minimum of results.
00:19:13
Speaker
increased output as the current solution. So 10 times better. That's how the moonshot was defined. And I really like to use that term because I think it helps you to to really quantify is is the vision, is that vision good enough? And and some many times entrepreneurs are being smiled at when they come out with moonshot.
00:19:34
Speaker
What is this guy talking about? And still, I think these moonshots need to be there. And So one of the moonshot, and it's probably even more than the moonshot, is is my vision of how drugs are being produced.
00:19:46
Speaker
And I do think that drugs currently are being produced for as if every human would be the same, which is not the case. So personalized drug development is, I think, where we should be aiming at. And I do know that there are organizations out there that are aiming towards that.
00:20:03
Speaker
And if you look at the capabilities that we have right now, and I don't i don't think we are there yet, we we need to go through those stages of Well, stage two, maybe be a bit of augmentation and starting to really implement a genetic AI and then really getting to autonomous solutions. That's what I think stage three could be. And if you, if I think at it from a patient perspective.
00:20:27
Speaker
My ideal scenario would be I go to the doctor, diagnoses me, prescribes me with a medicine, and that medicine is being produced, very personalized to my DNA.
00:20:40
Speaker
And while I'm walking home, this is being produced and the drone ep is delivering that personalized prescribed medicine to my door. And when I arrive at home, the drug is also there and I can start taking that treatment.
00:20:55
Speaker
And that's my moonshot that I have in mind. And I think with all the technology that we are developing right now, I do think that's possible. When you look at the improvements that we made during the last 10 years with AI, if you look at, if you ask any AI developer, would you have imagined that a large language model would be there and it it can resolve what it resolves right now. Would you have imagined that 10 years ago? every Everyone would have said, no way.
00:21:24
Speaker
No way is that possible. So that's the moonshot or that's the vision that I do think I have and try to work in small, tiny steps, support the path to there.
00:21:37
Speaker
Yeah, and that's certainly a fascinating moonshot. i use a different term. I call it Merlin's factor, you know, where you think about the end in mind when you start. And what occurs to me with your moonshot is, as you say, we're already making progress there. For me, how that's expressed.
00:21:54
Speaker
is with rare disease companies. One of them was my client, and I had to get educated. is like, Steve, you don't understand. we say rare disease, and they use the market of Japan.
00:22:05
Speaker
We have seven patients, seven patients in Japan. And instead of trying to get regulatory approval, they go with a compassionate use. But if you start looking at, even just with seven patients for this, you know obviously the DNA is different. Is there modification to, say, that base product?
00:22:24
Speaker
based on their DNA structure. Because back probably when we started, you know, I might have a decade on you, but just a lot of the drugs and compounds were mass, the pills, you know, the the capsules, the antibiotics, it's for broad application. And we've slowly as an industry, it's gotten a little more personalized and maybe in your moonshot, instead of incremental improvement in the next decade, it takes a major leap, which certainly would be very, very exciting.
Advice on AI Implementation Strategies
00:22:50
Speaker
So, um hey, the time goes real quick here, and I just will have one other question for our regulatory community here. Any other thing that you're seeing, you know, that you're either working on in the next or two, any other examples that should get the the typical regulatory professional excited if it's something you're working on or you're seeing that you want to share before we close it up?
00:23:14
Speaker
Yeah, I think my recommendation to everyone who starts looking into automating via AI is to to really set expectations properly and not over expect what AI can do.
00:23:31
Speaker
think that's very important. And a I think a professional AI provider will set these expectations realistically. And again, as we said before, at the moment, I do think we are in a phase where you should look at highly administrative tasks and try to automate those, have some more simpler things, start with the simple things and get familiar with the AI and then incrementally improve from there.
00:24:01
Speaker
i think that's very important. And that's really a paradigm shift from how projects or software has been developed in the past, because in the past you had a clearly defined spec and then you were developing according the specs, then you validated it, you implemented it, trained users and off you go.
00:24:19
Speaker
With ai it's very different because first of all, you don't know what is possible. So it's very hard to write the full specs. You maybe only want to write 5% of the specs, develop those and implement those and you get 5% savings immediately.
00:24:34
Speaker
And the other thing that's varies very specific in this industry right now, in the area of AI, is that it's so dynamic. yeah Almost every month you have new tools coming out and new large language models are emerging and so on. So you need to be very flexible and have an architecture that is adaptable to those changes.
00:24:58
Speaker
So be realistic, implementing phases and accept and expect that changes will come along this path of AI. I think that's what would be my recommendation from ah how to practically use AI right now.
00:25:16
Speaker
And I think just to add on to it, the I think the big takeaway for me, and we use different words, but I'm settling on kind of as you used the term stages, you know, because really now 26, 27 really is that entry stage with you. And it's a lot of learning, experimentation, getting some ah ROI, because everything's shifting. I mean, I've been covering...
00:25:37
Speaker
the ah The whole life sciences, part of life sciences when I was in Johnson & Johnson and then when I went in consulting for probably in the last 25 years covering this space and I've never seen anything like it. That's why it's like it's the start of a new era.
00:25:51
Speaker
So by thinking about it in stages and I mentioned this in one other podcast, and one of our clients last summer. We reviewed their digital roadmap, and it's the first time I saw it, and I tell any client I'm working with, because instead of, hey, we're going to turn this on, your typical five-year roadmap, right?
00:26:09
Speaker
But they talked about it and expressed it in accuracy. When we believe the accuracy is at this threshold, then we're going to bring it in. That's exactly what you're saying. You know but you have different stages and phases and unlike have to rethink about roadmaps because it's it's about accuracy and confidence in the accuracy. you know And it is going to keep on increasing probably exponentially.
00:26:32
Speaker
So I think that's a big takeaway for people that are listening and planning. Do it in stages. Communicate because that's when you you actually start getting the expectations align because instead of like when we're going to apply AI or not is when do we believe AI for whatever use case area will have the accuracy that we have confidence to put it in production. And I think you use just different words about you have to rethink about how you think about software and software implementation.
00:27:00
Speaker
So but it's exciting. Very, very exciting time. So thanks so much.
Contact Information and Further Discussions
00:27:05
Speaker
And if our listeners want to get a hold of you, Rainer, what's the best way? Yeah, getting hold of me is very, very easy. You can look up my profile on LinkedIn, or you can go on the clear.com website, clear.com and use the contact form there.
00:27:22
Speaker
And also i will be at the major conferences over the next years. So looking forward to see again, see you again, Steve, next year on the next conference and maybe you also some listeners that that are going to approach us.
00:27:35
Speaker
yeah Yeah. And I'm sure you're talking you know about the the big one here in the States, RSIDM. So we'll put a plug in for the DIA. That's always kind of an awesome show. So thanks again, Rainer. Very enjoyable discussion as always. I always love when we do meet at conferences, our sidebar conversations, exchange of ideas and thoughts.
00:27:56
Speaker
So thank you again. And maybe another year, will have you back and see where you're at. So thanks so much. Happy to do so. Thank you very much, Steve.