Episode: 8 - Jack Prior of Sanofi on AI Readiness in Biomanufacturing

In this episode of Bioprocessing Unfiltered, Jack Prior, head of process monitoring and data science and AI strategy at Sanofi, explores how AI has evolved from early expert systems to machine learning, generative AI, and agentic AI, and what that means for biomanufacturing today. He explains why digital transformation has not automatically created usable data, and why the real challenge is turning fragmented information from manufacturing systems into accessible, authentic, and analysis-ready data. Jack shares a practical framework for assessing data maturity and argues that the biggest near-term opportunity for AI is reducing friction in data engineering, troubleshooting, tech transfer, and decision support across bioprocessing organizations. 

Links from this episode:  
Bioprocessing Summit 
Sanofi 

Bioprocessing Unfiltered: Covering both upstream and downstream processing, analytics, AI and digitization, cell and gene therapy and more, Bioprocessing Unfiltered is your insider’s pass to the researchers tackling—and solving—the day-to-day challenges in the bioprocessing industry. 

Transcript

Welcome And Guest Introduction

Announcement 0:06

Welcome to the Bioprocessing Unfiltered podcast. Each month we host conversations with the researchers and leaders tackling and solving the day-to-day challenges of the bioprocessing industry.

Bill Whitford 0:19

Welcome everyone to this edition of Bioprocessing Unfiltered. We have a wonderful guest today that I've listened to before, so I'm excited to hear his thoughts on the topic this this morning. Jack Prior, is who is the head of data science at Sanofi. And he'll be speaking today on data acquisition, machine learning operation in biomanufacturing, and I'm so looking forward to his insights on the topic.

Jack Prior 0:55

Thank you, Bill. I am I'm a runner and I've listened to the Unfiltered podcast in my playlist, very high up. It's much more interesting than some of the news of the day some days. So I really appreciate the invitation to to join early on in this progression for the for the podcast.

Bill Whitford 1:13

Well, I've got a a few questions pre-prepared for you today, but I'm really interested in going off in the directions that that you find relevant, especially since the last message that I heard you deliver. So feel free to wander off my specific questions as we go forward. But I'd like to start with a little bit of your history in applying artificial intelligence and machine learning and and and proper data management to biological manufacturing.

From Chemical Engineering To AI

Jack Prior 1:43

Okay, sure. So I'm a chemical engineer, undergraduate at University of Connecticut in the US, and came of age really at the moment that personal computers came of age. So coming into the industry or going and going into eventually into graduate school, I thought, well, this is the thing that is novel in this day and age, right? How to help you know more established engineers use these new computers that were turning up in offices, and sometimes one in an office or something, eventually one per desk, but that was a novelty at the time. So I went to MIT for graduate school and interested in applying computers to chemical engineering or to process engineering, I was interested in process monitoring. I'd worked at Kimberley Clark making Kleenex faithful tissue, facial tissue as a as a co-op. So I arrived at MIT and was selected by Charles Cooney as a as one of his advisees. And his interest at the time was how do we apply artificial intelligence to bioprocess monitoring, to bioprocess understanding? And the caveat I like to add to that in retrospect is using five and a quarter inch floppy disk that could hold, you know, 360 kilobytes of data, 1.2 megabytes in the second iteration of those. So he had there was a graduate student there who had just finished working on expert systems. That was the kind of AI at the time. I would call it like AI 0.5, not even 1.0, but how can you apply, take a bunch of rules, take the expertise of experts, you know, put it all together into a computer system, have it iterate over that and draw inferences out of it.

Jack Prior 3:12

And so I couldn't find a way to really extend the work that had been done on that. And I took a more traditional nonlinear optimization approach to sort of searching for faults and failures and bioprocessed data. At that time, machine learning 1.0, I would say, was just emerging, use of principal component analysis, PLS, neural networks. And it kind of felt like cheating, like, oh, you're gonna just fit the data and have it give answers. It's not first principles. But the reality is we didn't really have enough computing power back then, and certainly not enough data in retrospect. Really, the it's you know, the Netflix, the Facebooks, the big, big data, you know, revolution of machine learning and artificial intelligence has really come with organizations and problems which have huge amounts of data.

Jack Prior 3:56

So I've gone from MIT into various roles at Genzyme, which became part of Sanofi in manufacturing science, manufacturing science technology. And I've always had data either entirely or partly my activity or something of supervising. So I've been kind of engaged in the space. You know, how do we learn from data over my entire career, you know, working with you know computational fluid dynamics, digital twins, that type of activity. And and now, 38 years later, suddenly the AI is ready. Yeah. So that's kind of the that's kind of my overall progression. I've always been interested in how do we make better use of data. And then you know in that journey, I would say the struggle has increasingly become how do I get the data necessary to launch into that activity. And so that kind of leads into the work that I presented at the summit a few months ago.

Bill Whitford 4:48

Oh, that you know, that little bit of bio there explains why you've got such depth. And when I hear you speak on a particular topic, it's obviously not just memorized from some chat GPT answer, but you know, you really provide this color, and that and that's why I've been so looking forward to this chat. So you've gone through a little bit of your

How AI Evolved In Practice

Bill Whitford 5:07

history. I wonder if you could give our audience a little bit of the history of the progression of of expert systems through our current machine learning models and ending up with with what our expectations are. Right now, if there were a complete fulfillment of what of what we have now, we're not looking to the future, but we know that what we can do is not being implemented very efficiently.

Jack Prior 5:30

So that has Yeah, I mean, I can and I don't know if I have exactly the frameworks academically, the way people would characterize the evolution of AI, but the way I think about it is kind of what I said expert systems, you know, trying to gather up a bunch of facts, expertise from a limited set of experts who might be retiring and shake them up and have an inference come out, maybe under be able to figure out why what to do when two DO pros diverge or if the viability of a cell culture drops, right? That that was kind of AI.5. And then I would say I one AI 1.0 was machine learning, neural network, sort of training from data in a kind of a brute force black box approach. But again, I think there you're limited by the amount of data you have. And actually, the art of that work is kind of subtraction, like take your full data set and then remove inappropriate amounts so you don't you don't train in the wrong direction. Use only your most recent process, take out batches that might have had some artifact that would confuse things. So it's it was really addition and subtraction.

Jack Prior 6:31

And then I would say I AI 1.5 in my mind is I remember going back to MIT to a conference and learning that things that evolve where you could have neural networks pre-train to understand an image, for example, probably the most impactful thing to me. Instead of having to train a neural network on all only on your images and have to take a thousand different angles, now there are these sort of starter models that are pre-trained on the world's information. So now it could see a duck or understand a graph. And so that ability to build a little bit on the world's information was exciting. I think probably found use in visual inspection, right? Now you could take a picture of a vial and and have it know that there was a spot there or a piece of debris. But in that area, you were a little bit limited by your training set. You only have so many defects. You know, so now you move to AI 2.0, more generative AI. Now you can generate images. You can say, give me a library of defects for training a more conventional sort of AI 1.5 neural network to identify things. And then in 2.0, again, we're getting this the LLM, the ChatGPT revolution, you know, the autocomplete, the ability to have harvested all the world's information, not just your small subset, but everything. It's having that come in. And now it's a little bit like AI.5. It's almost like the expert system concept, right?

Jack Prior 7:49

You instead of having a limited set of little facts and expert information, now you have a thousand forum discussions, a thousand repositories of you know, millions and millions of repositories of code and expertise. So it's almost like an expert system the way we had hoped for it to be, but with the with the data flowing in. And then you have I would call a sort of AI 2.5, which is you know beyond the initial Chat GPT experience people had in 2023, too, whenever that came out, whenever we saw it, of just autocomplete. Here's a question, here's a response, and kind of building on a an increasingly long autocomplete sequence. Now with Agentec AI, it's kind of a mix of the ability to have that chat discussion, but now that chat discussion makes a plan and executes that plan and then gives it back to you and the ability to interact. And now the instead of getting purely what's sort of coming out of the memory of the neural net or not, yeah, essentially the large neural net that is an LM, now you're getting the ability to go out into the internet, check facts, pull data from a process. And so it's another world.

Bill Whitford 8:53

Oh, that's great. Jack, you know what I appreciate is I know you know the the nuts and bolts of some of these approaches. And when you're describing the ability of machine learning to cut through all of the volume of data points and reduce it to salient, valuable information, you didn't use terms like dimensionality reduction and and contrasting to Nelson rules and things that biologists like myself don't understand. But you you you you're presenting it in story form, I think that's so valuable. If you'd like, I mean, and maybe we could take another term, but if you'd like, if you could talk a little bit about the distinction, this is what I picked up in your last presentation.

Digital Transformation Versus True ML

Bill Whitford 9:32

The distinction between the digital transformation, moving from paper to digital repositories, whether they're five and a quarter inch discs or or virtual, but the distinction between that and true machine learning models and what the the the the valuable massaging of data and and storage of it so it's accessible and the Alcorb type of rules, the distinction between that, because many biologists and engineers, especially who've been around for a while, I think are a little bit confused. We know that there's a digital revolution, but you really fleshed out some of the limitations of what we have been doing.

Jack Prior 10:15

Yeah, and maybe one point just to kind of follow up on the machine learning 1.0 era. I mean, I've honestly been a little bit of an AI 1.0 skeptic largely over my career, because I mean, if you think about if you have two points, you can fit a line, right? If you have three points, you can fit a curve, right? If you have five points, you can fit a curve, and maybe no one of the points is wrong.

Bill Whitford 10:34

With no with no machine learning. Yeah.

Jack Prior 10:36

Yeah, but I mean, essentially fitting a line is machine learning, right? It's an optimization that finds the minimum, you know, least squared error across the various points. And that's in a way what machine learning is doing, but just very large. But the when you do machine learning, if you go into what's behind a multivariate you know model or a or a neural network model, there are hundreds, thousands, millions of adjustable parameters. And you can't get water from a stone, right? You can only have you can only adjust as many parameters as you have experiments or manufacturing runs to lead to that number being set. And so we in our industry, at the time when we really need learning and the ability to discover things, have one batch, five batches, maybe twenty batches. Eventually we get 200, 500 for a mature process. By the time you get 500, though, you've probably already had to wrestle with most of the things. So I in general, I personally am a little skeptical when we I find that we it's extremely hard not to overfit a machine learning model, just limited to the data there. So I just I just want to make that point. But I think now what's exciting is you know, we're not getting just those hundred adjustable parameters from our own internal data set that we've organized. It's coming from the outside world in more in text form and in sort of knowledge and unstructured ways, you know, but but the opportunity is is changed. And I think we're with people, we almost need to have different words for this. I don't think AI 1.0 and 2.0 are the right word, but it's almost like it's almost like English and math or something, or addition and and multiplication. There's a lot of things getting brought back from 1.0 in the discussion and in the excitement. And we have with but the limitations exist in terms of what we can get from the raw data.

Jack Prior 12:18

In terms of transformation, I mean, I guess for the what I focus on in the talk I gave is digital transformation really hasn't yielded the data paradise that I would have thought 20 years ago, right? When people said we're gonna take all our paper batch records and we're gonna put them in manufacturing execution systems, MES systems, and there'll be no paper in the plan. I thought, oh, this is wonderful. It's just gonna come out in a nice big spreadsheet or for a jump or something you can analyze in various tools. And the people who purchased those MES systems, designed them, forgot to worry about me. They think they're worried about making medicine in a very safe, repeatable way. The the side effect of kicking out a spreadsheet that has all the data ready made is something that you have to work at. And that's true across the board. We have fundamental frontline kind of primary record systems across our manufacturing process. We have LIM systems, MES systems, historians, ERP systems, all having data that's key for really learning at the higher level, none of which set out to feed those systems. So the increasingly, as as we've gone into this era where we're fully digital, we've had to put more work into getting the data from those systems into usable systems. Whereas in the 90s or late 90s, you had a spreadsheet and it took a little time, but you more or less typed in what you need and got it. And now no one wants to type anything in again. It should be blowing to us. So I think that's the key learning is that just taking away paper, you know, doesn't lead to true kind of data transformation, digital in a in a way that you might expect.

Bill Whitford 13:53

It occurs to me in your description of the limited data sets we have from a true bioreactor run, for example, or fermenter run. But that AI is helping us in data governance and dimensionality reduction and preparing the the data that we have collected for its use ultimately as in building models. And you've alluded to this distinction between them. I wonder if you could we've talked about where we've been and some of the issues of applying machine learning to the systems, but in your talks you've referred to the work you're currently doing, the models you've built and and their application. I wonder if you could speak a little bit about that. How how you've if if not how you've developed them, but where you've implemented them and and what they've accomplished, what your successes have been like.

Jack Prior 14:48

Yeah, sure. And I can maybe talk, I could talk now or later, that I think one exciting thing about agentic AI, you know, vibe coding, this ability of these you know tools that are out there, and I won't just I won't name them, but to help us get over the hurdles of data wrangling and data engineering. Like I said, I think I feel like myself and my team at times have been consumed by the data engineering challenge. And then sometimes you don't have time left for the for the discovery and the exploration. And in my personal experience, I'm fairly hands-on with this stuff. You know, since November, December 2025, your ability to say, I would like to analyze this data data from this obscure thing, from my triathlon running watch, from my my sleep monitor, and just to get right to the data analytics is amazing, right? You're not having to go and figure out formats and and things like that. And that's gonna find its way into corporate work, right? Where we where we struggle to do that kind of work. It's gonna, it's we're gonna suddenly be able to do things overnight or in a week that might have taken months before. So I think that's that's an exciting thing.

Jack Prior 15:51

But I guess in terms of the work that I've done, you know, I guess the premise that I have is that as an industry, perhaps not every company, we've actually we're all struggling for the same fundamental reasons. And it's actually getting harder. It's not, you know, I take I took a little bit personally where I've seen our abilities not the uniformly getting better. You know, we you know, there's always kind of a a lighthouse that's okay, but you know, it's we struggle and I and I take it personally. Like what am I doing being responsible for in this area to not have it advanced? What am I not communicating correctly? And so I get number one, I think it's gotten harder because the if you look at most people's initial state is a spreadsheet and a small plant with a small team tailoring it to tailoring data to what they need. As you move to a larger company and globalize towards standards, now you have many processes around the world having to converge to a single standard, having that virtual spreadsheet be designed by people that are two or three parts removed from you. So I think it's getting harder. And so my approach has been how do we take the emotion out of that experience and you know, sort of you you know, stating what I just stated could be seen as a change management issue. You're resistant, you like things the old way, the way you're doing it. And it's not the case, right? So, how do we better convey what our requirements are? And so that was the first phase of this work.

A Six By Six Data Maturity Score

Jack Prior 17:13

You know, that to do bioprocess monitoring, troubleshooting, tech transfer, root cause analysis, all the things we need to do. We need six types of information. We need to know our quality trends, we need to know our productivity and yield trends. Those are two most important ones. Then we need our batch record offline data, we need historian data from our sensors, we need genealogy that ties everything together from end to end, from C train all the way to fill finish across the globe. And then we need the ability to calculate things from that raw data. And all six of those things often don't come together when you do and redo a system. In a spreadsheet, you tend to pull it together. So, first was that. And then having scored dozens of products internally with those metrics, the second hurdle I found was that you can have two processes measuring the same what, but very vastly different experience.

Jack Prior 18:01

And so what was the difference? And that led to kind of this how well dimension. So six what dimensions. The how well is can you get the data, is the data fresh? If you're working, I spent 10 years working in GMP manufacturing facility. You want the mornings, you want data now. If you're looking at a sensor, you want it a minute or less later. If you're looking at offline data, you want it that morning before the daily review. You don't want it the next day. So freshness is really key, and it can suffer when you start making data go around the world. The second part is frictionless. If you're tired on a Friday afternoon, but you have an idea, you want to be able to pour into the data without a lot of hurdles, without having to call someone. So is it fresh? Is it frictionless? The next dimension is is it accessible beyond the few people who have the data or who might, you know, is it can it go can it go out to broader groups of people? So is that and that can be a challenge. And is it authentic, validated? You know, in our industry, we have to work with data that has good data integrity. You mentioned earlier Al cola, right? So and that's a priority. And it's needs to be risk-based. The closer you get to real GMP decisions, the the quality has to go way up. Looking for hypotheses is a lower threshold. So authenticity, availability.

Jack Prior 19:11

And then the last two are is it's corporate standard? Is it the same across the company using the same software? Can an expert wander into various plants and be able to help without having to decode local specifics? And then finally, is it structured for data science? The when I say these things, systems aren't built for me, does the data ultimately put out a usable table of data? So those are six kind of different priorities with different stakeholders weighting them differently in practice. So a quality organization, 110% on authenticated or nothing, right? A an IT digital organization, extremely focused on standards, manufacturing site focused on freshness. And as as we move to a globalized system, that is if you're not making decisions to balance those however you'd like, but recognize it's a balance, I think we get in trouble. So my goal is score things six by six, numbered zero to a hundred, and then agree that higher number is better, and then work to improve it. What gets measured gets done. So that's the essence of the experience and the model I'm proposing. It's not rocket science, but I, you know, I think it resonates with people.

Bill Whitford 20:12

Well, not rocket science, but it brings true to me. You know, I'm I'm older and I recalled before any digitalization that there were very good operators, technicians, clerks involved in the in these systems who had beautiful paper records and knew what to do with particular information with respect to security, with respect to predetermined routing mechanisms. But it was slow. It was too slow. It was fine for me in RD, but it was too slow to be really useful in manufacturing. And not only that, but it seemed there was maybe only one person in the company that really understood the system, maybe a vice president of some of something. But the each individual terminal didn't really see the whole picture. But it that that I think is something we have now with an AL model is that it instantly knows the the entire system and and where things should be going and and what the deadlines are, which is you know which was rare you know 20 years ago.

Announcement 21:13

Are you enjoying the conversation? We'd love to hear from you. Please subscribe to the podcast and give us a rating. It helps other people find and join the conversation. If you've got speaker or topic ideas, we'd love to hear those too. You can send them in a podcast review.

Permissions Integrity And AI Regulation

Bill Whitford 21:30

Do you want to speak at all about you know the current status of of of there are more regulatory publications addressing artificial intelligence in pharmaceuticals and in manufacturing, and there are more best practices coming out from groups like the PDA or the ISPE that are are limiting or or directing or I and I heard you speak on this before unifying and co. Codifying particular approaches that they're recommending or insisting are the way we should approach particular systems. Do you do you want to go in this direction of the difficulty?

Jack Prior 22:11

I mean, I can talk a little bit about it, sure.

Bill Whitford 22:14

I know it's a very confusing time.

Jack Prior 22:16

Yeah, I mean, I guess to just a little bit on your first point, right? One of the questions I the model still has to answer when it when we ask the question, do we have the data is who is we? Right. And so sometimes there is a we, but it ain't me, right? And so, you know, there and there are probably you know, certainly there are people in local broad, you know, who are who are running a system who don't feel the need for corporate to, you know, or you know, help from other places to have as much data. So there's an interesting question around how what's appropriate data that for anyone to see. And it reflects a little bit on your second point, which is when you have a I had a vision in around 2008. I said, wow, if we could bring Wikipedia into the company and crowdsource editing of all this information that we tend to try to organize ourselves, wouldn't that be a wonderful thing? And but you realize really quickly that in a corporate environment, there's a lot of need to know philosophy.

Jack Prior 23:06

So you don't want one Wikipedia for your company to have every last secret in it, or and even things that are generally seen as as private. So I think we have that same trouble. We can't do the same, we can't just train an LLM on our all our internal information and have it tell us every bit of information from across the company. There are there are ways of doing that, of kind of having permissions cascade, but it's this you know, a basic LLM strategy won't do it. So what yeah, so those are two kind of what was back to the main question. Oh, the regulations. Yeah. So I think a lot of regulations, again, it's terminology. People are finally coming out to regulate artificial intelligence. A lot of what's happening is we're finally getting a handle on AI 1.0. How do you regulate a neural network traditional model being used for purpose X? Or how do you regulate the use of a multivariate model for per per purpose Y? And I think well, the ultimate conclusion of a lot of those, you know, I think they're reasonable regulations, you treat it like any model. You know, it's not y equals mx plus b, but it's y equals f of x. So, you know, fix it, you know, have accountability for it to work. Don't just have it auto-learn every day and potentially drift off the highway, you know, change it in a thoughtful way. So I think those are reasonable. I think where it's getting complicated is now everyone has used Chat GPT the month it was introduced. It was the fastest adopted technology in the history of the planet.

Jack Prior 24:29

And everyone saw hallucinations and said, oh my God, it's faking references. It's making up things. It said I wrote a book I didn't write. And so that kind of, you know, it is that kind of fear is going to flow into sort of AI 2.0 or even maybe affect AI 1.0. And I think that's where there's challenges. And I think we have a spectrum of opportunities. There's AI 2.0 or 2.5 frontline in our manufacturing processes, and then there's AI 2.5 helping us be better engineers and better process designers and better root cause analysis people. And the regulation there, you know, should naturally be less, at least in my opinion. I like to say that root cause has no Miranda rights, right? Getting to a root cause, however you get to it, can that you can prove in an authenticated way is a good outcome, right? So we sometimes get in the way when we focus too much, you know, particularly in root cause analysis and the systems required for that. You know, if we if we limit them too much, then you lose the ability to discover what's happening. So I mean, I'm still learning more about it, but I think we're it's re you know, we should behold ourselves to the higher standard and regulators should you know, they shouldn't be harder than what we would do ourselves.

Bill Whitford 25:37

You know, when I read in this topic, I in my mind, there's the issue of of people describe the limitations of a machine learning algorithm hallucinations and and stoichiometry versus the first principles. And but to me, it's what what's going on in the head of an individual making decisions, of an expert? I mean, we we we we see errors that that are that originate from algorithms, but individually we for I mean in the discussion I think we don't emphasize enough that there are individuals that make horrible calls, that make a mistake. They were up late the night before and they they call it wrong. And so I think there I think that these models don't have to be perfect. Yeah, we you know we we do want to work on safeguards to some extent, but they don't they certainly don't have to be perfect to be acceptable. You know, you you interested you introduced an idea that it just occurred to me. We talk about federated learning as a future of being able to data share. And I don't remember if you mentioned it in your work, but I'm interested in this to the degree to which we can limit the actual narrow data but share conclusions in building these models. But it occurs to me that maybe we could do this within an within a company too, have this type of federated learning of decisions and and including, for example, logistics and and and sales, so that we can gain the the appropriate information from the whole company or distribute it around the world, various locations, without sharing the details that that drive these conclusions. But my thought is we've been talking about a little bit of the history of ML and in pharmaceutical manufacturing development. You've talked a little bit about your applications, but if someone were to come to you, what what I've seen is there are many accomplished professionals that really work well in pharmaceutical manufacturing or product development that are very naive in artificial intelligence and they don't admit it, but they just haven't spent much time. They've used Chat GPT, but they're not mathematicians and some aren't even engineers, they're business people. So, my I thought if we could spend a little bit of time, what if someone says, okay, I'm sold.

Where To Start With AI

Bill Whitford 28:00

I see the value, I've read a couple of papers, I see what it can do, and I understand in cartoon form what's going on, but where do I start? How do I get how do I start applying this in my company? We we've got a couple of engineers good in math, we've got a couple of people who have been working with the data, but do I hire a consultant? Do I comp fire everyone and hire 22-year-olds? How do I get started?

Jack Prior 28:27

Oof, wow, that's a few questions there. So I just I mean, you raised a few points I wanted to speak to. I mean, I personally haven't had experience yet attempting federated learning. I've seen it mentioned a few times, but it's a little, it's gonna be a little scary to a company. You come in and say, hey, just give us your data, it'll all be abstracted and you're not revealing anything, and we'll combine it with a few other companies. And but I what I do think is we have to start to doing it internally by one mechanism or another, because you think about it, every process to us is unique. Oh, this process that has all these particularities. But ultimately in the industry, we're you, you know, and even within a company, we're using a platform process with similar purification steps. We're using Cho culture in many cases for similar times, yet we that we keep those data entirely separated, right? And the fact that you know lactate goes up when cells are stressed or the pH drops when lactate goes up, those kind of general kind of expert system rules aren't being applied to help us in the you know the the final thing. So I think there's an opportunity there. And if we really leverage that, then we might say, okay, this is great. Now let's now I understand how I what how this works, how to how I can anonymize my data. Now let's collaborate across the industry and and and grow the pie.

Bill Whitford 29:35

And have some failures internally. Clean it up, clean up your system so you you debugged it before you you go live, yeah.

Jack Prior 29:43

Right, right. So in terms of where do you get started, I mean, I guess one of the points I like to make is there is an inflection that's occurred in the power of you know agentic LLMs, right? Again, this is a sort of a dial spectrum between the basic LLM you saw on the first day in 2023 and what's in a fully automated dark factory automated agentic, right? So the ability for LLMs to do agentic things is going up. It was doubling every seven months. To me, it looks like it's now doubling every three months as the AI programs the AI that programs the AI in these and these big companies. So that's my first advice to people is this is not your father's, you know, board escort or whatever, right? That things are changing very quickly. If you formed an opinion, dip back in and understand personally in your personal life what's possible. I find right now our the limit is is our imaginations, right? It's this is not entirely overhyped technology.

Jack Prior 30:39

So yeah, and again, I think think out the box, think we have to think outside the box and not jump right to the worst case GMP use case, right, frontline, because I think we do have to think about when can a human in the loop do the work, right? Can we expect an operator on a Sunday morning to know if a machine learning model is not is outside of its assumptions, right? Almost anything that we could proceduralize, we could just computerize on that front. But we can do more around accelerate tech transfer. So I think there's a lot of interest in kind of root cause agents helping people get to the data they need. So I think that's how to, you know, I would get started at home by exploring what's possible, understanding how much things have changed. I was shocked even talking to some academics at a school I won't mention that they were convinced this was all going to plateau in 2023. It was overhyped. It was plateauing. I don't see the plateau yet. It could be imminent. I think if it stopped where it is now, we'd we could go for a decade before we'd really start to leverage what we have today. And then I think we agree. I would just circling back on the maturity model side, like how do we get how do you get a handle on your own internal data? I mean, the way that you know I get to set the weights in this model, to me, 30 % of the 100% score is having a great handle on your quality attributes, both drug substance and drug product.

Jack Prior 31:52

That is a floppy disk of information for an entire company. It can be surprisingly disparate. And someone asked me today, you know, what's what are what are our bioreactor titers across all our map processes? You know, not every company has that in one spot. So I would say, you know, and that's a 30% score. 30 %. So that's the second part. Yield, productivity, 30% quality, 30%. You get a 60% score, you're almost to a passing grade just if you do those two things well. So that's where I would start and then and then just gamify the situation, right? You you put a score on every product, every plant, and then have senior management under buy into it and make plans against it. That's where we are. Let's we've made good progress in a year since defining the scoring things last year, and we're making we now are making very concrete plans to to to improve it. So I say like make I say make it a game.

Bill Whitford 32:37

Okay, Jackie, you you've set that up. And now my question is so do you think that these are we're talking about people, intelligent people who are naive to machine learning?

Jack Prior 32:48

Oh, right, right.

Bill Whitford 32:49

Okay, so so should they buy a canned program or or or use one online? Should they hire a consultant to guide them through? Should they hire a half a dozen new graduates who are experienced in AI? How do they actually get into the nuts and bolts and apply it?

Jack Prior 33:07

So you're talking about a large organization or an individual? Yeah, with money. Yeah, yeah, yeah, yeah. Yeah. I mean, I think the interesting the interesting contrast and how an organization should pursue it.

Bottom Up Tools Versus Top Down

Jack Prior 33:18

There's a definitely schools of thought and a lot of news articles out there saying AI pilots fail 975% of the time or there's no return on investment. So there's a lot of on having top-down governed, funded, prioritize efforts in AI, right, to overcome that, right? Really define the problem, fund it adequately, govern it and track it and really work towards that. The other side of it is empowering tools, I think, are really important. I think you know, the most empowering thing a company can do is to take one of these commercial LMs that are out there, you know, on the broad internet, bring them and turn inside, ring fence them so you don't have data leakage, and empower from the bottom up for people to innovate in their work. I mean, I like to think about the the the reason we have a PC today in the phones is because the IBM PC brought Lotus 123, which became Excel, to the market.

Jack Prior 34:11

And if a PC only did only had Excel on it, you would still have them, right? And it but Excel was not done top down by governance. You didn't sit down and say, hey, corporate IT group, can I like to have an Excel spreadsheet? Can I have these fit 30 columns with these equations? Oh, they weren't quite right. Can you come back in two weeks, right? So how do we do this AI transformation in the same way with people having the autonomy, you know, kind of Excel for agents, I think is the key. And I'm seeing that, you know, now is become, you know, I think successful companies are doing that. They have a very strong internal L where no one would think to use their home account because they're getting the same power at work. And then you start to empower them a little bit of agentic coding or agentic ability. Obviously, we have to be careful, but I don't think you can have solutions come purely in a structured way.

Bill Whitford 35:00

You know, I remember the story of Digital Equipment Corporation who is competing with IBM, and there's a quote from the president who at a meeting famously said, why would everyone want a computer in their home? Shortly before DEC went out of business. Supporting your point about the natural grassroots development of these applications. Well, I think is there anything else you'd like to mention that I have I'm interested and excited about this topic.

Jack Prior 35:30

No, no, I think no great, great questions, and I certainly got all my favorite items out. I could do we could do this for hours. I'm really it is really it's an really exciting time to me to, you know, it's just something I didn't anticipate, right? The the the this really revolutionary ability to to have to take the friction out of a lot of the things we want to accomplish, right? To not have coding or UI or our data wrangling be the bottleneck, right? It's really imagination. I that sounds like a little bit overhyped, but I'm personally, in my personal experience, and I've tried about a dozen things at home to say, hey, could I finally do this? And the the hurdles are not there the way they were even two years ago, right? You you you don't get into these hallucination loops. So I think it's that's my main take is this is an exciting time. Don't let your your imagination limit you and get get to the hard use cases, you know, the hard GMP use cases that we really have to be careful about by regulatory incrementally by being productive in all aspects of what we do.

Bill Whitford 36:30

I remember at the turn of the century I had a flip phone and I hesitated on getting a smartphone because every time a new version came out, people criticized it and said, no, wait three months, there's a better one coming out. And so I hesitated and finally regretted it and just jumped in. And I think we're the same state with with applying all the the powers of AI and the and the hundreds of applications, specific applications, from from data acquisition and management to forecasting and and sales, that as you say, the you're it's only limited by the the limits of your imagination. There's so much power capability here today, today, not the future.

Jack Prior 37:14

Yeah, well, luckily we're renting, we kind of rent what we do versus buy what we do. I had a father-in-law who waited 40 years to get the perfect laptop, and when he retired, he finally got it. But he was shopping the whole way.

Bill Whitford 37:26

So I did that with phones. It's a flight, it's a it's a mistake.

Jack Prior 37:31

So, you know, here, you know, it's you know, personally, it's $20 a month. I mean, I think companies do have to be careful, you know, the companies are being careful not to lock in to a particular technology.

Staying Agile As AI Accelerates

Jack Prior 37:40

And I think that's one interesting thing that's is a challenge here. I was a professor at a talk last week at MIT said that the technology is increasing exponentially, organizations change linearly, right? And so I don't know if that maps directly, but one of my concerns is when the things that you would do today, like if you said, well, the project we need to do is X, Y, Z, use these tools to accomplish this, you know, imaginary, you know, fantastical goal. And so you launch the project to do that in a year, that's just a standard offering of the off-the-shelf tool. And so you have the the risk of that as you it's like it's like the GPS in your car. The GPS in your car was, if it was state of the art, it was state of the art three years before the model year. So no one says, I can't wait to use the GPS in my car. It's there in a crisis. So that's the thing I wonder about with AI. We have to be, and I've talked to folks about this, you have to be agile enough that whatever you're doing can re-plug into whatever's happening. And at a certain point, you have to say, you know what, I've the thing I was trying to do, it's just there in the menu.

Bill Whitford 38:40

Yeah. Oh, that's great. You know, Jack, I this has been very exciting. And as you say, we could go on. You're such a resource of practical information. As I said earlier, you not only see the vision and the applications, but you know a little bit of the nuts and bolts to help you get people going. I think this has been great. I'd like to thank you for your time. I'd like to thank CHI and and Bioprocessing Unfiltered for the opportunity to present this. I'd like to remind people that that you have some videos online, you've got publications that that better that go deeper into some of the points we've talked about today. And thanks very much.

Jack Prior 39:20

Oh, thank you. And yeah, you can probably navigate to most of it from my LinkedIn profile. So thank you. And again, a super honor to be here with you in this this podcast. It's a phenomenal addition to the to the world of things that you can listen to when you're out running or driving. So thank you for what you're doing. Great.