Anthony Mire-Sluis on Where AI Fits in Quality
September 1, 2026
Quality leaders are increasingly tasked with deciding where and how AI can add value without compromising quality. In this episode, Anthony Mire-Sluis, Ph.D., senior vice president of global quality at Gilead Sciences, and host Ran Zheng discuss navigating the intersection of quality, digital transformation, and AI in biopharmaceutical manufacturing. They explore how leaders can determine whether a process is ready for digitalization, avoid common implementation mistakes, and distinguish between AI that supports quality decisions and AI that makes them. Mire-Sluis also explains what meaningful human oversight looks like, how AI could help prevent human error and strengthen quality systems, and how digital transformation may reshape roles for operators, analysts, and quality professionals.
GUEST BIO
Anthony R. Mire-Sluis, Ph.D., Senior Vice President, Global Quality, Gilead Sciences Dr. Mire-Sluis is currently head of quality for Gilead Sciences. He was head of global quality for AstraZeneca and vice president, North America, Singapore, contract, and product quality at Amgen Inc. Dr. Mire-Sluis was also the head of the Cytokine Group in the division of immunobiology at the National Institute for Biological Standards and Control, Potters Bar, U.K. Dr. Mire-Sluis specialized in the development of assays for the characterization and quantitation of biological products. He then became director of bioanalytical sciences at Genentech. Prior to joining the FDA, he was executive director of analytical sciences at CancerVax Corporation, San Diego. He is the chairman of the IABS Biotherapeutics Committee, vice chairman of the USP Biologicals Characterization Expert Committee, an expert for the International Committee for Harmonization, and on the board of the Journal of Immunological Methods. He trained in genetics and biometry and has a doctorate in cell biology and biochemistry.
HOST BIO
Ran Zheng, Former CEO, Landmark Bio
Ran Zheng is a biotechnology industry veteran who brings more than 25 years of experience in biotechnology operations across multiple geographies to Landmark Bio. She most recently served as chief technical officer at Orchard Therapeutics, a commercial-stage global gene therapy company specializing in HSC based gene therapies. In this role, she established the technical operations function and manufacturing network and advanced the company’s product pipeline, including the approval of Libmeldy, the first gene therapy product for metachromatic leukodystrophy. Zheng has also held leadership positions at Genzyme (now Sanofi) and Amgen. At Amgen, Zheng played a key role in building differentiating capabilities in manufacturing for clinical supply and commercial product launch to enable speed to clinic and speed to market strategies for Amgen’s innovative products.
TRANSCRIPT
Welcome And Guest Introduction
Announcement
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.
Ran Zheng
Hi, I'm Ran Zheng. Welcome to Bioprocessing Unfiltered. Today I'm sitting down with Tony Maris-Sluis, Senior Vice President, Global Quality at Gilead. Tony, thank you for joining me.
Anthony Mire-Sluis
It's a pleasure and glad to be here.
Ran Zheng
Thank you. So,
Why Quality Leaders Want AI
Ran Zheng
Tony, you have a long career in quality and regulatory. What got you personally interested in the intersection of quality and the digital AI?
Anthony Mire-Sluis
You know, for me, quality is all about continuous improvement and making better quality decisions and using real data to base those decisions. And I think digital and AI is really enabling that to be much better than it has been in the past. So I'm excited about the changes that are coming with AI. And you know, many companies have been implementing it and it's already starting to make a difference.
Ran Zheng
That's great. So I'm sure you have some experience in this part. And when you start implementing AI or digital technologies, how do you tell if a process is actually ready for digital tooling versus need some rework? And see, you know, if it actually fits to the AI flow?
Anthony Mire-Sluis
Yeah, I mean, it's a great question. And my number one answer is speak to the people that use it. So, you know, if you're thinking about change control or an operator on the line, is to speak to them and say, is your process optimized? Is it easy to do? Is there an easy flow in it? Is there anything we would like to do to improve it before we digitalize it? Because my mantra always is if you take a complex process that the operator or quality professional finds complicated and digitalize it, all you've done is made a complicated digital way of working. So for me, the foundation of what you need to do before you digitalize or use AI is to simplify the flow of the process that you're looking at.
Ran Zheng
Well, that's good insight. So, what's the biggest mistake you have seen in the past while you're implementing AI digital technologies and a site or you know, organization could potentially make before fixing the underlying issues?
Anthony Mire-Sluis
Yeah,
Avoiding Failed Digital Rollouts
Anthony Mire-Sluis
I mean, I think like you mentioned, if you don't fix the process and digitalize it, you don't really help anybody. But also, I think when you look at the projects that people create to digitalize something, it really has to be with the end in mind. So talking to the users themselves, getting their opinions. It can't be a top-down process. The other thing what I find is that when it gets hard, people may be too busy working to a timeline as to getting it right first time. And again, if you roll out a digital process that is only half ready, you drive the business crazy. So to say to somebody, well, we promised to give you something in six months, but it's not quite ready, but you can use half of it, really doesn't help anybody. I actually always say it takes more strength to say no and extend timelines than it is to get to a timeline, finish, and then roll out something that isn't right first time. So it's really making sure that you've got everybody's input and trying it out too. One of the biggest mistakes I've seen in my career was where something was digitalized and it was rolled out on the floor without actually having real use data. And it almost brought that site to its knees because the systems didn't work. It was a packing site, orders were not getting filled because the system just didn't work properly when it was in the hands of the operators. So really making sure that you've done all the proper test cases and really used it in real time with real data or real machinery is extremely important.
Ran Zheng
Okay, so don't give us a half-baked solution. Absolutely. All right.
Anthony Mire-Sluis
Got it in one.
Ran Zheng
Yeah, that's that's that's a good tip.
Human In The Loop In GMP
Ran Zheng
You know, we we use the AI. AI is almost everywhere. We use AI in a lot of daily lives. Sometimes we ask questions, we search for information, and sometimes we use AI to assist in decision making. So, what where's the real line today between AI that assists a quality decision and AI that makes one?
Anthony Mire-Sluis
Well, so as we know, we're a regulated environment, and the regulators have been quite clear, AI cannot make a decision. So AI can only assist, it can provide data, it can clean up a report that you're writing, it can help you understand root cause, it can search databases, but it itself cannot make a decision. And that's the real concept of the human in the loop perspective is it's an aid, it can provide data, it can help smooth the process, but the ultimate ultimate decision has to be with a quality professional.
Ran Zheng
Great. So the new FDA EMA framework restrict generative AI to non-critical functions with documented human oversight. Practically, what does good human oversight look like versus a you know kind of a rubber stamp version of that?
Anthony Mire-Sluis
I mean, it's a great question. And personally, I think that from a risk perspective, AI is no different than what we were using autocorrect for or grammar checks. At the end of the day, the human has to read what the AI has come up with. Often, if it's used particular data sources, it's just a double check. Those data sources are real because as we know, AI can misinterpret data or make things up. And I know agencies are concerned that when we get more and more used to using AI, we're going to stop using our brains and just accept it. But you know what? I always say that can be exactly the same as a QA signature on a document now. If a QA person doesn't read the content and sign off on it saying, with the meaning of I've read it, I've understood it, we're no different now than we would be when we're doing the same thing with something that AI has generated. So part of it is going to have to be training. There always has to be a certain level of trust that a QA signature means somebody's actually read it and understood it and checked it. Because to me, it's no different from where we are now.
Ran Zheng
That's that's a really good analogy. And the human overset almost like a QA, a quality assurance step to AI. That's great. So
Proactive Quality Beyond Dashboards
Ran Zheng
what does a genuinely proactive quality system look like a day-to-day versus the one that just more a dashboard? And just in the context that now we have AI that can almost create infinite number of dashboards for you, right?
Anthony Mire-Sluis
No, I mean it's it's a great question. And I think proactive quality is where every quality professional should be looking towards to create. And I believe it starts right from the beginning of product development, developing your process, developing your procedures, particularly with humans in mind. So if you think about it, one of the most prolific reasons for having deviations is human error. And there is a new movement now called human and organizational performance that really looks at how you humans interact with the environment you give them. I always say nobody comes into work going, today I'm going to make a mistake. You know, if somebody makes a mistake, it's because we've created an environment that allows them to fall into a pothole, as an example.
Anthony Mire-Sluis
So using those tools that are now reasonably widely available to look at if I'm going to design a piece of equipment, how can I design it so somebody's not going to push the wrong button? Or the sequence that we've created for them is too complicated. Or even a procedure is to read a procedure with a human in mind, going, you know what, this paragraph isn't clear. Somebody could do A instead of B or whatever. And I think AI and dashboards can help. But again, it's using risk management as an example. Quality risk management in theory is about preventing errors and risks before they happen. In our industry, quite frankly, at the moment, we use risk assessments reactively. Oh, something's gone wrong. Better do a risk assessment. Did it impact product quality? What's the compliance issue? As opposed to creating risk assessments before you even start when you're designing a product or a process. So I think there's lots of changes we can make in the industry to truly be proactive. And AI and the dashboards and tools are just one small piece of that that will help with it moving forward.
Announcement
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.
Using AI To Reduce Human Error
Ran Zheng
So, what do you think if AI could potentially help to enhance the environment that the human perform their task so that will potentially reduce the tendency of making mistakes?
Anthony Mire-Sluis
Well, I think it's it's definitely a possibility. I mean, there is technology out there, which is really interesting, where you can actually watch, for instance, in a sterile suite, the AI can actually watch what a human is doing, which is amazing. It can watch how fast they're moving, it can watch whether they've touched something, it can watch whether they're about to knock into something and actually set alarms. And some companies have already started using that technology where, for instance, an alarm will go off if somebody's moving too quickly, or they haven't, , for instance, if they say spray your hands with disinfectant every 30 seconds, if you haven't done it, the cameras know, and they'll also cause an alarm. So there's it very many different interesting ways of AI looking at what humans are doing to try and reduce the possibility of somebody making a mistake. It can read a batch record and know, right, a human has to go from this side of the line to that side of the line. Oh, they're not doing it. Maybe something's wrong. Line clearance, it can follow somebody doing their line clearance and again come up and set an alarm if somebody's missed a step. So there are various aspects that people are coming up with now with technology that can help with that.
Ran Zheng
That's that's great. You already covered my next question, you know, thinking about how some of the you know technologies could really help catch human errors. You know, maybe switch gear a little bit. We have seen a lot of examples AI used in research, discovery research, even drug design, and you mentioned about risk assessment. Do you see any examples that using AI to predict product quality of a batch were help to identify the vulnerabilities of bioprocessing?
Anthony Mire-Sluis
Absolutely. I mean, you know, we know bioprocessing is a complex, complex process. There are you know multitudes of parameters we could be testing. And I think that's where PAT, process analytical technology, didn't quite reach its potential because there is so much data to analyze that nobody really knew exactly how to do it in a reasonable way. You know, I mean, in the in the days when it was first started, the equations people were using to predict if something was going to happen was so complex. Even trying to work with regulators to get those things approved was very, very hard. I think with AI now, able to look at multiple sources of data all in one go, looking at digital twins and all the data you have during development, the there's much more likelihood that we could use AI to be watching how a perfusion reactor is working, looking at oxygen levels and CO2 and acidic levels, that sort of thing, and say, oh, you know what? Things are starting to change. Maybe now you can quickly go in and alter something, or even alter it itself, alter the feedstock or order, change the way that the amounts are being added, maybe even affect temperature or flow, things like that. So I think the potential really is there now.
Ran Zheng
And then we would have to rethink how our control strategy is going to look like for a product or manufacturing process. Absolutely.
Anthony Mire-Sluis
I mean, the old mantra of the process is the product has always meant that people have always worked with him very tight in process controls, etc. That was always the promise of PAT is that you could move within your design space and end up with the same product quality. But again, I think it became so complicated that most people didn't really use it. But I think now with the use of AI and improved technology, we can probably start making inroads into that.
Ran Zheng
That's that sounds great. So talking about the technology implementation, particularly AI, machine learning, digital tools, it always starts people, right? It always
Data Strategy Jobs And Training
Ran Zheng
starts human. What you think just from your perspective? Should the transformation start from the top, or should it be taken from a bottom-up approach?
Anthony Mire-Sluis
So I believe it's a mixture of doing two different things. I think upper management has to support the funding and the time it takes to actually implement AI. I mean, this is not an easy thing. And one thing most people often forget is the data itself, is AI is only as good as the data you feed it. And trying to clean up our data sources is not a trivial matter. It's not seen as value-added or sexy because you're going into databases and trying to clean stuff up, but it takes time and it does cost money to do that. So you need upper management support to be able to do these projects properly. On the other hand, it's going to be the operators and the lab analysts and the QA staff that are going to be using these AI programs. So asking them, what do you believe is non-value added work? What do you see are the complicated parts of what you do? How much time are you searching for data? What databases are you looking at? Form the foundation of useful AI projects. Because you can AI anything. That does not mean it's truly going to add value, particularly if all you're doing is AIing something that didn't add value in the first place. You just created another way of working.
Ran Zheng
Very interesting. There are a lot of news now. People worry about AI gonna take their jobs. And well, in some cases, things actually happened. So you mentioned about early on that you know, AI or camera can watch people, can remind people not to make a mistake, and with robotic technologies, and one day, you know, the operator's job can potentially be replaced by technologies such as robotic and AI. What actual conversations is happening today with someone just like operators or analysts to help them to understand what AI means for their career?
Anthony Mire-Sluis
You know, it's it's always a difficult subject because there are going to be parts of people's jobs that will be replaced by technology and automation, et cetera, and AI. You know, I think it's twofold, particularly if you're in a in a growing company, is to grow without multiplying the number of staff by removing redundant work, by removing non-value added work. I always use the example of batch release, where some staff doing batch release can be searching six, seven, eight databases, trying to double-check data, getting the source data, particularly if if you're still on paper in certain areas of your business.
Anthony Mire-Sluis
You know, that doesn't help AI. AI doesn't read bits of paper unless you scan them in and do all that sort of work. So I think the conversations should be saying that AI is there to help you do your job, to make better decisions, to speed up the work you do, so that you're not having to spend your time searching through databases. It can present you the data up front and to actually be able to use your brain. I mean, that's what people would like to do when they come into work is actually be thinking and analyzing data. And I often find in our business we we can't see the wood for the trees. There is so much basic work that we're doing that we don't give people enough time to do the analysis that we were talking about earlier. Predictive analysis means people actually have to be able to look at data, present it in a way that they can understand so they can actually come up with answers. And I truly believe that using AI and other tools like that will allow us to make better decisions and then hopefully be proactive to be able to identify things before they go wrong rather than afterwards, and then we're all scrambling to fix stuff.
Ran Zheng
Great. Do you think today's quality professionals or manufacturing staff have the skills to help them to integrate AI into their daily work to become more efficient and to really help them to think at a higher level?
Anthony Mire-Sluis
I think it's a it's a journey. We could say the same thing about automation is the if you look at operators that maybe we had 20 years ago, I look at it like the mechanic in the in the auto industry. When you brought your car in, somebody would lift the hood up, a mechanic would be looking at all the all the bits and pieces to solve an issue. Let's be honest, these days a little light goes on at your dashboard. The technician, they don't like being called mechanics anymore, the technician will plug in a computer and will be analyzing how your engine is performing through a computer. So, yes, they still have to have the skills to change your oil and that sort of stuff, but it's raised it to another level. And I think the same thing has to happen with our staff, is we have to train them appropriately. Operators may now be called technicians, as more and more digital and automation actually occurs on the floor. They're not just pushing buttons, they're watching how it's working, they're looking at data in real time and being able to fix things as I say. If something starts going a bit awry, they can fix it before it goes wrong. But, you know, it's all about training, explaining to them about what's coming, and giving them a time to get used to it.
Ran Zheng
Great. So if you have a one piece of the YSO for quality leaders about kicking off digital AI initiatives, what it would be?
Anthony Mire-Sluis
Be patient
Prioritize Use Cases And Share Learnings
Anthony Mire-Sluis
because there are many, many different things we can use AI for. And I think it it's even publicly available that some companies went so overboard with introducing AI, it kind of caused chaos because you had multiple people creating very similar agents and ways of working, that they just flooded the business and their IT folks with multiple AI use cases where there had been no prioritization. So in Gilead, as an example, we've been very thoughtful about what we're doing. And each function in Gilead had to prioritize one or two of the what they believed were the most impactful user cases that AI would have. And in the whole company, that came to about 19. So we're not talking hundreds that were supported by upper management, that were totally funded, where the systems were created with IT hand in hand, training was built, et cetera, and then scaled across the organization.
Anthony Mire-Sluis
Functionally, when we talked about risk within a function, you can do the smaller AI use cases that aren't going to require a massive investment, et cetera. But again, everybody should be looking at prioritizing what is most important. Because, like we talked about, flooding the business with tools that operators and analysts aren't ready to use there yet is not helpful. So prioritization is important. And really looking at what is going to add the most value to the organization and pick those off first. Get your workforce used to it. And then I actually believe they will come up with their own answers. I've already started seeing that. Where once they get used to using it, they'll go, Oh, you know what? I could create an agent that would help me with this with this particular part of what I do. And then we can share it across all the different sites. So as I say, I think take it slowly in a stepwise, rational fashion and not go crazy.
Ran Zheng
So be patient.
Anthony Mire-Sluis
Be patient.
Ran Zheng
Take a stepwise approach and don't go crazy. Okay. My last question: where do you want to see the industry in three years from now on this AI initiative?
Anthony Mire-Sluis
I mean, I think, and I'm always hopeful, particularly with quality, is that we we share a lot in the pharma industry, that quality is to a certain extent a competitive advantage. But at the end of the day, it's it's what we do for patients. And so I'm hoping that as each company is coming up with useful AI tools that maybe not share the actual tool itself, but talk about what we're doing so that we can share our experiences, because that will help speed the implementation of AI. And then learn for when things go wrong. I always say in quality, why would we not want to learn from each other so that we don't all reinvent the wheel? If something doesn't work, let's discuss it. And as an example, I sit on a chief quality officers group where there's the top 25 pharma chief quality officers, and we often benchmark with each other confidentially. We don't mention, you know, come, you know, the actual company data, but share our learnings and that sort of thing. And it's extremely helpful. And as I say, that's what benefits the industry and benefits patients. But in three years' time, I can imagine that AI will be solidly embedded in our quality systems with the appropriate human oversight.
Ran Zheng
That's that's great. Well, thank you so much. It's such a pleasure talking with you and hearing about your your perspective on AI and the implementation in quality and power manufacturing. Thank you, Tony.
Anthony Mire-Sluis
Thank you, and I've really enjoyed it, and I'm gonna be able to do it.
Stay Connected
