“AI is marketed as such an approachable, ubiquitous thing that everybody thinks they can vibe code it. Unfortunately, a lot of companies are starting to think that way.” says Kat Kozyrytska.
Kat Kozyrytska is a portfolio strategy and technology leader in life science technologies who has operated across organizations, including Sartorius and Thermo Fisher, and now advises across the life sciences ecosystem. Her work focuses on translating emerging technologies, including AI, automation, and digital capabilities, into portfolio strategy, partnerships, and business growth.
Ahead of CDMO Live Americas, Kat explains why many of the assumptions shaping AI adoption in drug discovery do not translate into biopharma manufacturing. She covers the evolving questions around intellectual property and learning rights, why building AI in-house may be the wrong long-term choice for many companies, and why manufacturers need to move from experimentation toward portfolio-level prioritization if AI is to deliver sustainable business value.
Why Discovery-Stage AI and Manufacturing AI Solve Different Problems
Much of the excitement around AI in biopharma has come from drug discovery, where the technology has produced real, visible wins. Kat argues that manufacturing is a fundamentally different problem, and that treating the two the same is where a lot of strategies go wrong.
“The difference between discovery, where we’ve seen such excitement when it comes to AI, and manufacturing is that these two domains are solving vastly different problems,” Kat says. “With discovery, it’s about increasing the probability of getting a right answer. In manufacturing, what we need is the right answer every time, in every site, within a very controlled, regulated environment.” The specific challenge, she adds, is that AI is non-deterministic, which is in some ways completely contrary to what manufacturing requires: the same answer and reduced variability.
Kat also points out that this is not a new challenge. Companies have historically underestimated the distance between an exciting prototype and a compliant, scalable product, and AI is simply the latest technology to run into that gap.
The IP Puzzle: Who Owns the Learnings?
One of the more provocative points Kat raises is that AI is creating intellectual property and learning-rights questions that the industry’s frameworks were not designed to address. Historically, service providers have kept projects for competing sponsors separate at the level of people, systems, and contractual boundaries. Algorithmic learning complicates that model because the value of AI can come from identifying and aggregating patterns across data sets.
“This is a big responsibility for service providers. sponsors and technology providers, to understand how we will handle IP within this space,” Kat explains, “because our frameworks for IP handling are really designed for humans handling data and humans touching experiments. We don’t have a framework in place to understand what that looks like when an algorithm touches the data.”
She describes what’s happening today as inertia: companies are handling contracts the same way they did when it was just humans, even though it’s a brand-new space. She sees a real need for service providers, technology providers, and sponsors to make a true strategic decision about what stays within each party’s domain.
This question becomes sharper once CDMOs start building digital infrastructure for sponsors who operate primarily on paper records. Will CDMOs retool to become digital powerhouses within the ecosystem and enact the digital transformation for small and medium size biotechs? If they build digital infrastructure for customers and train a model across customer data, will they stand to own at least a part of that model? How does the historical process expertise of CDMOs translate into model ownership in the digital age?
These questions do not yet have settled answers. As CDMOs, sponsors, and technology providers develop AI capabilities, the boundaries among customer-owned process knowledge, provider know-how, and learning encoded into models are still being negotiated. Where that equilibrium ultimately lands could materially reshape the role—and value—of service and technology providers in the biopharma ecosystem.
Build, Buy, or Partner: The Bioreactor Lesson
Faced with AI hype, Kat says many companies default to building their own tools in-house, partly because AI is marketed as something anyone can pick up. She thinks that instinct will prove wrong for most organizations, and she uses a familiar piece of manufacturing hardware to make the point.
“AI is marketed as such an approachable and ubiquitous thing that everybody thinks everybody can vibe code it. Unfortunately, a lot of companies in the space are thinking that way, and in the long run, I don’t think that will be the default path,” Kat says. “There’s such a long road between something you make in-house that works for one team and something that works across sites, works for regulators, and can be updated over time. We see some companies building their own bioreactors, but most of the time, the bioreactor will be purchased from a specialist who makes bioreactors. I think we’ll arrive at a similar equilibrium for AI, where a company that specializes in making AI is whom you buy from most of the time.” Her expectation is that internal development will be reserved primarily for capabilities that are strategically differentiating.
She adds an important caveat for CDMOs specifically: building or acquiring AI capability is not simply bolting a new tool onto the services business. The development cycles, commercial approach, and quality systems for a software or AI business differ from a manufacturing services business, in much the same way that a CDMO deciding to sell media would be entering a genuinely different kind of business.
From Pilot to Portfolio: Making AI Prioritization Actually Work
The industry is reaching the point where isolated AI pilots need to give way to portfolio-level prioritization. Kat’s advice is to treat AI initiatives the way any mature corporate function treats a limited budget: with structured, portfolio-level trade-offs, not one-off approvals
“Moving from experimentation to scaled deployment, you really want to look at the entirety of the cost of that project,” Kat says. “What will it take to get the people on board? What will it take to get the data in place, standardized, AI-ready? What will it take operationally to run this month to month, since this is an ongoing project? And finally, what is the business value at the end of it?” Putting all of that together, she says, lets a company implement business ownership and actually prioritize between projects, including saying no to some of them. She compares it to a familiar problem in other corporate functions like IT, where the cost of a project sits with one function and the business value sits with another. Without a mechanism to bring those together, “there’s no real way to prioritize.”
Kat also flags a supplier-side version of the same problem. Emerging AI technology companies are often eager to land a pharma or biotech partnership to close a revenue target, and will take on projects that were never assessed for the broader market. Those one-off projects fragment a vendor’s roadmap and pull focus away from the work that actually serves its long-term strategy. Kat’s view is that both sponsors and technology providers need the same portfolio-level discipline, or the whole industry risks the pattern it has seen with previous corporate technology waves: excitement, disillusionment, and a move on to the next initiative before AI gets a fair test.
The Narrow Aperture Problem: Nobody Sees the Whole Picture
Perhaps the most striking idea from the conversation is one Kat says she rarely gets to fully unpack: every player in the biopharma manufacturing ecosystem sees a different sliver of the industry data.
“Pharma and biotech only see the molecules they’ve been developing, and they have the clinical context, which is very important,” Kat says. “But they’re only looking at a fairly small set of molecules that are theirs.” On the technology provider side, she notes, there’s often visibility into granular process details, but no information about the molecule going through the process, so that scientific context is missing. Big tech and AI specialists see model performance and infrastructure, but are missing the scientific context much of the time. CDMOs sit in a unique position: they see many molecules, so their aperture is much bigger and they have deeper context, though they’re still missing the clinical context. “It’s a multidimensional data set, but everybody’s looking at it through their own narrow aperture,” she says, “and in the end, we’re not getting the very big picture.”
Kat ties this straight back to the IP question: there is real value sitting in the sector’s collective data, but unlocking it depends on who ultimately gets the learning rights, and who gets to capture, productize, and commercialize those learnings.
Want to learn more? Kat Kozyrytska will be speaking at CDMO Live Americas (October 20-21, Boston) on what Build-Buy-Partner-Collaborate decisions look like now for AI in manufacturing. Find out more and register.
