AI in pharma and life sciences is currently easy to sell, but far harder to operationalize effectively. Some companies are eager to showcase “pilots purely for presentation,” while others are just as eager to buy them. However, in practice, it frequently becomes clear that these projects lack the essentials: proper access to data, clear rules for data exchange, and a culture in which AI is viewed not as a “replacement for humans,” but as their instrument.
In this interview, we dissect exactly why data sharing and data platforms are becoming the industry’s “bottleneck,” which metrics truly distinguish a working system from a demo, and why, in a live production environment, a pristine F1-score on a slide matters less than a model’s capacity to withstand data/model drift and remain explainable. This is especially important in a field where the cost of an error is measured not only in money, but also in patient outcomes..
Our interviewee is Marcin Wawryszczuk (PhD, MBA), Head of AI at Andersen, Assistant Research Professor in AI, and an expert on artificial intelligence in the Working Group of the Council of Entrepreneurs under the Ombudsman for SMEs in Poland. He operates at the intersection of fundamental science and the implementation of corporate projects, and without unnecessary mystique, explains where AI genuinely delivers results and where its impact is merely “fabricated” out of fear, fashion, and poorly chosen use cases.
2Digital: Which aspects of AI application in pharma/life sciences are delivering tangible results right now, and where do companies most frequently err when attempting to “fabricate” this effect?
Marcin: First and foremost, AI leverages shared data at scale or creates incentives for data sharing.a, which until now was impossible. Unfortunately, the culture of data exchange has not yet fully matured. And here lies the challenge: organizations that operate in or work within the life sciences and healthcare domains tell us about sensitive data, and in reality, we still lack well-structured mechanisms for their shared use.
Initial initiatives are emerging that attempt to establish proper data sharing at both the national and European levels. For instance, in Poland there is such an initiative, but it remains in its “infancy,” so data availability for scientific purposes is still very limited.
In other words, we theoretically have the data. It is localized, it is not shared, it is not utilized in all its “splendor.” I would say this is the first element that must be developed.
The second element is building human-machine collaboration. I frequently observe today either the fear that AI will perform better — “oh God, AI will become so intelligent that it will take my job” — or denial: “this will never work as well as a physician.” And all of this instead of recognizing the opportunity for cooperation and, essentially, building an assistant.
That is, if we are working in the life sciences field, it is essential to cultivate a working culture based on the principle of AI assistants: I remain the human who is still responsible for the final decision, and I have my AI assistant that performs a range of actions for me, for which I either lack time, or they simply fall slightly outside my competencies, yet I need someone competent to do them. Implementing such a culture is delivering results right now.
2Digital: If we asked you to name a few KPIs that best distinguish a “pilot for presentation” from a truly functional prototype, what would they be?
Marcin: If we are talking about models or building a demo system, we essentially rely on statistics, on KPIs — for example, the F1 score. It indicates a certain balance between the model’s accuracy and sensitivity: how many “good” cases it identifies, how many “bad” ones, but also how many “good” cases it classifies as “bad,” or how many “bad” cases it fails to detect entirely. This is the foundation if we are speaking about model training.
If we speak about solutions at the enterprise level, enterprise grade, everything becomes much more complex.
There are very many metrics. But what we subsequently begin to measure is of very great significance. If it turns out that the F1 score on training and validation data is correct, or better to say F1 is actually the harmonic mean of precision and recall, then we enter the realm of production launch and potential data drift — that is, the variability of data. And here — it depends on the system we are building and what it is intended for.
Depending on the data types, we apply different techniques for detecting this drift, but in my opinion, data drift is the key characteristic. Because the question is not whether this variability will occur (it definitely will). The question is when it will appear, and how quickly we will detect it so that we can react.
We fight against drift in order to consistently obtain consistent (in the sense of statistical accuracy and sensitivity) results over time. It is important to understand why the data changed, and then either retrain the model or perform some operation on the input data or the model’s output data to mitigate this drift. This depends on the specific case.
The second element is so-called explainability. And if we are talking about life science and healthcare, this is super-important, because if such a model, let’s say, makes a diagnosis, then this diagnosis must be verifiable in some way.
And here we enter the areas of so-called XAI — explainable AI, and the model’s capabilities to explain a decision in both textual and graphical form.
Of course, the metrics we use are far more numerous, but right now we are focusing on the most important ones so as not to create communication noise.
2Digital: What is currently the most significant brake on digital transformation — regulation in itself, how laws are interpreted internally within a company, cyber risks, data quality, or simply “people and processes”?
Marcin: I haven’t conducted specific studies to speak with 100% certainty, but I am willing to speculate empirically, drawing upon my observations.
In my view, a major impediment is the fear of the consequences of non-compliance, of failing to fulfill regulations, particularly when these regulations are often “murky” and unclear. And this concerns not only medicine or life sciences. Many organizations would like to implement AI and begin utilizing it, but they lack confidence in the regulatory environment. I am referring both to the interpretability of current regulations and to future regulation.
Therefore, they say: “hold your horses” — let’s pause for a moment, because we do not know what lies ahead. What if we do something that turns out to be a mistake?
Another point is the absolutely human factor: the fear that we might allow ourselves to be replaced by an AI system and lose our jobs.
I also previously mentioned limited access to data. We lack data concentration systems where information is available to every organization interested in conducting an experiment or training a model.
Finally, from a slightly different dimension, there is access to GPU infrastructure — that is, to video cards, which are extremely expensive. Now RAM is also beginning to emerge as an issue, so the costs for the infrastructure required to ensure correct GenAI operation are also rising.
2Digital: What is currently more critical for pharma: building unified “data platforms” and investing in data governance/MDM, or pivoting to more “local” solutions tailored to specific use cases?
Marcin: In my opinion, it is most critical to build data platform centers. This allows data to be jointly utilized by people, commercial entities, and state organizations that wish to work on this data to create something new. Globalization is intrinsically linked to the simplification of access.
2Digital: How does the risk landscape change when pharma begins actively utilizing clouds, external AI models, contractors, and platforms? What “gaps” do you most frequently see in real projects?
Marcin: Fundamentally, I do not perceive any specific risk in using the cloud. Often, this is fallacious reasoning along the lines of “my data in the cloud is less secure.” With all due respect: if a cloud provider of the caliber of Google, AWS, Oracle, or Microsoft is incapable of securing better security specialists than my own company, then who is capable?
Objectively, they possess some of the best security specialists in the world. So, however you look at it, the data is protected — and by mechanisms that have been refined over years. Therefore, storing data in the cloud is not about inferior security in terms of the possibility of unauthorized, uncontrolled access, however, cloud security failures are usually customer-side.
However, what does provoke certain risks and concerns is the physical location of this data, and this is tied to regulation (without even discussing whether these regulations make sense or not). I am referring, for example, to rules whereby if data is produced in my country, it must remain in my country or within the territory of the European Union, and cannot end up in the USA—in order to guarantee data processing in accordance with the law. In a sense, this is a challenge. And perhaps not so much a technological one, as a challenge of controlling data placement and creating a system that ensures its correct localization.
Potentially, price could prove to be a problem. We access models via the cloud or via API, but these are external, proprietary models; we simply pay for them “on a consumption basis.” If our systems are improperly safeguarded regarding limits related to data processing and budget consumption, we might one day wake up with a massive debit on the account. Therefore, one must monitor the volume of processed data and manage budgets correctly.
Another question is what happens to the information after the model has worked with my data via the cloud. Most often, there is an agreement that the data is logically separated, not commingled with other clients’ data, and not subsequently used to train these models. I do not wish to veer into the realm of conspiracy theories, but de facto, we do not know what actually happens to this data. If it “flies off” to a language model in inference mode—that is, figuratively speaking, the model is already trained: I ask a question, I get an answer — this data is leaving a trace and being recoverable enters speculative territory, even if technical recoverability is not possible..
At the same time, theoretically, it leaves a trace, meaning it can be recovered. And the question remains: in what form is it used, even though it shouldn’t be? And does the possibility exist to catch someone red-handed, to actually prove that this data is being used for training? Therefore, I would say there is a certain “grey zone” here where there is no total certainty. Despite the fact that contractually, we may have guaranteed logical data separation and a prohibition on its use for subsequent training.
However, if we keep data solely on our own premises, in our own data center, and work on our own language model, we may de facto significantly mitigate the risk that the data goes outside of our controlled environment in an uncontrolled way.. This is an element that may incline organizations — especially regarding sensitive data — to keep a portion of their information “in-house.”
2Digital: Which roles in pharma teams are becoming critical due to AI (Head of AI vs Head of Data, ML Ops/Model Risk, QA/CSV, Product Owner)? Whom do companies still not know how to hire or cultivate?
Marcin: I see a shortage across all AI roles. There is a massive gap in the market. The biggest problem is that there are many so-called “experts” who have read something, perhaps even done one project at home, and they want to view themselves as specialists, but they are not.
And this is a problem at the organizational level. Identifying the right talent very often hits a wall due to a lack of internal capabilities to verify such a person. That is, if a company lacks competencies—lacks AI competencies—and wants to hire someone from the market, it is unable to properly vet them. And this applies to practically all roles—it cannot thoroughly check the person.
The second point: speaking of the roles themselves, we often go down the path of “let’s just go and do it,” without thinking beforehand about what comes before “doing it.” There must be a clear strategy where we define matters related to AI; select candidates — process-based or product-based — on which we want to execute the first AI implementations, and only then begin to act.
Why? Because I have seen this with my own eyes many times: when action begins without a plan—”we have an idea, let’s do it.” Often, instead of becoming an internal evangelist and pivoting the organization toward AI, this idea turns into an internal enemy of AI: because it doesn’t work, it isn’t used, there is no attention paid to it. Someone came up with a project without alignment and without verifying how it fits into the company strategy.
This is how disappointment sets in: “AI doesn’t work.” No, AI works — it didn’t work in this specific instance because the case was poorly chosen: it wasn’t strategic, and it was executed by the wrong people. There was no proper adoption at the organizational level.
This is the challenge. But to address the question of “who is missing from the market” directly: generally speaking, everyone is missing. But what is especially lacking — and these are key roles to start moving — is strategy and architecture related to AI, to artificial intelligence: there are quite large voids here. And the third key role, in my opinion, is the analog to DevOps. That is, MLOps or LLMOps: the person responsible for the deployment and subsequent maintenance of the system.
2Digital: You are one of the key figures at the Life Sciences and Pharma Networking Event. What is the single most “practical” takeaway you want a participant to carry away from the event? One effect — as your ideal goal.
Marcin: I think it would be ideal to reduce the fear and uncertainty associated with the implementation of AI systems, and to cultivate a perception of understanding, of knowledge: where I can use it, why I need not be afraid, and why I need to move in this direction. That would be ideal for me.
2Digital: From our conversation, it is evident that people’s fear of AI is substantial—so it would indeed be good to alleviate it. The final question—you have chosen an informal format for the meeting at the Pharma Networking Event. Why is it sometimes worth it in your sphere to arrange for coffee, a beer, networking—to talk without a microphone? And why is there a rather small group at this specific event?
Marcin: Of course, it is always worth checking — even if we wish to attend webinars and trainings — who is conducting them, what their experience is, and what they have actually accomplished in life regarding AI. Not infrequently, people claiming the role of expert lack knowledge. Why is it worth going “for a beer” (in quotation marks) and talking informally? I think, if only to receive an honest answer: where one should go, whether one should go at all, whether my idea makes sense—to discuss this with a company or a person who has, as they say, “cut their teeth” on the market, has seen a lot, and knows a lot.
At Andersen, we execute many such projects; we have many genuine specialists, experienced in real projects” I myself, one might say, have been connected with AI and GenAI since time immemorial — both scientifically and professionally: because besides working at Andersen, I also work as an adjunct professor at the Academy of Silesia [Akademia Śląska], where we are engaged in research on AI and GenAI in medicine. It is a medical university.
And furthermore, Andersen offers a service called AI Strategic Evaluation. In this strategic evaluation, we state very honestly where one should go and where one should begin. My team and I adhere to the maxim “No AI first.” That is, if we can build a system without AI, we do so. If we see that there is no application for AI or GenAI in the proposed project, we say it straight: there is no point.
