Asking whether AI can be trusted in pharma is really asking a question about architecture. What the system retrieves from, whether it learns from what it’s given, and whether any given sentence can be traced back to a source — those answers determine the trust, not assurances about the model. 

In this episode of Pharma Talks, Nataliya Andreychuk spoke with Anna Shavurska, an AI solutions architect with over a decade in technology and 14 years at Viseven, working across front-end, back-end, and AWS environments. 

The Data Question 

Anna’s starting position is that client caution here is well-founded rather than something to be talked past. The concern is about how AI systems use and process proprietary data, and the answer has to be structural. 

The approach she describes is built on retrieval-augmented generation. Rather than drawing on the general data a model was trained on, the system retrieves exclusively from approved, client-provided material held in secure knowledge bases. 

The second half matters as much as the first: these systems don’t learn from client data. It stays isolated and isn’t used to train the underlying models. Nataliya’s question was whether that amounts to a short memory. Anna confirmed it does — which is the tradeoff being made deliberately. 

Traceability Over Accuracy Claims 

Hallucination is the other standing concern. Anna’s answer isn’t that the risk disappears; it’s that high-quality validated data reduces it substantially, and that traceability handles what remains. 

Every output carries references back to source — what she describes as complete linkage of all documents used to generate each sentence. In an environment where claims have to be verifiable, that’s the property that matters. A system that’s usually right without showing its work is less useful than one whose reasoning can be inspected. 

Structure Instead of Prompting Skill 

Nataliya raised the practical version of the problem: using general-purpose tools produces inconsistency, and the manual correction eats the time saved. 

Anna’s answer is that consistency comes from constraining the workflow rather than improving the prompt. Predefined prompt cards and guided actions walk users through content creation instead of starting from a blank field. Master templates — HTML structures with approved design elements and content blocks — hold both visual consistency and regulatory compliance in place. 

Prompting itself gets addressed the same way. Crafting a good prompt is a real skill, and best practice increasingly means long, complex instructions that most users won’t write. Prompt libraries configured per client remove that requirement: pre-built prompts aligned to specific workflows, whether generating an email, adapting existing material, or building a campaign. Users select from the library or follow the system’s suggested next step. 

There’s a reuse argument underneath this too. Pharma companies already hold enormous quantities of produced content, and Anna’s emphasis is on giving those assets another life by adapting them across channels and formats rather than generating from zero. 

Why Who Builds It Matters 

Nataliya raised the underrepresentation of women in AI development. Anna’s response moved past representation as a fairness question into what it does to system behavior. 

AI is not just an algorithm. It is a system designed by humans and trained on human data. 

Her examples are documented ones — systems trained on historical hiring data reproducing a preference for male candidates. The data reflected what actually happened, which is precisely the problem. Without diverse input at the design stage, historical inequality gets encoded as forward-looking logic. 

Her conclusion is direct: if only men build these systems and only male-centered data feeds them, the output will reflect that. 

A Note on Getting Into the Field 

Anna’s own path started at university, with encouragement from her father — his prediction being that if she finished her studies she’d be an excellent engineer within a few years. She found the work absorbing once she started. “It’s its own world.” 

What Trust Actually Rests On 

The conversation lands on a straightforward point: trust in AI systems doesn’t come from claims about them. It comes from architecture that isolates data, outputs that can be traced to sources, workflows structured so users don’t have to be experts to get compliant results, and development teams diverse enough to catch what a narrower group wouldn’t. 

Those are engineering decisions and hiring decisions, which is what makes them checkable.