“The level of comfort with AI use in this space is pretty low right now.” Mike Boyson’s observation cuts to the heart of pharma’s AI dilemma. Hesitation is justified—but it shouldn’t paralyze action. His answer: build practical, compliant digital twins that operators can create in Claude without waiting for IT.
Mike is a supply chain and pharmaceutical operations leader with over 20 years of experience spanning CMC strategy, external manufacturing, and digital transformation. He spent more than a decade at Takeda leading oncology and biologics supply chain operations, served as Executive Director of Value Chain Leadership at Moderna, and now advises organizations at Azure Biopharma Consulting on adopting AI-driven solutions.
In the latest episode of the PharmaSource podcast, Mike shares concrete use cases for digital twins, explains how to build them without deep technical expertise, and addresses the GXP compliance, data confidentiality, and human-centered challenges that will define pharma AI adoption.
Regulatory Caution vs. Practical Implementation
The EU’s Annex 11 expressly prohibits generative AI and LLMs in certain manufacturing contexts. That regulatory reality creates legitimate hesitation. But Mike argues it’s not a blocker—it’s a design constraint.
“The industry has to think about how it will implement these tools to better our systems, processes, and organizational design,” Mike says. His solution: follow ISPE’s GAMP 5 AI Guide. Keep humans in the loop. Show all the work AI is doing. Build in security controls.
AI and machine learning in pharma aren’t new—exploratory work dates back to 2016. What’s different now is the accessibility. Domain users can build models in Claude without being computer scientists.
From Spreadsheets to Interactive Twins: Two Concrete Examples
Mike didn’t know Python when he started. Now he builds digital twins with Claude and GPT. The key difference: he starts with domain problems, not coding frameworks.
His first project tackled a real supply chain question: What happens when a company switches from traditional vials to auto-injectors globally? Market adoption rates vary. Payer willingness differs. Regulatory pathways diverge. Mike built a twin that lets stakeholders input assumptions about market adoption and instantly see implications for pen assembly, cartridge sourcing, and multi-supplier mix.
“These twins aren’t making GMP or GXP decisions,” Mike clarifies. “They’re making supply chain decisions—modeling what happens when demand shifts, when capacity must expand, or when you should divest assets.”
His second example: EU regulatory compliance requires detailed supply chain maps showing physical, documentation, and financial routing for every market. For a 60-market product, that’s complex. Mike created an interactive mapping tool in an afternoon. Users click on any supply chain node to see routing, inventory position, and data lineage. The system pulls directly from SAP and updates automatically.
“The data pulls from SAP,” Mike explains. “There’s no reason you can’t have a direct connection updating daily, depending on how often your refresh runs—seven hours, five hours, whatever it might be.”
Audit Trails and Human-in-the-Loop Design
The biggest pharma question: How do we maintain compliance if AI is making recommendations?
Mike’s answer is direct: humans stay in the loop. Everything gets logged. If an agent collects data, runs it through a model (say, Monte Carlo analysis), and recommends a scenario, the audit trail shows all three steps. A separate tab shows exactly where that data came from.
“You need a running tab of every single action the system takes to produce output,” Mike says. “If it’s collecting data, putting the data through a model, and selecting a scenario recommendation, then you’ll see the audit trail list all of those things. A person performing that work that day can then validate the recommendation before acting on it.”
On hallucinations—AI confidently asserting things that aren’t true—guardrails are essential. But the principle is familiar: human judgment backed by auditable data.
“If there’s a particular decision point or action that’s going to be taken, the data has to be made available so the human can help make the right decisions,” Mike notes.
Data Confidentiality: Enterprise Safeguards Are Available
Pharma leaders worry about sending sensitive manufacturing data to cloud-based AI systems. That concern is valid. But enterprise-grade protections exist.
Current enterprise instances of Claude, GPT, and Perplexity explicitly state they won’t use your data for model training. Both Claude and Perplexity are HIPAA compliant. “It’s not a switch you have to turn off,” Mike says. “It’s something that comes immediately available.”
Mike acknowledges the trade-off. On one hand, companies can’t advance without adopting these models. On the other, there’s inherent risk. “Each company has to weigh that,” he says. “But privacy safeguards are evolving and getting much better at compartmentalizing data usage.”
His recommendation: start with non-sensitive use cases, build organizational competency, and escalate as confidence grows.
Getting Started Without Technical Expertise
The barrier to entry is lower than most assume. Mike didn’t know Python. He learned by doing.
“I had chat GPT write Python for me, and I’d cut and paste it into my laptop terminal,” he recalls. “I don’t know Python, really at all. But it seemed that chat GPT was comfortable with it, so it was easier for me to do the cutting and pasting.”
For teams building AI-driven solutions:
- Start conversationally. Ask Claude or GPT to help you develop the right prompt or skill.
- Learn the basics. Understand why skills matter, what Markdown is, and whether Python or a no-code tool (Langflow, n8n) fits your use case.
- Embed domain expertise. Mike created Markdown prompts that taught Claude about ISPE’s GAMP guide, so it could verify computer systems assurance compliance.
- Pick the right tool. Claude Code is powerful for rapid prototyping. GPT 5.4 is a close competitor. Perplexity Computer works as an orchestrator for parallel workflows.
The Future: Augmentation, Not Replacement
Mike is optimistic but realistic. “The word vibe coding was coined 12 months ago,” he notes. “It’s almost impossible to predict where this is going.”
His thesis: massive workforce displacement won’t happen. Some job functions will shift. But growth will follow. “Looking back at history, it’s very difficult to see that we wouldn’t be in a better place over the next couple of years.”
Real risks exist: prompt injection attacks, changing workforce skill requirements, confidentiality concerns, and unresolved GXP/regulatory gaps. Geopolitical tensions around AI add another layer of uncertainty.
But uncertainty shouldn’t paralyze action. Winning companies will be those that experiment thoughtfully, build compliance into their DNA, and see AI as augmentation, not replacement.
Mike Boyson will be speaking about AI in Pharma Manufacturing at CDMO Live Americas (Oct 19-21, Boston)
