Engineering Trust in a Life Sciences Culture

The Challenge

This was an organization built by engineers — mechanical and electrical specialists who controlled every variable in their experiments, trusted what they could measure, and viewed anything “heuristic” with suspicion.

But the business had evolved. As both a manufacturer and an outsourced R&D partner for pharma and chemical clients, they now managed layers of complexity that didn’t fit neatly into spreadsheets:

• Dual confidentiality: Their own IP as machine manufacturers, plus client-specific recipes and processes for outsourced production
• Compartmentalized access: Strict separation between internal knowledge and client-specific secrets
• Regulatory weight: EU GMP compliance, FDA requirements, and a constant stream of regulatory changes across multiple jurisdictions
• Knowledge concentration: Decades of institutional memory — especially around regulatory change processes and regulator expectations — lived in the heads of a few long-tenured specialists

The result: regulatory change assessments took three weeks. Manual overhead consumed expert time. And the engineers who could navigate cross-country regulatory nuances were approaching retirement with no structured way to transfer what they knew.

The Knowledge Landscape

All three dimensions were in play:

• Structure: Structured experiment results, semi-structured experiment setups, unstructured policy documents and regulatory texts
• Provenance: Internal IP, client-confidential processes (requiring compartmentalized access), and external regulatory frameworks
• Knowledge Types: Deterministic formulations and results, stochastic process optimization, and — critically — experiential knowledge about how regulators actually behave, how long reviews take, and what triggers scrutiny in different jurisdictions

The Approach

We started where openness existed: the Strategy team and a small group of champions willing to experiment.

The focus:

1. Identify where missing knowledge transfer and manual administration hurt most
2. Establish knowledge extraction and automation with robust Human-in-the-Loop controls
3. Gradually increase agent autonomy as trust was earned

The autonomy ramp was deliberate:

• Start: ~90% Human-in-the-Loop review (only trivial checks excluded)
• Progression: Each guardrail removed from full HitL was replaced with routine spot checks — catching hallucinations and model drift without bottlenecking every decision
• Current state: ~30% HitL, continuing to evolve as knowledge bases mature

The Impact

Time to value: 5-6 months to measurable, consistent impact — deliberately paced to get the ontology and guardrails right.

The Human Story

The “not invented here” syndrome was real. Senior engineers didn’t oppose the project — they simply didn’t believe it would work. Their stance: “Let’s see where this takes us in 6 or 12 months.”

The Strategy and Regulatory teams couldn’t convince them with arguments. They convinced them with results. Faster turnaround times meant engineers got what they needed sooner. The value became undeniable, even if the method remained unfamiliar.

Some long-timers came around. Others maintained skeptical respect. Convincing a culture built on control and precision is a long-term task — and that’s fine. The system earns trust the same way the engineers do: by being right, consistently, over time.