Reproducible AI as a property of the platform

Q: What makes AI reproducible in a regulated environment?

AI becomes reproducible when data, environment, and logic are bound together into a single immutable package at the moment of execution rather than afterward. That package, the scientific unit of work, lets any result be rerun and audited later without relying on someone remembering to record what they did. Reproducibility becomes a property of the architecture, not a discipline scientists maintain by hand every time a model runs.

Q: Why do scientists in life sciences work around compliance requirements?

Scientists in life sciences circumvent compliance requirements when the compliant option is more painful than the alternative. Well-trained scientists in well-designed environments want to do the right thing, but they will find a route around a barrier the way they solve any research problem. The fix is lowering friction on the compliant path rather than raising barriers on the noncompliant one, so the easiest path and the compliant path are the same.

Q: What does compliance for AI agents look like in a regulated environment?

Compliance for AI agents means applying the same framework used for human-initiated work, not building a separate track after the fact. AstraZeneca's AI gateway, model hub, and Domino runs that power agentic workflows operate under identical governance rules as scientist-driven projects. Self-service infrastructure originally built for human practitioners now treats agents as first-class consumers, so governance architectures designed only with human actors in mind need structural rework as agentic systems take on more of the work.

Reproducible AI means knowing exactly what data, environment, and logic produced a result, not just repeating the same output twice. Removing friction makes science faster, more trustable, and more compliant all at once, not a trade-off between them.

Justin Lecher

Sr. Director, Data & Platform Engineering

AstraZeneca

What you'll take away from this session

A unifying framework beats point fixes

Defining a single reusable unit of work, data, environment, and logic bound together beats a decade of incremental fixes that only added complexity. Point solutions solved immediate complaints without addressing why complexity kept compounding.

Binding data, environment, and logic builds trust

Binding data, environment, and logic into a form that cannot change afterward makes a result auditable and reproducible on its own. Documentation alone depends on someone remembering to record what happened.

Fast deployment reflects platform design

AstraZeneca's patient-data agent reached production in weeks because the environment was already qualified, approved for that data classification, and wired into audit infrastructure. Instead of being the constraint, the model was the last step.

Model leaderboards double as governance evidence

A foundation model benchmarking platform built on real clinical tasks gives a ranked, auditable answer to which model fits a given use case. That ranking documents the basis for model selection decisions during regulatory review.

Self-service cost governance needs new mechanisms

Self-service opex fixes the visibility problem centralized IT had, but creates a new one. Closing projects and releasing resources is harder to build into the platform than spinning them up.

Self-service turns a barrier into a path

A generative app built and deployed in a single afternoon, with no IT ticket and project overhead, runs on three things working together: a coding agent cyber-approved to operate inside AstraZeneca’s own boundary, an AI Gateway offering risk-based, self-attest access to approved models, and Domino’s self-serve compute and publishing.

AstraZeneca's data and platform engineering team built its AI infrastructure around the idea of a scientific unit of work. Binding data, environment, and logic into one immutable package turns a result into something anyone can rerun, audit, and trust, without a separate documentation step layered on afterward. Chaining these units together extends that property from a single experiment to a full clinical trial workflow, carrying the provenance regulatory submissions require.

The same logic explains why a patient-data agent handling clinical trial data reached production in weeks. The development environment was already qualified, approved for that data classification, and wired into audit infrastructure before the model was chosen. Across life sciences, governance as infrastructure rather than friction is what separates organizations that move fast from those that stall, and it's the same principle that lets agentic AI systems operate safely at scale. Extending that same self-service model to AI agents as well as human scientists is the next structural shift that’s underway.

FAQ

What makes AI reproducible in a regulated environment?

Why do scientists in life sciences work around compliance requirements?

What does compliance for AI agents look like in a regulated environment?

Transform the work that matters most

See how Domino helps the world’s most regulated enterprises build, scale, and govern AI-powered applications.

Watch demo Talk to an expert