Why AI Is Too Generic for Regulatory Conversations in Life Sciences

Regulators Focus on Intended Use and Impact

Regulatory bodies are generally less concerned with the fact that a system uses AI and far more focused on what the system is intended to do, the impact it may have, and the controls surrounding it.

The same AI model can be considered low risk or high risk depending entirely on the context in which it is deployed.

For example:

• An algorithm used for administrative document sorting carries a very different risk profile from one supporting clinical decision-making
• A model generating internal operational insights does not require the same governance as a model influencing product quality or patient safety
• An AI assistant supporting review workflows is fundamentally different from an autonomous system making unsupervised GxP decisions

Regulators therefore evaluate more than the technology itself. They assess intended use, potential consequences of failure, oversight mechanisms, traceability, validation strategy, and risk controls.

This distinction is often where organizational fear originates.

When a quality leader hears: "We want to deploy AI in manufacturing," the assumption may immediately shift toward unpredictable self-learning systems making autonomous decisions. In practice, the proposed solution may simply be a static deterministic model supporting batch review, a completely different governance and validation scenario.

Without this level of specificity, teams end up discussing entirely different systems while using the same terminology.

AI Risk Evolves Across the Pharma Value Chain

AI risk is not static. It evolves as technology moves through the pharma value chain.

In preclinical research, AI is primarily used to accelerate discovery and generate scientific insights. At this stage, the main concern is typically scientific validity and reproducibility rather than direct patient impact.

Within clinical trials, AI may influence patient selection, protocol optimization, or trial design. Here, considerations around bias, transparency, patient protection, and Good Clinical Practice (GCP) become significantly more important.

In manufacturing and quality environments, the stakes increase further. When AI systems can influence product quality, release decisions, or patient safety, Good Manufacturing Practice (GMP) expectations demand predictable behaviour, validated performance, and continuous oversight.

The same underlying AI capability may therefore require completely different governance approaches depending on where it is deployed.

Understanding this progression allows organizations to begin with lower-risk applications and gradually expand their AI maturity as internal trust, governance capability, and operational experience grow.

Static Deterministic Models Are the Practical Starting Point

A common misconception is that only highly autonomous or continuously self-learning systems create value.

In reality, static deterministic AI models already deliver substantial benefits in regulated environments.

These models operate with fixed parameters and locked behaviour, making them significantly easier to validate, monitor, and control within GMP-regulated processes. Deterministic behaviour also ensures outputs remain predictable and reproducible, which is essential when product quality and patient safety are involved.

As Evelien Cools from delaware highlighted during our joint webinar, current regulatory expectations in GxP environments strongly favour static deterministic models. Annex 22 guidance is particularly clear on this point.

Human-in-the-loop oversight also remains a critical principle, not only from a regulatory perspective but also from an organizational one. Maintaining human review and accountability supports change management, user adoption, and confidence in the system's outputs.

This mirrors what is already well established within the medical device sector, where static models remain the dominant approach for many high-risk applications.

Rather than waiting for fully autonomous AI systems, organizations can already achieve meaningful gains in efficiency, accuracy, and operational consistency using controlled deterministic models today.

Building AI Maturity Step by Step

Successful AI adoption in life sciences is rarely about making one massive leap toward autonomy.

The organizations seeing the most progress are typically the ones approaching AI maturity incrementally:

• Starting with lower-risk use cases
• Embedding governance early
• Defining intended use clearly
• Establishing validation and monitoring frameworks upfront
• Expanding capabilities gradually as trust and experience increase

Treating AI as a spectrum of technologies rather than a single category enables far more productive conversations between quality, regulatory, digital, and operational stakeholders.

And ultimately, that clarity is what allows AI initiatives to move beyond proof-of-concept and into compliant production environments.

Looking to Assess Your AI Use Cases?

Jonathan Boel, Evelien Cools, and Pieter Smits recently explored this topic during a webinar on compliant AI adoption in life sciences.

Watch the on-demand session to learn how intended use, risk classification, governance, and validation frameworks influence AI readiness in regulated environments.

Looking to assess your own AI initiatives against the right regulatory framework? Get in touch with the QbD Group team to discuss your roadmap toward compliant AI adoption.

Watch On-Demand: AI in Life Sciences

Explore AI maturity models, governance, validation frameworks, and human-in-the-loop principles with Pieter Smits and Evelien Cools (delaware).

Watch the on-demand webinar