AI exists in life sciences, but broad manufacturing scale remains rare

CRB's Horizons: Life Sciences 2026 surveyed more than 400 global R&D and manufacturing leaders.

Detailed results reported by BioSpace show 1% using AI across most or all regulated-manufacturing applications and 8% across many applications, for a combined 9%.

Another 17% said AI was scaled in a few applications. The signal is therefore not 'no AI in pharma manufacturing' but 'broad production scale is still uncommon.'

The industry is not rejecting the technology itself

BioSpace reported AI use in administrative workflows at 56% and in drug discovery at 45%.

CRB's own release says 68% expect AI, robotics and automation to reduce manufacturing cost of goods.

Interest exists. The harder question is how to move from technical capability into regulated operations.

AI Capability ≠ Validated Operation

Standard AI evaluation emphasizes accuracy, speed and cost reduction.

Regulated manufacturing adds different questions: what is the context of use, what happens if a recommendation is wrong, who reviews it, and what must be revalidated when the model changes?

A model that works technically is not the same thing as an operating system that can be repeatedly used inside a GxP environment.

Performance alone does not close the gap

Reporting on the CRB survey highlighted concerns around accuracy and quality, but also insufficient validation standards, unclear accountability, and weak transparency or explainability.

Even a stronger model cannot solve unclear responsibility, change control or traceability by itself.

In regulated manufacturing, the surrounding operating architecture can become the bottleneck.

FDA and EMA emphasize context, risk and lifecycle management

In 2026, FDA and EMA set out ten Good AI Practice principles for drug development.

They include human-centric design, a risk-based approach, clear context of use, multidisciplinary expertise, data governance and documentation, risk-based performance assessment and lifecycle management.

The direction is clear: validation, oversight and documentation should be proportional to context and risk, and AI needs management across its lifecycle.

Between the model and production scale sits a validation architecture

Banseog structures the operating path as AI Model → Context of Use → Risk Assessment → Validation → Human Oversight → Documentation → Change Control → Monitoring → Production Scale.

This is not an official CRB or FDA process name. It is an analytical sequence built from the public regulatory and industry principles.

The gates are not just regulatory friction; they are part of how production trust is created.

Human-in-the-loop can be a scale architecture, not a temporary weakness

Human-in-the-loop is often treated as a temporary patch until models improve.

CRB argues that human-in-the-loop approaches may be important for expanding AI in highly regulated settings.

A structure in which people review recommendations, approve actions and leave an auditable record can make a broader deployment more acceptable.

Validated semi-autonomy may scale before full autonomy

Consider AI that detects anomalies, proposes causes and recommends an intervention while an experienced operator reviews and approves the action and the full path is logged.

That is less autonomous than an AI system that changes a process on its own.

Yet the less autonomous system may scale first in a regulated environment because responsibility and evidence are clearer.

The product competition shifts beyond model performance

A vendor saying 'our model is 99% accurate' may not be enough.

A manufacturer may also need to know the validated context of use, the evidence package, audit trail, approval rights and the impact of future model updates on prior validation.

The unit of competition can therefore expand from Model Performance toward a Validated Operating System.

The scarce talent may not be an AI engineer alone

Regulated-manufacturing AI sits at the intersection of AI and data, manufacturing process, GMP and QA, validation and regulatory knowledge.

FDA and EMA explicitly include multidisciplinary expertise among their Good AI Practice principles.

A particularly scarce profile may be someone who can define where a model is allowed to act inside GMP manufacturing and how responsibility and evidence are maintained.

The 43% with no formal plan points to an operational trust gap

CRB says 43% of respondents had no formal plan to implement AI in regulated manufacturing.

That does not have to mean the industry is simply behind. In regulated production, moving fast without responsibility and validation can carry unusually high costs.

Banseog calls the gap between technical promise and deployable trust the Operational Trust Gap.

BANSEOG VIEW | AI Capability ≠ Validated Operation

Only 9% of respondents had AI broadly scaled across many or most regulated-manufacturing applications.

The central question is not just whether AI can perform the task, but how the decision is validated, supervised, documented and managed as the model changes.

In regulated industries, the next competition may be less about who has the strongest model and more about who can build Validated Operation fastest.

BANSEOG VIEW

Banseog View — AI Capability ≠ Validated Operation

Broad regulated-manufacturing AI scale was 9%, distinct from the additional 17% scaled only in a few applications.

Human-in-the-loop can function as a scale architecture for validation, oversight and auditability rather than merely a temporary limitation.

Scarce talent may emerge at the intersection of AI × Manufacturing × GMP/QA × Validation × Regulatory.

SOURCES

Primary sources and references

Broad scale is defined here as 1% most/all plus 8% many applications, and is kept distinct from the 17% reporting scale in a few applications. AI Capability ≠ Validated Operation, Validation Architecture, Operational Trust Gap and Validated Operating System are Banseog analytical frames, not official CRB or FDA terminology.