AI in CMC strategy: why senior scientific review still matters
- 4 days ago
- 7 min read
In April 2026, the FDA issued a warning letter to a manufacturer after investigators found the company had relied on AI-generated content within manufacturing and quality activities without appropriate oversight. During the inspection, the company stated that a required process validation activity had not been performed because “the AI never told us it was required”.¹
The FDA addressed the issue directly in a section titled Inappropriate Use of Artificial
Intelligence in Pharmaceutical Manufacturing. The agency stated that where AI is used to
support document creation, firms remain responsible for reviewing the resulting content to ensure it is scientifically accurate and compliant with cGMP requirements under 21 CFR 211.22(c).¹
That warning reflects a broader shift already taking place across the industry. AI tools are
now embedded in many parts of pharmaceutical development, including technical writing, regulatory summarisation, data review, protocol drafting and early-stage CMC planning.
Their usefulness is obvious. They can process large volumes of information quickly, structure documents efficiently and support teams working under increasingly compressed development timelines.
The challenge is that technically coherent output can still contain critical omissions, weak
assumptions or phase-inappropriate decisions that only become visible under deeper
scientific review.
In CMC development, that creates a specific risk.
Why AI struggles with CMC decision-making
Most CMC problems are not caused by a complete absence of data. More often, the
challenge lies in interpreting incomplete or conflicting information while understanding the regulatory and manufacturing consequences attached to it.
That judgement develops through direct experience with scale-up, process transfer,
validation strategy, analytical variability, regulatory interactions and manufacturing
deviations. Two development programmes may present similar datasets while requiring
completely different decisions because the operational context differs.
AI systems can identify patterns across published information and historical datasets. They do not interpret process behaviour in the context of manufacturing history, scale-up
constraints or prior deviation patterns. A technically acceptable decision during early
development may still create downstream problems during PPQ, comparability assessment or regulatory review.
This becomes particularly important during early development, where many CMC decisions are made before process understanding is mature.
The process validation example in the FDA warning letter illustrates the issue clearly. The
missing activity was not hidden behind fabricated data or an obvious technical error. It was absent because the AI system never identified it as required. The resulting documentation may still have appeared complete to someone reviewing it superficially.
The modality landscape is increasing CMC complexity
The CMC decisions described above were already difficult in programmes built around
conventional monoclonal antibodies and recombinant proteins. Newer therapeutic modalities introduce additional layers of analytical, manufacturing and regulatory complexity.
Recent examples illustrate the direction of travel. In January 2026, the FDA granted Fast
Track Designation to IBI3003, a trispecific T-cell engager targeting GPRC5D, BCMA and
CD3 in relapsed or refractory multiple myeloma. Regeneron’s linvoseltamab, a BCMA/CD3 bispecific built on a tandem scFv architecture, received FDA approval in the same indication during 2025. In vivo CAR-T programmes such as Capstan Therapeutics’ CPTX2309 entered first-in-human studies using lipid nanoparticle-delivered mRNA constructs rather than ex vivo engineered cells. Bispecific antibody–drug conjugates are also expanding rapidly across clinical pipelines.
Each modality introduces CMC problems with limited historical precedent.
A trispecific construct must maintain consistent binding behaviour across three target
interfaces while controlling aggregation, charge heterogeneity and chain mispairing that
become more difficult as structural complexity increases. Bispecific ADCs combine the
challenges of dual-target architecture with payload-related quality attributes such as drug-to-antibody ratio distribution, linker stability and site-specific conjugation control. In vivo CAR-T platforms introduce an entirely different framework where the drug product is an LNP-mRNA system and biological activity occurs after intracellular reprogramming.
These programmes generate larger volumes of data, more complex comparability questions and a broader range of potential failure modes. That increases the practical value of AI within CMC workflows because literature synthesis, technical drafting and cross-referencing activities expand faster than available expert time.
At the same time, the need for senior scientific review becomes greater rather than smaller. Many of these modalities involve development decisions that cannot yet rely on established regulatory or manufacturing precedent. Determining whether a comparability strategy is phase-appropriate, whether an analytical gap is acceptable or whether a process risk will become visible only during PPQ still depends on experience with how programmes behave under scale-up and regulatory scrutiny.
AI can assist with information handling and documentation at a speed no individual reviewer can match. But the judgement required to distinguish manageable uncertainty from programme-level risk remains a scientific responsibility rather than a computational one.
AI limitations in batch review and analytical assessment
At 3Biotech, we recently reviewed a monoclonal antibody clinical batch released by a
CDMO, with analytical confirmation performed by an external laboratory.
The batch met specification on paper. The reported drug substance concentration sat
comfortably within the release limits. The certificates of analysis were complete, the
documentation trail was intact and no obvious deviation had been identified during batch
review.
An AI-based review system would almost certainly have accepted the dataset without escalation. The numerical values were internally consistent and aligned with the expected formulation target.
The senior reviewer identified a concern immediately. The reported concentration appeared unusually precise given the sequence of upstream processing steps and the known variability associated with that process configuration. That observation triggered a more detailed investigation into sample preparation and process history.
The root cause was residual Water For Injection (WFI) carried over from an upstream
operation that had altered the analytical sample matrix. The concentration result had been systematically overestimated. As a consequence, the calculated patient dose would have been lower than intended.
No single value in the dataset appeared implausible in isolation. The issue emerged from
recognising that the overall result looked cleaner than it should have based on practical
manufacturing experience.
This type of judgement is difficult to reduce to pattern recognition alone because it depends on understanding how process artefacts appear in real manufacturing environments and how apparently acceptable analytical data can mask underlying process issues.
AI-generated CMC assessments can miss critical development risks
A second example emerged during an audit assessment performed for a biological scale-up project supported by a French public innovation fund.
The development package was extensive. It included process descriptions, analytical
summaries, development reports and a multi-year scientific rationale supporting the
programme. Structurally, the dossier appeared comprehensive.
An AI-assisted assessment would likely have produced a balanced technical summary
identifying strengths, development gaps and recommended next steps. Many of the
individual issues were detectable at document level.
The more important task involved assessing which issues materially threatened programme progression.
During senior review, more than fifteen unresolved questions were classified as blocking points requiring resolution before further investment. These included concerns related to
genetic stability of the production system, unsupported productivity assumptions,
downstream process mass-balance inconsistencies and culture medium choices carrying unresolved regulatory implications.
The central challenge was not identifying isolated technical observations. It was determining their severity in the context of development stage, manufacturing strategy and regulatory expectations.
Triage remains one of the core responsibilities within CMC leadership. Development
programmes rarely progress with perfect datasets. Decisions depend on understanding
which uncertainties are acceptable within the current phase and which create unacceptable technical or regulatory exposure if left unresolved.
Regulatory expectations for AI in GMP and CMC activities
The FDA warning letter is important because it places AI oversight directly inside existing
GMP expectations rather than treating it as a separate future framework. Under 21 CFR
211.22(c), quality units remain responsible for reviewing and approving procedures and
records affecting product quality regardless of how those documents were generated.
Regulatory agencies including the FDA and EMA have both published principles addressing AI use in drug development and manufacturing environments. 2 Across these publications, several themes appear consistently:
accountable human oversight
defined governance structures
documented review processes
data integrity controls
traceability of AI-generated content
clarity regarding intended use
The EMA’s 2024 reflection paper on AI in the medicinal product lifecycle also addresses the risks associated with generative AI in regulated environments. 3 The paper states that
language models are “prone to include plausible but erroneous or incomplete output” and recommends that AI-generated medicinal product information should remain under close human supervision.
The MHRA’s existing expectations around GxP data integrity become increasingly relevant as AI-generated material enters quality systems and regulatory documentation.
Organisations still need to demonstrate that records remain attributable, accurate,
contemporaneous, reviewable and scientifically justified regardless of how the first draft was produced.
In practice, this means firms need to understand where AI contributes to CMC
documentation, how outputs are reviewed, who approves them and whether reviewers have sufficient scientific and manufacturing expertise to identify missing or phase-inappropriate content.
Using AI in CMC without losing scientific oversight
At 3Biotech, AI already forms part of our internal workflow. It supports technical drafting,
information organisation, literature synthesis and early-stage document structuring. Used
appropriately, it reduces administrative burden and accelerates parts of the development
process that do not require repeated manual reconstruction.
However, scientific accountability still sits with experienced reviewers.
Every CMC recommendation ultimately needs to withstand technical scrutiny, manufacturing reality and regulatory review. Early-stage biotech companies are often working with incomplete datasets, evolving process understanding, limited development timelines and significant investor pressure. Under those conditions, apparently small technical assumptions can influence programme viability much later in development.
AI systems reduce the time required to generate technically coherent documentation. That increases the importance of experienced review because scientific omissions and phase-inappropriate assumptions become harder to detect once presented in polished technical language.
Producing technically coherent documentation is becoming easier. But recognising where that documentation is incomplete, operationally unrealistic or strategically unsafe still depends on experienced review.
Scientific judgement in these situations develops through direct experience managing
investigations, defended regulatory positions, resolved scale-up failures and seen how early technical decisions behave under commercial manufacturing conditions.
AI can support drafting, analysis and information handling within CMC workflows. But it still takes experienced scientific review to identify missing validation activities, weak process assumptions, phase-inappropriate strategies and technically plausible conclusions that do not withstand manufacturing or regulatory scrutiny.
References:
1. FDA Warning Letter: Purolea Cosmetics Lab, April 2026 https://www.fda.gov/inspections-compliance-enforcement-and-criminal-investigations/warning-letters/purolea-cosmetics-lab-722591-04022026
2. FDA, Guiding Principles of Good AI Practice in Drug Development
3. EMA, Reflection Paper on the Use of Artificial Intelligence in the Medicinal Product
4. MHRA Guidance on GxP Data Integrity