EU GMP Annex 22 and AI in Pharma and Biologics

What is Annex 22 on artificial intelligence

EU GMP Annex 22 is a proposed annex addressing artificial intelligence in pharmaceutical manufacturing. This white paper explains the draft Annex 22 AI guidelines and examines how BioMedAna’s scientific models, explainable AI and human oversight support implementation planning for pharma and biologics organizations. The discussion focuses on process optimization, quality prediction and deviation investigation. [1–3]

Biologics, also called biological medicines or biologicals, require analytical approaches that preserve the relevant product and process context. BioMedAna combines physics, chemistry, biology, data and machine learning through modality-specific models. Its agentic AI assists experts in preparing and reviewing scientific work, with operational authority retained by people. [5]

Regulatory context: Annex 22 remains a draft in the sources reviewed. References to Annex 22 requirements in this paper mean proposed expectations, not adopted requirements or regulatory approval of BioMedAna. [1–3]

Executive summary

Biologics leaders need AI that improves the quality of scientific decisions while preserving accountability for how those decisions are made. Process optimization, quality prediction and deviation investigation each require a clear connection between data, scientific reasoning, model outputs and expert judgment.

BioMedAna combines mechanistic and deterministic models with in-house hybrid models built for specific modalities. Explainable AI supports review of model behavior, while agents assist with analytical work and preparation of findings. BioMedAna reports that humans review work at every stage and that agents do not execute operational actions. This architecture provides a strong foundation for alignment with the expectations being developed through EU GMP Annex 22.

The business case is a controlled route from scientific analysis to a reviewable decision. A useful output includes the applicable model, the evidence used, its limitations, uncertainty and the basis for a proposed next step. The accountable scientist or quality professional can then decide whether further testing, investigation or an approved operational change is warranted.

What this paper establishes

This paper explains the regulatory boundary, presents BioMedAna’s scientific and governance approach, and maps company-reported capabilities to the draft’s subject areas. It also proposes the evidence a biologics organization should examine for its own deployment. Architectural alignment is a foundation for use-specific qualification; it is not a platform-wide certification or a substitute for the manufacturer’s quality system.

Regulatory baseline: the July 2025 consultation draft and the June 2026 EMA workshop report. The Commission’s published GMP index does not list an adopted Annex 22. [1–3]

Capability basis: BioMedAna management’s description of implemented controls and the public website. The mapping is not an independent audit of validation records. No quantified customer outcomes are asserted. [4–5]

Annex 22 guidelines on AI and the boundary for GMP use

What AI does draft Annex 22 cover and restrict

The draft covers static, deterministic ML in critical GMP applications. It says adaptive models, non-repeatable outputs and generative AI or LLMs should not be used there. Non-critical generative uses require qualified human responsibility. These are draft positions, not an adopted authorization framework. [1]

In June 2026, EMA convened discussion of broader inclusion of adaptive, probabilistic and generative models under robust risk controls. The report records regulatory questions and industry proposals intended to inform further drafting. It does not itself amend the published draft or authorize a deployment. [2]

How to apply this distinction

For BioMedAna, the practical assessment should begin with a workflow and the decision it supports. A research scenario comparison, a quality prediction used in a manufacturing decision and a narrative prepared for a deviation record may have very different consequences. The product label “assistant” cannot resolve those differences.

A useful deployment review asks: What could an incorrect output cause? Can the reviewer detect the error using independent evidence? Does the output influence product quality, patient safety or the integrity of a regulated record? These questions should be answered before choosing the permitted role of any generative component.

Preventing an agent from changing a process setting is an important safeguard. It does not remove the possibility that a persuasive but incorrect recommendation influences the person who changes that setting. Human review must therefore be substantive and supported by accessible source evidence.

A repeatable prediction can still express uncertainty

A fixed model can return the same estimated probability or prediction interval for identical inputs. That is different from an output that changes between repeated executions. For implementation review, BioMedAna should demonstrate repeatability of the relevant deployed pipeline, including preprocessing and model configuration, while separately explaining the uncertainty of its predictions.

Mechanistic, deterministic and machine-learning describe different properties. A mechanistic model expresses process relationships; deterministic behavior concerns reproducibility; a hybrid model combines approaches. None of these labels, on its own, establishes performance for a particular manufacturing use.

AI in biologics using modality specific scientific models

BioMedAna describes its differentiation as modality-native modeling: mechanistic models, deterministic models and internally developed hybrid models tailored to the science of the process. Management identifies eight supported groups: monoclonal antibodies; recombinant proteins; microbial and precision fermentation; chemical and pharmaceutical APIs; cell-culture viral vaccines; cell therapies; gene therapies and viral vectors; and RNA/oligonucleotides. These are BioMedAna’s portfolio groupings, rather than a claim that all eight share the same regulatory or biological characteristics. [5]

Why the distinction matters to an executive

A portfolio can share data infrastructure and governance while requiring different scientific assumptions for each process. An optimization objective is only meaningful alongside the constraints that protect the product. Improving a predicted yield, for example, is insufficient if the proposed conditions undermine an important quality attribute.

The following are illustrative design questions, not assertions of validated BioMedAna performance: Does an antibody process model account for the applicable culture and product-quality context? Does a recombinant-protein investigation preserve the applicable expression, purification and activity evidence? Does a cell therapy analysis preserve the significance of starting-material variability? The model dossier should explain which relationships are represented and which remain outside its scope.

Combining scientific knowledge and learned patterns

The value of a hybrid approach is the opportunity to combine established process relationships with patterns learned from data. Its credibility depends on demonstrating where each component contributes. Scientific structure can make assumptions easier to inspect, but it cannot guarantee that a model generalizes to another product, site, scale or operating range.

We recommend presenting each deployed model through a concise scientific dossier: process and product context; modeled relationships; training domain; applicability limits; assumptions; and comparison with a suitable baseline. Any claimed advantage of the hybrid approach should be measured on the actual intended task.

Explainable AI and SHAP in pharmaceutical manufacturing

The reviewed BioMedAna Twin implementation exposes global feature importance, which is distinct from per-prediction SHAP attribution. Where a suitable, evaluated SHAP method is available for the selected model, it attributes a model output to input features; it does not, by itself, demonstrate that changing a feature will cause the desired process outcome. Correlated inputs and confounding can complicate interpretation. [6]

For an investigation or optimization proposal, feature attribution should be examined alongside mechanistic plausibility, the source measurements and competing explanations. The useful question is whether the evidence supports the proposed next step, rather than whether an explanation chart is available.

AI in pharma for process optimization quality prediction and deviation investigation

The following examples illustrate how BioMedAna’s stated capabilities could be applied. They are proposed workflow patterns, not customer case studies or claims of achieved performance.

AI for biologics process optimization

A process-development team compares candidate operating conditions using an approved model and a defined objective. The analysis preserves relevant quality constraints and identifies scenarios that extend beyond the available evidence. An agent can coordinate analysis and prepare a study proposal for review; the responsible scientist determines whether the proposal warrants an experiment.

The review package should make the baseline, candidate conditions, assumptions and uncertainty visible together. A promising scenario becomes a hypothesis to evaluate through an authorized study or change process. A model-predicted improvement should not be presented as an experimentally demonstrated improvement.

AI for pharmaceutical quality prediction

A team estimates a quality attribute from available process measurements. Before relying on the estimate, the reviewer checks that the correct model and input context were used and examines the supporting evidence. A missing measurement, unfamiliar operating regime or weakly supported estimate should trigger an appropriate alternative path rather than an unsupported conclusion.

The use-specific assessment should distinguish an early development signal from an estimate used in a critical quality decision. The same analytical method can demand different controls depending on this role. A prediction does not itself authorize replacement of a required assay or a release decision.

AI assisted deviation investigation in GMP workflows

An agent organizes authorized records, constructs a chronology and prepares candidate explanations for expert review. The analytical model helps identify unusual relationships; mechanistic knowledge helps the investigator assess plausibility. The investigator checks the original records and tests competing explanations before accepting a conclusion.

A reviewable investigation should distinguish observations, model inferences and unresolved hypotheses. An agent-generated narrative must not quietly become the factual basis for closing an investigation. The accountable investigator and quality organization retain responsibility for conclusions and any subsequent corrective action.

The common decision record

Across these workflows, we recommend a compact record that identifies the question, authorized inputs, model version, analytical results, limitations, reviewer and approved next step. This makes the relationship between scientific work and organizational accountability inspectable without treating every application as identical.

Agentic AI in pharma with human oversight

BioMedAna reports that its agents can retrieve information, coordinate analyses, prepare experiments and draft recommendations, while humans review all stages and agents do not execute operational actions. Management also confirms auditability, guardrails, explainability and context boundaries. [5]

Separate analytical work from operational authority

The proposed control pattern is straightforward: an authorized request defines the task; approved data and tools support the analysis; the system prepares a review package; an accountable expert evaluates it; and any operational action proceeds through the organization’s authorized workflow. This is an implementation recommendation for making the stated separation demonstrable.

Agent permissions should reflect this boundary. Being able to prepare a model configuration or draft a study does not imply permission to deploy a model, change a manufacturing setting, approve a quality record or close a deviation. The organization should test those boundaries directly, including attempts to bypass them.

Make review an observable activity

A qualified reviewer needs the original evidence, time and authority to challenge or reject the output. Record material corrections, the reason for the decision and the accountable person. An approval click alone is weak evidence of scientific review.

Test the review arrangement under realistic workload and automation bias. Representative exercises should show whether reviewers detect plausible errors; a software review step cannot compensate for a workload that prevents meaningful assessment.

Test the boundaries of the agent

Challenge the agent with wrong-product context, missing records, conflicting measurements, an unauthorized tool request and misleading instructions in retrieved material. These are proposed tests, not completed test results. Record whether scope holds, how the limitation is shown and how the reviewer responds.

Retain the request, sources, tools and resulting decision. If a generative component materially influences a critical GMP decision, assess that role against the published regulatory position; human oversight is not a general exemption.

Annex 22 AI requirements mapped to BioMedAna capabilities

The clause labels below summarize the draft. [1] The capability column reflects management statements. [5] The evidence column is our proposed deployment assessment, not a finding that the records have been inspected. “Implemented” refers to the reported capability; adequacy remains specific to the deployment.

This is a navigation map, not a full conformity checklist. Review the complete source text for each clause and the applicable computerized-system and pharmaceutical quality-system obligations. [1,3]

A deployment gate for the draft Annex 22 assessment

Use the clause mapping to examine a particular workflow, not to issue a platform-wide readiness verdict. The following is a proposed review worksheet. It organizes the evidence discussion; the source draft and the regulated organization's quality system determine the applicable obligations.

The review outcome should be explicit: proceed within the supported scope, hold for missing evidence, or redesign the proposed use. A hold is accompanied by the missing record, its owner and the condition for reopening the gate. A proceed decision identifies what it permits and what remains excluded. Neither outcome substitutes for the organization's established approval procedures.

This also sharpens the BioMedAna evaluation. Ask to see the relevant source context, model configuration, retained test evidence and a reviewer handling a credible error. Where a capability is described as implemented, inspect the records for the offered configuration and intended use. A demonstration of a control is distinct from evidence that the complete customer workflow is suitable.

Evaluating Annex 22 readiness for a biologics AI deployment

BioMedAna’s reported controls provide the elements for a substantive assurance discussion. The next step for a biologics organization is to connect those controls to a defined use and a reviewable body of evidence. The following approach is a recommended diligence framework rather than a statement of additional Annex 22 requirements.

Select a bounded first application

Choose one process, product context and decision workflow. Establish the current method, the users and the consequences of a wrong result. Agree in advance how scientific usefulness will be judged. An initial deployment should be narrow enough that reviewers can explain both its value and its limits.

Review the scientific and operational evidence together

Scientific leaders should assess the model’s assumptions and relevance. Quality leaders should assess the decision pathway and retained records. Data and technology leaders should examine the implementation and permissions. Reviewing these perspectives together can reveal gaps that are invisible in a model accuracy presentation alone.

The evidence package should connect the deployed configuration to its evaluation results and the actual user workflow. A strong evaluation of a different version, population or process is not evidence for the proposed deployment. Any transfer of conclusions should have an explicit technical justification.

Measure benefits without inventing outcomes

Potential measures include the time required to prepare a reviewable analysis, the frequency of reviewer corrections, the performance of predictions in the intended population, and the proportion of investigation hypotheses supported by subsequent evidence. Select measures that reflect the workflow rather than a generic platform score.

BioMedAna should publish quantitative benefit claims only with a defined baseline, relevant sample, evaluation period and explanation of how results were obtained. Illustrative website values should remain labeled as illustrations. This paper does not claim a demonstrated reduction in deviations, development time or manufacturing cost.

The executive conclusion

BioMedAna’s positioning is strongest when it connects modality-specific scientific reasoning with explainable analytics and accountable human decisions. Its reported combination of implemented controls is consistent with an architecture intended to support Annex 22 alignment. The confidence of a customer or quality organization should then rest on the evidence for its particular deployment.

For biologics leaders, the appropriate next step is a focused demonstration of one of the three workflows, accompanied by the relevant model and governance records. This converts an architectural claim into an assessable implementation proposition.

Sources and basis of assessment

Regulatory sources were. Links below identify the exact documents used. The regulatory discussion distinguishes the published consultation draft, subsequent workshop discussion and the current Commission index.

1 European Commission — Draft Annex 22 Artificial Intelligence

July 2025 consultation text. Primary reference for the scope summary and clause mapping.

2 EMA — Multistakeholder workshop on Annex 22

Report dated 30 June 2026. Records discussion to inform future guidance; not an adopted replacement text.

3 European Commission — EudraLex Volume 4

Published GMP guidance index. No adopted Annex 22 was listed at the time of review.

4 BioMedAna — Public product website

Supports the public positioning around modality-specific science, evidence-linked work and expert oversight. Demonstration values are illustrative.

5 BioMedAna management statements

Supplied product and control descriptions: three target workflows; mechanistic, deterministic and in-house hybrid models; eight portfolio groups; human review at all stages; agents without operational execution authority; auditability, guardrails, explainability and context boundaries; and implemented versioning, independent test data, acceptance criteria, access controls, change approval, monitoring, uncertainty assessment and validation documentation. These statements were not independently audited for this paper.

6 SHAP — Predictive explanation and causal interpretation

Technical explanation of limitations when interpreting predictive models as causal evidence.