Global life sciences AI · Published 16 November 2026 · HERMS Consultancy DMCC
See how HERMS delivers integrated life sciences solutions.
AI in pharmaceutical manufacturing can support process understanding, monitoring, investigation and operational planning, but it must sit within an effective pharmaceutical quality system. Define the intended use, establish reliable data and meaningful acceptance criteria, validate the configured workflow, control changes and retain accountable human oversight. An anomaly score is not proof of a deviation; a proposed root cause is not an approved conclusion; and a model prediction is not automatic authority to release a batch. The control strategy must remain scientifically justified throughout development and commercial supply. Process Validation: General Principles and Practices Artificial Intelligence in Drug Manufacturing Q9(R1) Quality Risk Management Q10 Pharmaceutical Quality System EudraLex — Volume 4 — Good Manufacturing Practice guidelines
Start with the quality decision—not the AI technology
Specify the manufacturing problem, product, process stage, users and decision. Will AI identify unusual sensor patterns, estimate an unmeasured process attribute, prioritize maintenance or prepare an investigation summary? These activities have different consequences and evidence needs. Record what the system may read, recommend, write and control.
Distinguish predictive models, generative assistants, conventional statistical methods and deterministic automation. Predictive monitoring estimates patterns or outcomes; generative AI produces content; a workflow engine routes actions. A hybrid can be useful, but a conversational interface does not establish process understanding or GxP fitness. FDA’s 2023 manufacturing paper discusses possible AI uses; it explicitly states that it is neither draft nor final guidance. Artificial Intelligence in Drug Manufacturing
Map controls across the manufacturing lifecycle
FDA’s process-validation guidance uses a lifecycle approach: process design, process qualification and continued process verification. AI-related evidence should connect to that lifecycle rather than becoming a parallel demonstration project. Validating a model does not independently validate the manufacturing process. Process Validation: General Principles and Practices
| Lifecycle stage | Potential AI contribution | Essential quality boundary | Evidence to retain |
|---|---|---|---|
| Development | Explore relationships and propose experimental priorities | Confirm hypotheses through appropriate scientific work | Data, assumptions, experiments and limitations |
| Scale-up / transfer | Compare process behavior and readiness gaps | Demonstrate relevance at the receiving scale/site | Equipment, material and process comparability |
| Qualification | Support data review and test analysis | Approved qualification/validation protocol and acceptance | Requirements, results, deviations and approvals |
| Commercial monitoring | Detect patterns and predict quality risk | Established control strategy and escalation | Model/version, inputs, alarms and responses |
| Investigation / CAPA | Assemble records and propose hypotheses | Quality-approved conclusions and effectiveness review | Sources, rationale, actions and outcomes |
| Batch disposition | Compile records and flag exceptions | Retain applicable authorized release responsibility | Controlled evidence and required certification |
| Supply / change | Identify resilience risks and support planning | Approved change and distribution controls | Impact assessment, decisions and supply records |
This is a proposed implementation map, not a universal mandatory AI workflow. The appropriate controls follow the product, process, jurisdiction and degree of reliance. ICH Q9(R1) emphasizes risk-informed decisions and appropriate formality; ICH Q10 places lifecycle activities within the pharmaceutical quality system. Q9(R1) Quality Risk Management Q10 Pharmaceutical Quality System
Development and scale-up: use models to test understanding
During development, AI may help identify material/process relationships or suggest informative experiments. Keep the link to critical quality attributes, or CQAs, and the variables that may affect them. A model correlation is not evidence that changing a parameter will cause a desired quality improvement. Consider experimental design, mechanistic knowledge and confounding in historical production data. Artificial Intelligence in Drug Manufacturing Q9(R1) Quality Risk Management
Scale-up and technology transfer change the operating environment. Mixing, equipment geometry, sensors, raw-material sources and sampling can differ. Demonstrate which relationships remain applicable and where additional evidence is required. Do not assume that a model trained on development batches is qualified for commercial use or transfers unchanged between CDMOs. Process Validation: General Principles and Practices Q10 Pharmaceutical Quality System
Data integrity is a prerequisite for reliable monitoring
Connect the model to trustworthy, contextualized records. Relevant inputs may come from historians, MES, LIMS, electronic batch records, equipment systems and quality records. Align batch, material, equipment and sample identifiers; units; calibration status; timestamp references; and process phases. A synchronized file transfer does not establish that signals represent the same event.
Assess sensor gaps, maintenance periods, manual entries, repeated records, changed methods and process interventions. Distinguish a missing value from a confirmed zero. Preserve raw data and metadata needed to understand transformations. FDA’s data-integrity guidance expects reliable and accurate CGMP data and supports meaningful, risk-based strategies to prevent and detect integrity problems. Data Integrity and Compliance With Drug CGMP: Questions and Answers — Guidance for Industry
Prevent leakage between development and evaluation datasets. Repeated measurements from one batch are not independent batches. Split at a level appropriate to the intended claim, and assess time, equipment and site effects. Ensure variables are available at the actual decision time rather than only after completed testing.
Process monitoring: separate alerts, evidence and actions
An anomaly detector can identify a departure from learned patterns without knowing whether product quality is affected. Conversely, an apparently familiar pattern may still be unacceptable. Interpret alerts against process knowledge, validated limits and the control strategy—not only a probability or distance score.
| Signal | What it may indicate | Required distinction |
|---|---|---|
| Sensor anomaly | Process change, instrument fault or data issue | Unusual does not necessarily mean out of specification |
| Predicted CQA | Estimated quality attribute based on an evaluated model | Prediction is not a direct assay or automatic release criterion |
| Multivariate trend | Combined movement across variables | Association needs process interpretation |
| Maintenance forecast | Potential equipment degradation | Schedule/action must fit qualified maintenance procedures |
| Alarm increase | Changed process, configuration or false-positive burden | Do not silently weaken thresholds to improve a dashboard |
PAT is a broader framework for process understanding and control, not a synonym for AI. An AI model may operate within a PAT strategy, but adoption of PAT does not by itself authorize a new release approach. Define what happens when a model is unavailable, inputs are unsupported or predictions conflict with other evidence. PAT — A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance EudraLex — Volume 4 — Good Manufacturing Practice guidelines
Advisory monitoring and closed-loop control need different safeguards
An advisory system informs a person. A closed-loop system can adjust a process automatically. As direct influence grows, the consequences of error, delay, inappropriate inputs and control instability become more important. AI model performance alone is not enough: evaluate the complete sensor-to-action chain and its interaction with the established process. Artificial Intelligence in Drug Manufacturing
As proposed engineering safeguards, use authorized parameter boundaries, independent protections where justified, deterministic checks, defined stop states and tested fallback. Assess cumulative adjustments and response time, not just individual recommendations. Changes to a registered control strategy or manufacturing process may have regulatory consequences; obtain the relevant assessment before implementation.
Deviation investigation: a proposed cause is not a root cause
AI can assemble chronologies, find related deviations and summarize maintenance, material and batch records. It can also produce a plausible but unsupported causal narrative. Require each material assertion to link to evidence and separate observed facts, hypotheses and established conclusions.
Do not treat a lower model score, a convenient operator explanation or correlation with a raw-material lot as sufficient justification to invalidate a result. Preserve original records, follow approved investigation procedures and assess affected product and potentially related batches. The pharmaceutical quality system must govern the investigation and decision. Q10 Pharmaceutical Quality System Data Integrity and Compliance With Drug CGMP: Questions and Answers — Guidance for Industry
For out-of-specification and out-of-trend events, retain the distinction between an approved specification failure and a trend signal. AI may assist triage, but it should not silently recategorize an event to avoid investigation. This is a proposed governance boundary; exact procedures follow the applicable quality system and regulations.
CAPA: measure effectiveness, not document completion
A corrective and preventive action plan needs an evidence-supported rationale, responsible owner, implementation record and effectiveness assessment. AI-generated actions can be repetitive or disconnected from the actual cause. Completion of training or an updated SOP does not by itself demonstrate that the failure mechanism is controlled. Q10 Pharmaceutical Quality System
Define effectiveness criteria before closure, proportionate to the problem. Review recurrence, related events and whether the control introduced new risks. Avoid using a reduction in reported deviations as the sole success measure: it can reflect fewer events, weaker detection or changed reporting behavior.
Validate the configured AI-supported workflow
Define intended use, system boundaries, data sources, outputs, action permissions and acceptance criteria before testing. Assess product-quality impact, patient risk, record integrity, uncertainty and detectability. A supplier benchmark or generic “validated AI” claim does not establish suitability for the site’s process. Process Validation: General Principles and Practices Q9(R1) Quality Risk Management EudraLex — Volume 4 — Good Manufacturing Practice guidelines
| Test area | Illustrative challenge | Acceptance evidence |
|---|---|---|
| Analytical performance | Representative batches, rare failures and relevant operating ranges | Task-specific results and uncertainty |
| Data handling | Missing, delayed, corrupted or mislabelled signals | Detection, safe handling and escalation |
| Workflow | Rejected, stale or unauthorized recommendations | No unintended execution; correct role routing |
| Traceability | Reconstruct a sampled prediction and decision | Source, configuration, result and reviewer history |
| Operational resilience | Network/provider outage and restore | Documented fallback and reconciled records |
| Human review | Seeded errors under realistic workload | Reviewers detect consequential issues with usable evidence |
| Change behavior | New raw material, equipment or model version | Impact assessment and justified retesting |
This test matrix is illustrative. Validation of the computerised system, assurance of the model and qualification/validation of the manufacturing process are related but distinct. Keep requirements, risks, tests, deviations and acceptance decisions connected. EU GMP Annex 11 and Annex 15 address computerised systems and qualification/validation respectively. EudraLex — Volume 4 — Good Manufacturing Practice guidelines
Human oversight, batch disposition and electronic records
Assign the process owner, system owner, technical owner, Quality approver and relevant authorized release role. Reviewers need evidence, competence, time and authority to reject a recommendation. A human approval button adds little if the system has already changed a parameter or the reviewer cannot inspect the source.
EU Annex 16 addresses Qualified Person certification and batch release; Annex 17 addresses real-time release testing and parametric release. Do not equate model-assisted record review, predictive quality or RTRT with replacing the required authorized responsibility. The applicable product authorization and approved release strategy remain important. EudraLex — Volume 4 — Good Manufacturing Practice guidelines
Determine which electronic records and signatures fall within 21 CFR Part 11 and underlying predicate rules. FDA’s scope guidance is nonbinding and describes limited enforcement discretion; Part 11 remains in effect. It does not provide a blanket exemption from trustworthy required records, access controls or signature requirements. Part 11, Electronic Records; Electronic Signatures — Scope and Application
Retain decision-relevant source data, model/configuration versions, preprocessing, material tool actions, reviewer changes, approvals and final execution results. Protect identities, timestamps, retention and record availability. An audit trail helps reconstruct a decision; it cannot make an unsupported conclusion scientifically sound.
Change control and drift: no silent retraining
Track model code, weights, thresholds, features, preprocessing, sensor configuration, dependencies and interfaces. For generative tools, prompts, retrieval sources and external service versions can also affect outputs. Assess changes before release and bind acceptance to the actual configuration.
Differentiate process drift, sensor drift, changes in input distribution and deteriorating model performance. A changed distribution is a reason to investigate, not automatic permission to retrain. Preserve a controlled reference, evaluate whether the underlying process has changed and determine the appropriate quality and regulatory action. Process Validation: General Principles and Practices Q9(R1) Quality Risk Management Q10 Pharmaceutical Quality System
Continuous learning is not an exemption from change control. Define when monitoring reduces autonomy, triggers deviation handling or requires suspension. A rollback must account for affected batches, records and downstream actions; it is more than restoring a software file. Medical-device PCCP guidance should not be assumed to govern drug manufacturing model changes.
Biologics, sterile products and advanced therapies need product-specific controls
Bioprocesses involve raw-material variability, complex measurements and scale-dependent behavior. AI may support trend interpretation or process development, but sparse datasets and limited failure examples can weaken evaluation. Examine whether the data support the claimed operating range and product context.
Do not assume software can replace contamination control, aseptic assurance, appropriate testing or identity controls. EU GMP provides specific guidance for sterile manufacture and ATMPs, including Part IV for advanced therapies. Use the applicable product framework and site procedures rather than a generic AI checklist. EudraLex — Volume 4 — Good Manufacturing Practice guidelines
Commercial supply: quality intelligence must follow the product
Commercial readiness includes validated operation, maintained equipment, controlled materials and reliable handoffs between manufacturer, quality function, MAH and distribution partners. A forecast of stock availability cannot compensate for unresolved quality impact or unauthorized release.
AI may support demand, maintenance or supply-risk planning, but distinguish these uses from GMP quality decisions. Define ownership of excursions, holds and communications, and evaluate outsourced activities through the pharmaceutical quality system. ICH Q9(R1) explicitly addresses product-availability risks from manufacturing quality issues; resilience and quality should not be managed as unrelated agendas. Q9(R1) Quality Risk Management Q10 Pharmaceutical Quality System EudraLex — Volume 4 — Good Manufacturing Practice guidelines
Hypothetical scenario: an unexpected bioreactor trend
A monitoring model flags an unfamiliar bioreactor pattern. The operator does not automatically change the recipe. The approved workflow checks sensor status and phase alignment, preserves the data and routes the event to process and Quality personnel.
Review identifies a possible material-related difference, but evidence is insufficient to establish the cause or product impact. The team follows approved investigation and testing procedures, assesses related batches and documents the decision. If a model or control-strategy change is proposed, it enters change control before deployment.
This scenario illustrates a proposed response, not a reported HERMS deployment. The useful boundary is the separation between detection, investigation, scientific conclusion, process adjustment and authorized batch disposition.
Regulatory landscape: distinguish existing GMP from proposals
| Resource | Status as checked | What it means for manufacturing AI |
|---|---|---|
| FDA process validation | January 2011 final guidance | Lifecycle validation principles; not model-only approval |
| ICH Q9(R1) / Q10 at FDA | Final guidance, May 2023 / April 2009 | Risk management and pharmaceutical quality-system principles |
| FDA AI manufacturing paper | 2023 discussion paper, not draft/final guidance | Discusses possibilities and policy questions; no blanket authorization |
| EU GMP Annexes 11/15/16/17 | Current index lists established guidance within EU GMP framework | Computerised systems, validation, QP responsibilities and release strategies |
| Proposed Annex 22 / revised Annex 11 | 2025 consultation drafts; final adoption not established | Do not describe proposed AI-specific text as effective final requirements |
| FDA integrity / Part 11 guidance | Final nonbinding guidance; underlying regulations apply where relevant | Accurate reliable data and appropriate records/signature controls |
The Commission consultation ran from 7 July to 7 October 2025 and identifies new Annex 22 and revised Annex 11/Chapter 4 as draft texts. The current EudraLex index reviewed still lists Annex 11, January 2011. Treat Annex 22 as proposed unless a final official publication establishes otherwise. FDA’s 2023 AI manufacturing discussion paper must likewise not be promoted to guidance or law. Artificial Intelligence in Drug Manufacturing EudraLex — Volume 4 — Good Manufacturing Practice guidelines Stakeholders’ Consultation on EudraLex Volume 4 — Good Manufacturing Practice Guidelines: Chapter 4, Annex 11 and New Annex 22
A practical 90-day readiness roadmap and KPIs
| Planning window | Proposed deliverables | Acceptance question |
|---|---|---|
| Days 1–30 | Intended use; data/decision map; risk assessment; roles and source review | Is the problem well-defined and data fit for the task? |
| Days 31–60 | Configured controls; test plan; representative evaluation; fallback and training | Can the workflow detect and safely handle consequential failures? |
| Days 61–90 | Independent review; resolved critical findings; controlled pilot/release; monitoring | Is residual risk acceptable under approved procedures? |
These windows are illustrative, not implementation promises. Define critical-event false negatives, false-positive burden, source-to-decision traceability, post-review corrections, overdue changes, time to investigate alerts and CAPA effectiveness. Report denominators and severity. Pair efficiency with quality; no universal numerical target is defensible across all processes.
HERMS perspective: connect process evidence to accountable decisions
HERMS positions AXION around manufacturing quality, SYNTRIX around bioprocess/scale-up, INNEXIS around technology transfer, VERIQIS around evidence integrity and ARQON around supply/cold chain. These are company descriptions, not independently validated features, regulatory acceptance or measured outcomes. A useful demonstration should show how the proposed workflow preserves sources, configuration, decision rights and relevant approvals. AI Life Sciences Software | 25 HERMS Solutions
Explore HERMS manufacturing quality and bioprocess solutions to discuss your process-monitoring, validation and quality-oversight requirements.
Frequently asked questions
How is AI used in pharmaceutical manufacturing?
Potential uses include process-design analysis, trend monitoring, anomaly detection, maintenance support and evidence assembly. The risk depends on intended use and whether the system only recommends or directly controls actions. FDA’s manufacturing discussion paper describes possibilities, not a blanket approval for any deployment. Artificial Intelligence in Drug Manufacturing
Does every AI application require the same validation?
No. Define the activity, regulated impact, records and reliance. Assurance should be proportionate to risk, uncertainty and potential quality consequences. A draft-writing aid and a model influencing process control need different evidence. Existing applicable process and computerized-system controls still matter. Process Validation: General Principles and Practices Q9(R1) Quality Risk Management EudraLex — Volume 4 — Good Manufacturing Practice guidelines
Can AI approve a pharmaceutical batch?
AI may assist record checks, predictions and exception review. It does not replace required authorized quality decisions or QP certification. EU Annex 16 addresses QP certification/batch release; Annex 17 addresses distinct controlled release strategies. Apply the jurisdiction, product authorization and approved procedures. EudraLex — Volume 4 — Good Manufacturing Practice guidelines
Is PAT the same as AI?
No. Process Analytical Technology is a broader framework supporting process understanding and innovative control. AI may be one component, but PAT can use other analytical and statistical methods. Neither label independently validates the process or authorizes real-time release testing. PAT — A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance EudraLex — Volume 4 — Good Manufacturing Practice guidelines
Can AI identify the root cause of a deviation?
It can suggest hypotheses and assemble evidence, but a plausible explanation is not an established cause. Qualified investigators and Quality must assess supporting facts, alternatives and product impact through the approved process. Preserve the original evidence and distinguish generated suggestions from accepted conclusions. Q10 Pharmaceutical Quality System Data Integrity and Compliance With Drug CGMP: Questions and Answers — Guidance for Industry
Is EU GMP Annex 22 already effective?
Final adoption was not established in the reviewed sources. The official consultation identifies Annex 22 as a new draft and closed in October 2025; the current GMP index still lists the established Annex 11. Confirm a final official publication before asserting effective AI-specific obligations. EudraLex — Volume 4 — Good Manufacturing Practice guidelines Stakeholders’ Consultation on EudraLex Volume 4 — Good Manufacturing Practice Guidelines: Chapter 4, Annex 11 and New Annex 22
Can a manufacturing model retrain automatically?
Do not assume retraining is allowed merely because it is technically possible. Define the model’s role, relevant authorization, quality controls and change procedure. Assess data, process changes, performance and impact before approving a new configuration. Continuous improvement is not an exemption from controlled change. Process Validation: General Principles and Practices Q10 Pharmaceutical Quality System
How should teams measure business value?
Compare accepted, quality-checked work against a baseline, including review, rework, integration and maintenance. Track critical misses, false alarms, traceability and post-approval corrections alongside time or cost. Faster output is not value if it weakens detection, scientific conclusions or product-quality assurance.
The strongest manufacturing AI strategy makes the process more understandable without obscuring accountability. Keep the intended use explicit, evidence traceable, changes controlled and required quality decisions assigned to authorized people. That discipline matters from early development to the final commercial handoff.
Discuss a quality-readiness assessment or focused demonstration through HERMS Solutions , or explore related HERMS manufacturing and life-sciences insights .
Sources and further reading
- Process Validation: General Principles and Practices
- Artificial Intelligence in Drug Manufacturing
- Q9(R1) Quality Risk Management
- Q10 Pharmaceutical Quality System
- EudraLex — Volume 4 — Good Manufacturing Practice guidelines
- Data Integrity and Compliance With Drug CGMP: Questions and Answers — Guidance for Industry
- PAT — A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance
- Part 11, Electronic Records; Electronic Signatures — Scope and Application
- Stakeholders’ Consultation on EudraLex Volume 4 — Good Manufacturing Practice Guidelines: Chapter 4, Annex 11 and New Annex 22
Explore the pharmaceutical AI series
- AI decision authority and human oversight
- Agentic versus generative AI in pharmaceuticals
- AI governance in life sciences
- Human approval controls for pharmaceutical AI
- Real-world evidence for regulatory and market access
- AI medical device validation and change control
- Global pharmaceutical market access, HTA and pricing evidence
- Preparing healthcare data for AI
Explore HERMS life sciences solutions and AI governance and traceability, or request a workflow-focused demonstration.
Relevant HERMS solutions
Explore the HERMS solutions relevant to this article and assess their fit for your workflow, evidence and approval requirements.
