AIforPharma.org

Introduction

⇒ FDA reports more than 1,500 AI-related drug submissions as FDA and EMA regulators call for greater international harmonisation of AI standards, lifecycle controls and evidence requirements across pharmaceutical development.

Artificial intelligence is no longer a niche regulatory topic in pharmaceutical development.

At RAPS Convergence 2026, held in Charlotte in September, representatives from FDA, EMA and industry discussed how rapidly AI use is expanding across the medicines lifecycle and why international harmonisation is becoming increasingly important.

One number is particularly striking.

According to Gabriel Innes, Assistant Director for Data Science and AI Policy at FDA’s Center for Drug Evaluation and Research (CDER), the Center has now received more than 1,500 regulatory submissions involving AI.

These submissions include:

• Investigational New Drug applications (INDs);
• New Drug Applications (NDAs);
• Biologics License Applications (BLAs);
• Abbreviated New Drug Applications (ANDAs);
• Emergency Use Authorisations;
• Drug Development Tools.

The implication is important:

AI is moving from experimental innovation toward routine regulatory practice.

The challenge for regulators is therefore changing.

The question is increasingly not whether AI will be used in pharmaceutical development, but how regulators can establish sufficiently consistent expectations for its use.

More Than 1,500 AI-Related Submissions Changes the Regulatory Context

FDA had previously reported experience with more than 500 submissions containing AI components between 2016 and 2023.

The figure reported at RAPS Convergence indicates how rapidly this activity has expanded.

This matters because regulatory frameworks designed when AI use was relatively uncommon now need to support much larger numbers of applications.

For pharmaceutical companies, this also means that regulatory experience with AI is accumulating quickly.

AI can no longer be treated solely as:

an innovation project → a pilot → a future regulatory problem.

It increasingly needs to be treated as part of normal pharmaceutical development and regulatory strategy.

Context of Use Remains Central

FDA’s current approach is based strongly on Context of Use.

Its January 2025 draft guidance on AI supporting regulatory decision-making proposes a risk-based credibility framework in which the necessary level of assurance depends on:

• what the AI model does;
• how its output is used;
• how important that output is to the regulatory decision;
• what could happen if the output is wrong.

This creates a fundamentally different question from:

“Is this AI model validated?”

The more useful question becomes:

“Is this AI model sufficiently credible for this particular Context of Use?”

A model used to identify information for expert review does not necessarily require the same evidence as a model generating information that materially influences a safety, efficacy or quality conclusion.

FDA and EMA Already Agree on the High-Level Principles

In January 2026, FDA and EMA jointly published ten Guiding Principles of Good AI Practice in Drug Development.

The principles cover:

• human-centric design;
• risk-based approaches;
• adherence to relevant standards and GxP;
• clear Context of Use;
• multidisciplinary expertise;
• data governance and documentation;
• model design and development practices;
• risk-based performance assessment;
• lifecycle management;
• clear communication of essential information.

Importantly, these principles apply across the pharmaceutical lifecycle, including:

nonclinical development → clinical development → manufacturing → post-marketing activities.

This means that international regulatory convergence has already started at the level of general principles.

But Principles Are Not the Same as Standards

The discussion at RAPS Convergence highlighted an important gap.

Regulators can agree on principles such as:

reliability → transparency → lifecycle management → human oversight

while companies can still face uncertainty about how to demonstrate compliance in practice.

Examples include:

• Which performance metrics should be used?
• How much independent test data are necessary?
• What constitutes sufficient robustness?
• How should uncertainty be documented?
• When does a model change require revalidation?
• How should third-party AI models be controlled?
• What evidence is needed for foundation models or externally hosted AI?

This is where international technical standards and more detailed regulatory guidance become important.

EMA: Accuracy, Robustness and Explainability Remain Priority Questions

During the RAPS discussion, EMA highlighted findings from its recent regulatory science work.

Three questions remain particularly important:

• How can AI accuracy and reliability be demonstrated?
• How can models remain robust and trustworthy as data and conditions change?
• When is explainability important for regulatory decision-making?

These questions point toward a lifecycle concept of AI assurance.

Initial validation may demonstrate acceptable performance at deployment.

But regulated pharmaceutical use may also require evidence that the model continues to perform adequately over time.

This introduces:

performance monitoring → drift detection → change assessment → reassessment / revalidation.

International Convergence Is Already Broader Than FDA and EMA

EMA also described collaboration with organisations including the International Coalition of Medicines Regulatory Authorities (ICMRA) and the World Health Organization.

This is important for multinational pharmaceutical companies.

AI systems are rarely developed for one regulatory jurisdiction only.

A global pharmaceutical company may use the same AI platform across:

• the United States;
• the European Union;
• Canada;
• Japan;
• other PIC/S or ICH jurisdictions.

Substantially different national AI-validation expectations would create considerable complexity.

International convergence therefore has practical value beyond regulatory policy.

It could eventually allow companies to develop a more consistent global evidence package for AI systems.

Human Decision-Making Remains Central

Another important point concerned the regulator’s own use of AI.

FDA is itself adopting AI technologies internally.

However, the model described by FDA remains augmentation rather than replacement of expert regulatory judgment.

This principle has obvious relevance for pharmaceutical companies.

AI can support:

• information retrieval;
• data analysis;
• document review;
• regulatory intelligence;
• signal identification;
• drafting;
• workflow automation.

But the existence of an AI output does not transfer accountability away from the person responsible for the regulated decision.

This principle is especially important when AI is used in GxP environments.

AI Controls Should Depend on Capability, Integration and Risk

The RAPS discussion also highlighted the importance of distinguishing AI systems according to their actual technical characteristics.

Not every AI system creates the same regulatory problem.

Relevant differences include:

• static versus adaptive models;
• deterministic versus probabilistic outputs;
• task-specific models versus general-purpose foundation models;
• standalone models versus AI embedded in larger systems;
• advisory AI versus systems capable of initiating actions;
• internally controlled versus externally hosted models.

The required control architecture should therefore depend not only on the label “AI”, but on what the technology actually does.

This supports a risk-based sequence such as:

Intended Use → Context of Use → Technical Characteristics → Decision Impact → Risk → Controls → Evidence

What Pharmaceutical Companies Should Do Now

Companies do not need to wait for complete international harmonisation before establishing sensible AI governance.

Several foundations are already becoming consistent across FDA and EMA thinking.

A practical approach should include:

• an inventory of AI applications;
• clearly defined intended use and Context of Use;
• GxP and regulatory-impact classification;
• documented data provenance;
• risk-based performance requirements;
• independent testing where appropriate;
• human oversight;
• model and configuration version control;
• change management;
• ongoing performance monitoring;
• defined reassessment and revalidation triggers.

The future technical standards may change some details.

The fundamental control principles are already becoming relatively clear.

Why This Development Matters

The most important signal from RAPS Convergence is not a new regulation.

It is the scale of actual regulatory experience.

More than 1,500 AI-related submissions means that AI is increasingly entering real regulatory dossiers rather than remaining a theoretical future technology.

At the same time, FDA and EMA increasingly share the same high-level principles.

The next challenge is therefore:

turning regulatory principles into internationally consistent technical evidence standards.

For pharmaceutical companies, this suggests that successful AI governance should be designed with global convergence in mind rather than built independently for each jurisdiction.

Regulatory Status Note

The statements reported from RAPS Convergence represent discussion by regulators and industry experts and do not constitute new FDA or EMA guidance.

The figure of more than 1,500 AI-related submissions was reported by an FDA CDER representative during the RAPS Convergence panel.

The FDA/EMA Guiding Principles of Good AI Practice in Drug Development are official joint regulatory principles, but they do not replace applicable legislation, GxP requirements or application-specific regulatory guidance.

Sources

RAPS – Convergence: Regulators emphasize need for harmonized approach to AI, published 21 September 2026:

https://www.raps.org … -approach-to-ai.html

FDA – Guiding Principles of Good AI Practice in Drug Development:

https://www.fda.gov/ … ice-drug-development

EMA/FDA – Guiding Principles of Good AI Practice in Drug Development:

https://www.ema.euro … g-development_en.pdf

FDA – Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products:

https://www.fda.gov/ … -drug-and-biological

The UK Medicines and Healthcare products Regulatory Agency has opened a new initiative focused specifically on the use of artificial intelligence in medicines safety evaluation.

On 16 September 2026, MHRA launched the Beyond ADMET: AI for Medicines Safety stakeholder survey.

The initiative will support development of a future regulatory sandbox examining how AI could be used to improve prediction and assessment of medicines safety.

This is important because the regulator is not simply publishing general principles for responsible AI.

MHRA is actively asking pharmaceutical, biotechnology, toxicology and AI organisations:

What prevents AI models from being used confidently in regulatory medicines-safety evaluation?

Moving Beyond Classical ADMET Assessment

ADMET traditionally refers to:

• Absorption;
• Distribution;
• Metabolism;
• Excretion;
• Toxicity.

These characteristics are fundamental to understanding how a medicine behaves in the body and whether potential safety problems may occur.

Traditional safety evaluation relies on multiple sources of evidence, including:

• preclinical studies;
• toxicology;
• clinical trials;
• pharmacokinetic information;
• post-marketing information.

MHRA’s initiative explores whether AI can integrate a broader range of information, including traditional preclinical data together with clinical and real-world data.

The objective is to improve understanding of how medicines behave:

• across different populations;
• in different disease states;
• under different clinical conditions;
• in real-world use.

Potentially, this could help identify safety risks earlier.

The Regulator Is Explicitly Asking About Validation

One of the most important aspects of the initiative is the scope of the questions MHRA wants industry and researchers to address.

The Agency specifically identifies areas including:

• model development;
• validation;
• access to data;
• data sharing;
• regulatory considerations.

This is significant.

The central problem with AI in regulated pharmaceutical science is increasingly not whether sophisticated models can be built.

The more difficult question is:

What evidence is necessary before the regulator can rely on their output?

That requires moving from model development toward model credibility.

AI Performance Alone Is Not Enough

A highly accurate AI model may still be unsuitable for regulatory use if the organisation cannot adequately explain:

• what data were used;
• which population the data represent;
• how the model was validated;
• where its limitations are;
• whether performance generalises beyond the development dataset;
• how model changes are controlled.

This is particularly important for medicines safety.

A model can perform well overall while performing poorly for a specific:

• patient population;
• disease state;
• rare adverse event;
• drug class;
• clinical context.

Average performance therefore does not necessarily demonstrate regulatory credibility.

Data May Be the Bigger Problem Than the Algorithm

MHRA specifically identifies access to data and data sharing as areas of interest.

This reflects a broader problem in pharmaceutical AI.

Medicines-safety information is distributed across multiple sources:

• preclinical datasets;
• clinical trials;
• electronic health records;
• pharmacovigilance systems;
• real-world evidence;
• literature;
• laboratory data.

Connecting those datasets creates significant challenges involving:

• data quality;
• standardisation;
• missing data;
• population bias;
• privacy;
• interoperability;
• provenance.

A sophisticated AI model cannot compensate automatically for poor or non-representative data.

This leads to an increasingly important regulatory principle:

AI model credibility begins with data credibility.

Why a Regulatory Sandbox Is Important

The stakeholder survey will inform development of the Beyond ADMET regulatory sandbox, which MHRA announced earlier in 2026.

A regulatory sandbox is different from conventional guidance development.

Instead of attempting to define all regulatory expectations in advance, a sandbox allows regulators and innovators to explore real technologies and identify regulatory problems through practical cases.

This can help answer questions such as:

• What evidence is sufficient?
• Which validation approaches are practical?
• How should uncertainty be communicated?
• What documentation does a regulator need?
• How should model changes be controlled?
• Which data-quality limitations are acceptable?

For rapidly developing AI technologies, this iterative regulatory approach may be particularly valuable.

The Important Shift: From AI Principles to Regulatory Evidence

Over the last several years, many AI frameworks have converged around principles such as:

• human oversight;
• transparency;
• risk-based governance;
• data quality;
• lifecycle management.

Those principles are important.

But pharmaceutical regulation eventually requires something more concrete:

evidence.

The question becomes:

What evidence must a company provide to demonstrate that an AI model is sufficiently reliable for a particular regulatory purpose?

The Beyond ADMET initiative appears to be moving directly toward that problem.

This Is Closely Related to Context of Use

The acceptable level of AI uncertainty depends heavily on how the model is used.

For example, AI could be used to:

• prioritise compounds for additional evaluation;
• identify potential safety signals;
• support toxicology assessment;
• integrate multiple data sources;
• predict possible adverse effects;
• support regulatory evidence.

These applications do not carry identical consequences.

The validation strategy should therefore reflect:

Intended Use → Context of Use → Decision Impact → Required Credibility

The closer the AI output is to a consequential regulatory or safety decision, the stronger the evidence supporting its reliability is likely to need to be.

Lifecycle Control Will Also Matter

AI models do not necessarily remain static.

Performance can change because of:

• new data;
• changing populations;
• model retraining;
• algorithm updates;
• software changes;
• changes in clinical practice.

This creates another important regulatory question:

When does a change require reassessment or revalidation?

For regulated pharmaceutical AI, initial validation cannot necessarily be treated as the end of the assurance process.

A more realistic model may become:

Model Development

↓

Validation

↓

Defined Context of Use

↓

Deployment

↓

Performance Monitoring

↓

Change Assessment

↓

Revalidation Where Necessary

Why This Matters Beyond Safety Prediction

The Beyond ADMET programme is not GMP guidance.

However, the regulatory questions being investigated are closely related to questions facing AI throughout pharmaceutical development and manufacturing:

• How should AI models be validated?
• What data are acceptable?
• How should model limitations be documented?
• How should uncertainty be managed?
• How should changes be controlled?
• When can a regulator rely on AI-generated evidence?

These questions are also relevant to AI used in:

• manufacturing models;
• quality systems;
• process monitoring;
• analytical applications;
• regulatory submissions.

The specific application changes, but the underlying credibility problem remains similar.

MHRA Is Asking Industry for Practical Experience

Importantly, MHRA is not limiting participation to organisations already using advanced AI systems.

The Agency is seeking input from organisations across the medicines-development pathway, including:

• toxicology;
• preclinical research;
• clinical research;
• life sciences;
• biotechnology;
• AI technology.

Organisations can contribute experience concerning current use, future plans, barriers and unmet needs.

The survey remains open until:

22 December 2026 at 23:59 UK time.

For pharmaceutical companies developing AI-based safety models, this represents a direct opportunity to influence the regulator’s understanding of practical validation and data challenges.

Why This Development Is Important

The Beyond ADMET initiative is another indication that regulators are moving from general AI policy toward application-specific regulatory science.

The progression can be seen as:

AI Principles

↓

Specific Use Cases

↓

Regulatory Sandboxes

↓

Evidence and Validation Methods

↓

Future Regulatory Expectations

This is a significant development for pharmaceutical companies.

Future AI regulation may not consist of one universal AI validation standard.

Instead, regulators may develop different evidence expectations for different contexts of use.

Regulatory Status Note

The Beyond ADMET stakeholder survey is an MHRA call for evidence.

It does not establish a new regulatory requirement and does not represent final guidance on validation or acceptance of AI models for medicines-safety evaluation.

Its findings will inform development of the Beyond ADMET regulatory sandbox and future MHRA work in this area.

The initiative should therefore be viewed as an important regulatory-development signal rather than a current compliance requirement.

Source

MHRA – Beyond ADMET: AI for medicines safety stakeholder survey, published 16 September 2026:

https://www.gov.uk/g … y-stakeholder-survey

Introduction

⇒ FDA and industry discussions at the 2026 PDA/FDA Joint Regulatory Conference point toward risk-proportionate AI assurance in CGMP, combining data readiness, validation, human oversight, source traceability and lifecycle control.

Artificial intelligence was no longer treated as a peripheral technology topic at the 2026 PDA/FDA Joint Regulatory Conference.

The conference, held in Washington, D.C. from 14 to 16 September 2026, included dedicated discussions on AI in regulatory oversight, inspection readiness, manufacturing data and validation of digital tools in a CGMP environment.

One message stands out:

AI does not remove existing CGMP responsibilities. The level of validation and control should reflect how the AI is actually used and the impact of its output.

The official PDA/FDA programme explicitly states that appropriate controls, human oversight and validation remain foundational requirements when digital tools are introduced into CGMP processes.

This is an important direction for pharmaceutical companies because it moves the discussion beyond the simplistic question:

“Does every AI tool require the same validation?”

towards:

“What evidence and controls are necessary for this particular AI use case?”

AI-Ready Manufacturing Starts with AI-Ready Data

One session, Advancing Digital Tools and Data Utilization in a CGMP Environment, brought together FDA and industry representatives.

Sanofi presented a practical framework for assessing whether manufacturing data are sufficiently mature to support advanced analytics and AI.

The concept evaluates not only which data exist, but also how those data are made available.

Relevant considerations include:

• coverage of critical quality attributes;
• process and yield data;
• batch and sensor information;
• equipment and material genealogy;
• derived process values;
• data freshness;
• accessibility;
• authenticity;
• standardisation;
• structure.

This is an important point for AI implementation.

An organisation may have very large quantities of manufacturing data without actually having data that are suitable for regulated AI use.

AI readiness therefore cannot be measured simply by:

How much data do we have?

A more relevant question is:

Are the data complete, contextualised, reliable, accessible and sufficiently controlled to support the intended AI decision?

CGMP Principles Still Apply

FDA participation in the session focused on the relationship between digital transformation and established pharmaceutical quality requirements.

The underlying message was not that AI requires abandonment of traditional CGMP principles.

Quite the opposite.

AI-enabled processes still need to operate within established expectations relating to:

• Quality Unit oversight;
• data integrity;
• documented procedures;
• validation;
• record governance;
• Quality Risk Management;
• appropriate human control.

This reinforces an important principle for pharmaceutical AI:

The technology may be new, but accountability remains within the Pharmaceutical Quality System.

An algorithm cannot assume responsibility that CGMP assigns to the Quality Unit, system owner, process owner or qualified personnel.

A Risk-Proportionate Validation Model Is Emerging

One of the most interesting discussions concerned the level of validation appropriate for different digital and AI applications.

Conference reporting described a distinction between tools that contribute directly to final regulated decisions and tools used primarily for investigation, exploration or information support.

This suggests a potentially useful continuum.

Higher-impact application

AI directly influences or supports a final GxP decision.

Examples might include:

• product-quality decisions;
• critical process decisions;
• batch disposition support;
• critical specification or release assessments.

Such uses may require extensive CGMP validation and strong evidence demonstrating fitness for intended use.

Lower-impact or exploratory application

AI helps users investigate information or identify possible signals while a qualified person independently evaluates the evidence.

Examples might include:

• investigation support;
• regulatory intelligence;
• trend identification;
• searching quality records;
• exploring historical inspection findings.

For these applications, a different level of evidence may be appropriate, provided that the use is clearly bounded and the output cannot bypass the responsible human decision-maker.

The important point is not that lower-risk AI requires “no validation”.

Rather:

The depth of assurance should be proportionate to intended use, GxP impact and the consequence of an incorrect output.

This Is Different from a One-Size-Fits-All AI Validation SOP

Pharmaceutical companies may be tempted to create one standard validation package for every AI application.

The conference discussion suggests that this could be unnecessarily rigid.

AI applications can perform fundamentally different functions.

Compare:

• AI that searches historical deviations;
• AI that drafts an investigation summary;
• AI that recommends a possible root cause;
• AI that predicts a critical process parameter;
• AI that influences a final batch disposition decision.

Calling all of these simply “AI systems” hides important differences in risk.

A more mature approach could therefore start with:

Intended Use → GxP Impact → Decision Impact → Failure Consequence → Required Assurance

Only after those questions are answered should the validation strategy be defined.

Inspection-Ready AI Regulatory Intelligence

A separate conference session provided an interesting practical example.

Eli Lilly described an AI-supported regulatory intelligence system connecting information that historically sits in different quality repositories, including:

• regulatory inspection findings;
• internal audit observations;
• deviation information;
• responses to observations;
• after-action reviews.

The architecture combines retrieval-augmented generation for unstructured documents with structured-data exploration.

The objective is not simply to ask a chatbot a compliance question.

It is to allow users to trace quality signals across:

time → site → health authority → observation → response → organisational learning

while retaining connection to the underlying evidence.

This represents a potentially important model for pharmaceutical Quality Systems.

AI may provide the greatest value not by replacing Quality experts, but by connecting information that humans currently review in isolated systems.

Source Grounding Becomes a Critical Control

For regulatory or compliance applications, an AI answer is only useful when the user can determine where it came from.

This makes source grounding particularly important.

A useful architecture may therefore require:

• controlled source repositories;
• retrieval of relevant source documents;
• citations linked to the generated answer;
• traceability to the original record;
• defined document versions;
• human verification of the cited evidence.

This can substantially change the validation question.

Instead of asking only:

“Is the generated answer correct?”

the organisation can also ask:

“Can the user reconstruct the evidence on which the answer was based?”

For GMP applications, that second question may be equally important.

Hallucination Risk Does Not Disappear with Better Models

The conference discussion also highlighted the continuing risk of hallucination and model variability.

General-purpose AI models can differ substantially in their behaviour, and performance can deteriorate depending on question framing, context length and the information supplied to the model.

For pharmaceutical compliance use, this means that selecting a more powerful model does not eliminate the need for controls.

A robust architecture should assume that:

the AI can sometimes be wrong.

The control strategy should therefore determine:

• how incorrect outputs are detected;
• when source verification is mandatory;
• when the AI must defer;
• what requires independent human review;
• what decisions AI is not authorised to make.

Human Oversight Must Be Real, Not Formal

The official conference programme explicitly identifies human oversight as a foundational requirement.

However, simply inserting a human approval step into a workflow may not provide meaningful control.

Effective human oversight requires that the reviewer:

• understands the intended use of the AI;
• understands its known limitations;
• has access to source information;
• can challenge the AI result;
• has sufficient expertise to identify an implausible output;
• has authority to reject the recommendation;
• remains accountable for the final decision.

This distinction may become increasingly important as AI becomes embedded inside routine quality-system software.

From AI Validation to AI Assurance

The conference discussions point toward a broader model than traditional one-time software validation.

A possible lifecycle could be:

Intended Use

↓

Risk Classification

↓

Data Readiness

↓

Validation / Fitness-for-Use Evidence

↓

Controlled Deployment

↓

Human Oversight

↓

Performance Monitoring

↓

Change Control

↓

Periodic Reassessment

The exact depth of each element should depend on risk.

This is increasingly consistent with the broader concept of maintaining an AI-enabled system in a demonstrable state of control rather than merely declaring the system “validated” at implementation.

Why This Is Important

The 2026 PDA/FDA conference is significant because AI is now being discussed within mainstream CGMP topics such as:

• inspection readiness;
• data governance;
• manufacturing data;
• validation;
• quality systems;
• regulatory oversight.

AI is therefore moving from an innovation discussion into normal pharmaceutical quality-system governance.

The emerging direction appears to be:

do not validate AI because it is AI; validate and control the regulated function according to intended use, risk and decision impact.

This may ultimately provide a more practical approach than attempting to apply one universal validation standard to every AI technology.

Regulatory Status Note

The observations described above originate from presentations and panel discussions at the 2026 PDA/FDA Joint Regulatory Conference.

The conference includes FDA participation, but individual presentations, panel discussions and interpretations should not be treated as new FDA guidance or formal regulatory policy.

In particular, the concept of proportional levels of AI validation discussed during the conference should be understood as emerging regulatory and industry thinking rather than a newly issued CGMP requirement.

Sources

FDA – 2026 PDA/FDA Joint Regulatory Conference:

https://www.fda.gov/ … -conference-09142026

PDA – PDA/FDA Joint Regulatory Conference 2026, programme and session details:

https://www.pda.org/ … tory-conference-2026

BioPharm International – FDA, Sanofi, Gilead Weigh AI Readiness and Validation Risk at PDA/FDA Conference:

https://www.biopharm … k-pda-fda-conference

Pharmaceutical Technology – Lessons For AI Governance in Quality Systems:

https://www.pharmtec … e-in-quality-systems