A.B. MODI

Seven Principles for Pharma 5.0

An independent framework for human and machine collaboration in regulated drug development, aligned with the FDA and EMA Guiding Principles of Good AI Practice in Drug Development (January 2026).

Version 1.1, October 2026. Written by A.B. Modi. Not endorsed by any regulator.

The principle behind the seven

AI in pharma exists to multiply what experts can do for science and patients, with humans and machines accountable to each other through evidence.

I

Purpose

Read principle I

Select a principle on the wheel, or use Tab and Enter, to see what it commits you to.

What Pharma 5.0 means here

Pharma 4.0 digitized the work. It moved paper to systems, batch records to electronic logs and manual review to dashboards. That work is necessary and still unfinished: many of the calculations and checks that decide things still happen in spreadsheets and by hand.

Pharma 5.0 is human and machine collaboration, with accountability on both sides. Machines take on calculation, cross-checking and first drafts. Experts keep every consequential decision, and spend their judgment where it matters most.

Pharma 4.0 asked whether machines can talk to each other. Pharma 5.0 asks whether humans and machines can work with each other.

The name borrows from Industry 5.0, which the European Commission described in 2021 as human-centric, sustainable and resilient. Pharma 5.0 is not a regulatory term. In this framework it names a direction: the point where the people who know the work can build for it, on a platform that makes what they build defensible.

Why these principles, and why now

Since 2024, practitioners in industry forums have asked the same question: what evidence is enough to trust an AI system in GxP work? Two years on, after the FDA and EMA principles, draft EU GMP Annex 22 and the first FDA warning letter to cite a firm's use of AI agents to write its procedures and master production records, it is still the open question. Guidance describes what good looks like. It does not yet say how to get there, or how to prove it.

These principles were tested against thirteen concerns that came up again and again in those discussions. Select a concern to see how the principles answer it.

Safe, and transformative

The FDA and EMA ten principles describe how to make AI safe across the drug product life cycle. These seven build on them to make Pharma 5.0 happen. Every regulator principle maps into the seven. Purpose, Partnership, People, Capability and Resilience carry what the regulators leave to industry.

FDA and EMA guiding principles

January 2026

Seven Principles for Pharma 5.0

This framework

Select a principle on either side to see where it connects.

How to read each principle

Each principle is a single commitment, followed by the problem it solves and the practices that make it concrete. Every practice carries a basis tag, so you can see what rests on regulation and what is a proposal.

Each principle is written from the side of the partnership. Control is the discipline that makes it trustworthy, not the point of it.

  • Regulator-derived grounded in guidance, regulation or enforcement
  • Established practice widely used in GxP or engineering
  • Emerging practice in use, not yet standard
  • Author's position my proposal, open to challenge

Purpose

AI exists to lift what experts can do, and we measure that lift against how the work is done today.

The problem it solves. AI adopted for novelty, pilots that never scale, and two opposite failures: quality teams who assume any AI means an audit finding, and teams who adopt AI without understanding its limits or their own responsibility.

  • Name what the expert and the machine will do together that the expert could not do alone.Author's position
  • Use AI where work is repetitive, high-volume and error-prone, and name the decision it supports.Established practice
  • Define what a good output is, and record the expert baseline, before building anything.Established practice
  • Measure success as expert time redirected to judgment and errors reduced, not minutes saved.Author's position
  • Run the system in shadow mode, alongside the existing process, before anyone relies on it.Established practice
  • Start with one bounded workflow and widen only on evidence.Established practice

FDA and EMA principles: 1 Human-centric by design4 Clear context of use

Partnership

Humans and machines each do what the other cannot, and answer to each other for it.

The problem it solves. Automation bias, review that becomes a rubber stamp, and attempts to hand accountability to the tool. And a one-way relationship, in which only the machine is held to account. Fluent AI output disarms reviewers, so stacked reviews tend to fail together rather than independently.

  • The human holds the decision and the signature. Accountability for a GxP record cannot be delegated to an AI.Regulator-derived
  • The machine signals uncertainty only where it can measure it, and escalates or abstains rather than guesses.Established practice
  • The expert gives the machine clear context and corrects it honestly, through the controlled update route.Author's position
  • The machine's reasoning is open to inspection, so the expert can challenge it rather than only accept or reject it.Established practice
  • Autonomy is granted in steps, each step earned with evidence.Established practice
  • Review depth follows risk, with anomaly flags directing attention instead of spreading it evenly.Established practice
  • Record where AI contributed to a document or decision, so human ownership stays honest.Author's position
  • Human corrections feed controlled, versioned updates. Models do not learn live in critical GMP use.Regulator-derived

FDA and EMA principles: 1 Human-centric by design8 Risk-based performance assessment

Integrity

Build the machine so its core cannot be corrupted, and so its work can always be questioned.

The problem it solves. Non-deterministic output, errors that look right, and attacks that work through content rather than networks. A property enforced by design is verified once; a property left to model behaviour has to be proven statistically, again and again.

The two-layer architecture An agent layer plans, retrieves and drafts. It invokes a deterministic core, which returns every value of record with its source. The agent reads systems of record but cannot write to them. It passes a labelled draft to the expert, and only the expert approves and writes to the record. Expert decides and signs Agent layer probabilistic plans, retrieves, drafts Deterministic core computes every value of record refuses inputs out of range Systems of record LIMS, QMS, ELN invokes returns value with its source labelled draft read only only the expert approves and writes to the record
The two-layer architecture. The agent plans and drafts, and every value of record comes from the deterministic core it invokes. The agent can read systems of record but never write to them.

Containment

  • A deterministic core computes every value of record. The agent invokes it and never replicates the calculation.Author's position
  • Actions the agent must never take are blocked by least-privilege permissions. The agent is read-only to systems of record.Established practice
  • Least privilege also covers what the AI reads. The reference corpus is qualified from ingestion onward and version-controlled, regulations included.Established practice
  • Each agent has its own role-bound identity and credentials, never a borrowed human account, so every action is attributable.Established practice
  • Retrieved documents and tool outputs are treated as untrusted data, never as instructions.Established practice
  • Output is validated before it reaches another system. Retrieval enforces the permissions of the person asking. Secrets and controls never depend on the system prompt.Established practice
  • Security is part of performance: prompt injection and poisoned retrieval are correctness risks, tested alongside accuracy.Established practice
  • AI trace logs are kept distinct from the GMP audit trail of the record they informed.Author's position

Transparency

  • The support for every claim is verified, not just the existence of its source. Simulated output carries labels that cannot be removed.Established practice
  • Make the work machine-ready: facts as fields, decisions as closed lists, judgment labelled as judgment.Author's position

FDA and EMA principles: 3 Adherence to standards6 Data governance and documentation7 Model design and development practices

Evidence

Prove that the human and the machine together get it right, and keep proving it.

The problem it solves. No agreed standard for what evidence is enough, no shared metrics, and control applied either nowhere or everywhere. An accurate model paired with an uncritical reviewer performs worse than it looks on paper. You cannot make an agent deterministic, but you can make most of its evidence deterministic.

Four kinds of evidence, scaled to risk

Choose a risk band to see how much of each kind of evidence it calls for.

Bands follow the Risk Profile Index (RPI), a proposed scoring aid that extends the influence and consequence model in FDA's draft credibility framework with autonomy, non-determinism, reversibility and detectability. It has not yet been validated across a broad set of systems.

  • Write the context of use. The burden attaches at reliance, and output that steers a GxP judgment counts as relied upon.Regulator-derived
  • Prove oversight works by seeding known defects into review, segregated so they never reach a real record, and tracking catch rates over time.Emerging practice
  • Treat risk as graded, not binary, and align risk bands with recognized model categorisation.Regulator-derived
  • Be proportionate in both directions: control the new risks AI introduces, use qualified vendor evidence, and add nothing where no new risk exists.Established practice
  • Keep a frozen reference set: expert-adjudicated, with known negatives and repeated runs, against acceptance criteria set in advance and at least as good as the baseline.Established practice
  • Include prompt injection and red-team cases in the reference set, rerun them on every relevant change, and monitor live use for attack patterns.Established practice
  • Design monitoring up front: metrics for each layer, tolerance limits that trigger requalification, named owners, and risk-based retest triggers.Regulator-derived
  • State purpose, performance and limits in plain language at the point of use, where the user acts.Regulator-derived

FDA and EMA principles: 2 Risk-based approach3 Adherence to standards4 Clear context of use8 Risk-based performance assessment9 Life cycle management10 Clear, essential information

People

Grow the people who can both use the machine and judge it.

The problem it solves. Only qualified experts can tell whether an AI is right, and they are scarce, senior and stretched. As AI produces more of the work, the human baseline erodes, and training alone cannot create an expert overnight.

  • Bring together a multidisciplinary team with real authority to stop the work, not just sign it.Regulator-derived
  • Qualify reviewers specifically for reviewing AI output, and document their competence.Emerging practice
  • Train people to work with AI and to question it, not only to supervise it.Author's position
  • Spread judgment across more people, so oversight does not rest on a few senior heads.Author's position
  • Preserve manual expertise deliberately, and grow new experts on purpose.Author's position
  • Redesign roles honestly. Reviewing AI is real work, not a side duty.Author's position
  • Train everyone on what must never go into an unapproved AI tool. It is a common route for confidential data to leave an organisation.Established practice
  • Monitor human competence over time, alongside the system.Emerging practice

FDA and EMA principles: 5 Multidisciplinary expertise

Capability

Let those who know the work build for it, on a platform that keeps it defensible.

The problem it solves. IT backlogs, shadow tools and an intake bottleneck. Anyone can now build a working tool with an AI assistant in days. The risk is not that AI wrote the code; it is that a personal tool quietly becomes one that a GMP decision depends on.

The reliance gate On the left, people build and draft freely. In the middle, a reliance gate is passed when output enters a GMP record, a quality review or a decision, detected at data access and in record templates. On the right, relied-upon tools are registered and assured with evidence scaled to risk. Build and draft freely personal tools, prototypes, first drafts Reliance gate output enters a GMP record, a quality review or a decision detected at data access and in record templates Registered and assured context of use, risk band, evidence scaled to risk
Gate reliance, not building. Tools built on the governed platform inherit its controls, so at the gate they prove only what they add.
  • Fit the tool to the work, not the work to the tool. An organisation's procedures should shape its systems.Author's position
  • Specify first: write the intended use and requirements before AI writes any code, so evidence accumulates during the build.Established practice
  • Gate reliance, not building, and keep an inventory of relied-upon tools with named owners.Established practice
  • The platform carries the controls. Each tool built on it proves only what it adds.Established practice
  • The platform supplies the security baseline every tool inherits: an information security management system aligned with ISO/IEC 27001 or the NIST Cybersecurity Framework.Established practice
  • The platform sets limits on use and controls what can be installed, so cost, capacity and dependencies stay governed.Established practice
  • Distinguish tools built with AI, which are ordinary software, from tools that contain AI in a critical function. They need different controls.Established practice
  • Share what works, and publish measured results rather than claims.Author's position

FDA and EMA principles: 3 Adherence to standards5 Multidisciplinary expertise

Resilience

Stay in control of everything the partnership depends on.

The problem it solves. Confidential data reaching places it should not, vendor-side model changes, deprecations and outages, evidence that vendors cannot supply, and regulation that is still moving. A system qualified today can stop being fit for purpose without any change on your side.

  • Own your data and your models where it matters. At minimum, control the model version and prove it.Author's position
  • Inventory the whole harness, and put prompts, retrieval indexes and tool definitions under change control.Established practice
  • Data classification decides which model and which host each kind of data may reach. Vendor terms on training, retention and location are verified, not assumed.Established practice
  • Model weights, licences and software dependencies have verified provenance.Established practice
  • Know which model produced each output.Established practice
  • Every agent has a kill switch and a tested fallback, so it can be stopped without stopping the work.Established practice
  • Keep the reference set as a portable regression test, with an exit plan and a manual fallback.Established practice
  • Build so that compliance holds however draft Annex 22 is finalised.Author's position
  • Require vendors to supply evidence a regulated customer can actually use.Established practice
  • Right-size models to the task. The smallest adequate model is easier to test, explain, host and afford.Author's position

FDA and EMA principles: 6 Data governance and documentation9 Life cycle management

Security across the seven

Security is not an eighth principle, because for AI it cannot be separated from the other seven. Content an agent reads can change what it does, without any breach of a network control. That makes security part of correctness, of control and of the platform itself.

Select a threat to see which principles defend against it. Threats are aligned with the OWASP Top 10 for Large Language Model Applications where a category exists.

From principles to practice

The principles state what must be true. The practices make each one concrete. The how comes next, as a set of frameworks now in development, in the order a team would use them:

  1. Self-assessment. Where your organisation stands against the seven.
  2. Adoption pathway. The route from first pilot to governed platform.
  3. Use-case triage. Which work should, and should not, use AI.
  4. Reference architecture. The two-layer design in detail.
  5. Risk classification. The Risk Profile Index, with worked examples.
  6. Validation and evidence lifecycle. The four kinds of evidence, stage by stage.
  7. Roles and governance. Who owns what, from builder to quality unit.

Shared language

Data scientists, IT and quality teams often use the same words to mean different things. These are the meanings used here.

Context of use
The role and scope of an AI system: what it does, for whom, with what authority, and what it must not be used for.
Reliance
The point at which an output enters a GxP record, a quality review or a decision, or materially steers one.
Deterministic core
Software that gives the same output from the same input, and refuses inputs outside its validated range.
Harness
Everything around a model that shapes its behaviour: prompts, retrieval sources, tool definitions and configuration.
Reference set
A frozen, expert-adjudicated set of cases with known answers, including cases where the right answer is that nothing is wrong, used to measure performance before and after every change.
Seeded defect
A known error placed deliberately in a review stream, tagged so it can never reach a real record, to measure whether reviewers catch it.
Prompt injection
Instructions hidden in content an AI reads, such as a document, email or web page, that try to change what it does.
Misgrounding
Citing a real source for a claim that the source does not support.
Abstention
The system declining to answer and routing to a person when it cannot support an answer.
Validation
In GxP, documented evidence that a system is fit for its intended use. In data science, the same word names a dataset used while training a model. The two meanings should not be confused.
Critical use
Draft EU GMP Annex 22 separates critical GMP applications, with direct impact on patient safety, product quality or data integrity, from non-critical ones. These principles treat risk as graded within that distinction.

Status and independence

These principles are an independent proposal by A.B. Modi. They are not endorsed by the FDA, the EMA or any other regulator or standards body, and they are not a validation standard. They draw on public regulatory documents, published research and summarised practitioner discussion.

EU GMP Annex 22 was still in draft at the time of writing, and the FDA guidance on AI for regulatory decision-making was still a draft. Practices tagged as regulator-derived should be rechecked against final texts as they are published.

This is version 1.1. It will change as regulation, evidence and practice move. Challenges are as welcome as agreement.

What changed in 1.1. Version 1.0 framed each principle around controlling the machine. Version 1.1 reframes them around the partnership between human and machine. The practices are largely unchanged.

How this was made. The research, conversations and decisions behind these principles are the author's. A large language model was used to help synthesise the material and test the reasoning.

Send a comment

References

  1. U.S. Food and Drug Administration and European Medicines Agency. Guiding Principles of Good AI Practice in Drug Development. January 2026.
  2. European Commission. EudraLex Volume 4, Annex 22: Artificial Intelligence. Draft for consultation, 7 July 2025.
  3. European Commission. EudraLex Volume 4, Annex 11: Computerised Systems. Revised draft for consultation, 7 July 2025.
  4. European Commission. EudraLex Volume 4, Chapter 4: Documentation. Revised draft for consultation, 7 July 2025.
  5. U.S. Food and Drug Administration. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. Draft guidance, January 2025.
  6. U.S. Food and Drug Administration. Computer Software Assurance for Production and Quality System Software. Final guidance, September 2025.
  7. U.S. Food and Drug Administration. Warning Letter to Purolea Cosmetics Lab. 2 April 2026.
  8. U.S. Code of Federal Regulations. Title 21, Part 11 (Electronic Records; Electronic Signatures) and Part 211.22 (Responsibilities of the quality control unit).
  9. International Council for Harmonisation. ICH Q9(R1): Quality Risk Management. 2023.
  10. International Council for Harmonisation. Points to Consider for ICH Q8/Q9/Q10 Implementation (R2). 6 December 2011.
  11. International Society for Pharmaceutical Engineering. GAMP 5: A Risk-Based Approach to Compliant GxP Computerized Systems, Second Edition. 2022.
  12. International Society for Pharmaceutical Engineering. Comments to the European Commission on draft EU GMP Chapter 4, Annex 11 and Annex 22. 7 October 2025.
  13. European Commission. Industry 5.0: Towards a sustainable, human-centric and resilient European industry. January 2021.
  14. OWASP GenAI Security Project. OWASP Top 10 for Large Language Model Applications, 2025 edition (v2.0). Published November 2024.
  15. International Organization for Standardization and International Electrotechnical Commission. ISO/IEC 27001:2022: Information security, cybersecurity and privacy protection. Information security management systems. Requirements.
  16. National Institute of Standards and Technology. The NIST Cybersecurity Framework (CSF) 2.0. 2024.
  17. Parasuraman R, Riley V. Humans and automation: use, misuse, disuse, abuse. Human Factors. 1997;39(2):230–253.
  18. Practitioner discussions in the ISPE Engage Artificial Intelligence Community of Practice, 2024 to 2026. Summarised; individual contributions are not quoted.