The Latest FDA Guidance on AI in Drug Manufacturing: April-May 2026 Update

The Current FDA AI Landscape for Manufacturing

For most of the past decade, FDA’s published positions on AI in pharma have been fragmented: GMLP principles for software-as-medical-device in 2021, discussion papers on drug development and drug manufacturing in May 2023, and a steady cadence of public engagements through CDER and CBER. As of May 2026, the picture is materially more coherent. Three published documents now constitute the operational reference set for AI in drug manufacturing.

The first is the January 6, 2025 draft guidance titled Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, which provides a risk-based credibility assessment framework that may be used for establishing and evaluating the credibility of an AI model for a particular context of use, as described in the FDA’s January 2025 press announcement. The 90-day comment window closed April 7, 2025, with finalization expected during 2026.

The second is the December 4, 2024 finalization of Predetermined Change Control Plans for Machine Learning-Enabled Medical Devices. Although scoped to SaMD, the PCCP construct articulates a regulatory framework for managing model lifecycle changes that has become a reference point for pharma quality leaders thinking about how to manage vendor-driven AI updates in regulated workflows.

The third is the March 2026 proposal from CDER for new ICH guideline work to facilitate the adoption of advanced pharmaceutical manufacturing, including AI. This signals where the FDA’s manufacturing posture is heading: toward international convergence with EMA, MHRA, PMDA, and other regulators on a harmonized framework rather than a US-only approach.

Quality leaders who have been tracking the AI regulatory landscape for years will recognize the pattern: the agency does not legislate in a single comprehensive document. Instead, it publishes principles, discussion papers, and targeted guidances that, taken together, define operational expectations. The 2024-2026 cycle has produced enough material that “we’re waiting for FDA to clarify” is no longer a defensible posture for active manufacturing AI deployments.

The 7-Step Credibility Framework Explained

The credibility framework is the single most consequential conceptual contribution of the 2025 draft guidance. Rather than prescribing universal validation rules, it asks sponsors to assess and document the credibility of an AI model in a specific context of use. The seven steps, as described by DLA Piper’s January 2025 analysis of the guidance, run from defining the question of interest through documenting the activities undertaken to establish credibility.

The seven steps, in order:

1. Define the question of interest. The specific decision the AI is being used to support. A model that flags batch records for human review is a different question of interest than a model that releases batches autonomously.
2. Define the context of use. Where in the manufacturing or quality workflow the AI operates, what its inputs and outputs are, and how its output is consumed downstream.
3. Assess the model risk. A combined function of model influence and decision consequence. High influence plus high consequence drives Tier 3 expectations; low influence plus low consequence permits a lighter posture.
4. Plan credibility activities. The specific evidence the sponsor will generate to establish credibility, calibrated to the model risk.
5. Execute the credibility plan. The actual validation, testing, performance assessment, and documentation work.
6. Document credibility evidence. Producing artifacts that a sponsor, an internal QA function, or an FDA reviewer can use to assess whether the model meets the bar for its context of use.
7. Determine adequacy and use. A formal go/no-go decision on whether the model has been established as credible for the proposed use.

The framework’s most important practical feature is that it is not a checklist. Sponsors must exercise judgment about what credibility activities are appropriate for the specific question of interest and context of use. As a FDA guidance landing page indicates, this approach is being extended across the drug product lifecycle, including manufacturing-adjacent applications where AI affects safety, efficacy, or quality.

The scope clause has been widely discussed. The draft guidance does not address AI models used in drug discovery, or when used to streamline operations such as drafting regulatory submissions, in cases where those uses do not impact patient safety, drug quality, or the reliability of results from nonclinical or clinical studies. In practice, manufacturing applications generally do fall within scope because AI used to support manufacturing decisions almost always touches quality, even when the use case looks operational on its surface.

PCCP Finalization and Its Manufacturing Implications

The December 4, 2024 finalization of the PCCP guidance for AI/ML-enabled devices was technically a SaMD action, but its conceptual reach extends into pharma manufacturing. A PCCP allows manufacturers to pre-define specific AI/ML software changes and the methods to develop, validate, implement, and monitor them, so those updates can be made without a new marketing submission when executed exactly as an authorized PCCP. The August 2025 joint document with Health Canada and the UK MHRA articulated five principles for PCCPs covering focus, risk-basis, evidence requirements, transparency, and lifecycle governance.

For pharma manufacturing quality leaders, the PCCP framework matters for two reasons. First, it provides a reference model for thinking about vendor-driven AI changes in regulated workflows. The structure of a PCCP — description of modifications, modification protocol, impact assessment — can be adapted to internal change-control SOPs for AI components in manufacturing systems, even when the regulatory pathway is not a 510(k). Second, the PCCP construct signals the FDA’s broader appetite for change-control mechanisms that accommodate the iterative nature of AI without requiring a regulatory action for every model update. As the agency’s PCCP guiding principles page indicates, the principles are being designed to scale beyond their initial SaMD application.

Manufacturing quality teams that adopt PCCP-like thinking for internal AI change control reduce two real risks. First, they avoid the trap of treating every model update as a deviation, which exhausts QA capacity. Second, they avoid the opposite trap of allowing model updates to flow through without documented assessment, which is precisely the gap inspectors probe.

The five PCCP principles in practical terms

The five principles articulated in the August 2025 international document — focused, risk-based, evidence-based, transparent, and lifecycle-aware — translate into operational language pharma manufacturers can use:

• Focused means the change scope is defined narrowly enough that a reviewer can determine in advance whether a proposed change falls within the plan.
• Risk-based means the depth of validation evidence is calibrated to the consequence of the change.
• Evidence-based means the plan specifies what data will support the change, not just that data will be collected.
• Transparent means the plan is documented in a way that internal QA, external regulators, and downstream consumers of the AI output can reason about.
• Lifecycle-aware means the plan does not assume static models, but rather articulates how the changes will be governed over time.

CDER’s FRAME Initiative and the ICH Bridge

CDER’s Framework for Regulatory Advanced Manufacturing Evaluation (FRAME) Initiative provides the operational scaffolding through which the agency evaluates advanced manufacturing technologies, including AI. The FRAME initiative was launched to articulate how CDER reviews emerging manufacturing approaches that do not fit neatly into existing CMC review categories.

The March 2026 CDER proposal for new ICH guideline work to facilitate the adoption of advanced pharmaceutical manufacturing connects the FRAME initiative to international harmonization. By proposing ICH-level work, CDER signals that the AI-in-manufacturing direction is being pursued in coordination with the EMA, MHRA, PMDA, and other regulators rather than as a US-only effort. For multinational manufacturers, this is materially important: a US-only framework would mean parallel compliance work in each jurisdiction, while an ICH-harmonized framework permits a single set of practices to satisfy multiple regulators.

The relationship between FRAME, the credibility framework, and ICH M15 (the model-informed drug development guideline that reached step 5 in early 2026) deserves to be understood as a single coherent direction. As EMA’s ICH M15 scientific guideline page indicates, ICH M15 explicitly recognizes AI/ML as a category of computational modeling methods that fall within MIDD. The convergence between FDA’s domestic guidance and ICH harmonization work is not accidental; it reflects deliberate coordination by the agencies involved.

What Actually Changed in April-May 2026

The April-May 2026 window did not produce a single dramatic new document. What it did produce was a recognizable consolidation of the direction that has been emerging since late 2024:

Practical Actions for Pharma Quality Leaders

For pharma manufacturing quality leaders, the operationally relevant actions in the next two quarters fall into five buckets.

1. Adopt the credibility framework language in QMS documentation

Even before the January 2025 guidance is finalized, the seven-step framework provides defensible language for AI use case validation in QMS documentation. Tier classification SOPs can map directly onto the model risk step. Validation protocols can articulate credibility activities. Adopting this language now produces documentation that ages well as the guidance finalizes and that reads as forward-aware to inspectors.

2. Build PCCP-style change control for AI components in manufacturing

Internal change control SOPs for AI components should articulate, in advance, the categories of model change that will occur, the validation that will accompany each, and the conditions under which changes require revalidation versus tracked acknowledgment. This is the PCCP construct applied internally, regardless of whether the AI components are part of a regulated submission. The discipline is the same; the regulatory framing differs.

3. Inventory vendor-embedded AI in manufacturing systems

AI capabilities embedded in MES, LIMS, EMS, and other manufacturing platforms — by Siemens, Rockwell, Honeywell, AVEVA, and others — are often active without being explicitly classified as AI use cases. The credibility framework applies to these no less than to standalone AI deployments. Inventorying these features and bringing them into the framework is now a baseline expectation, not a forward-looking optimization.

4. Engage CDER or CBER early for material AI deployments

For Tier 2 and Tier 3 manufacturing AI deployments, early engagement with the FDA’s Emerging Technology Program (ETP) or Center for Advanced Tissue Therapies (CATT) is highly encouraged before submitting a regulatory application or implementing AI technology. The agency has been explicit that pre-submission meetings on AI components produce smoother review pathways than post-hoc disclosure.

5. Align with the ICH direction even before ICH M15 finalizes

ICH M15 reached step 5 in early 2026 and provides a harmonized vocabulary for model-informed approaches. Manufacturing quality leaders whose AI use cases reach across jurisdictions should align documentation language with ICH M15 terminology, as confirmed in the EMA scientific guideline page for ICH M15. This produces documentation that satisfies FDA, EMA, and PMDA reviewers from a common base.

What to Watch Next

Three near-term signals will shape the next twelve months.

First, finalization of the January 2025 draft guidance. The April 2025 comment closure produced substantive industry feedback. Finalization is expected during 2026 and will solidify the credibility framework as enforceable rather than draft expectation.

Second, EMA Annex 22 finalization. The EMA’s draft Annex 22 (the first comprehensive EU regulatory framework for AI in pharmaceutical manufacturing) was published for stakeholder consultation on July 7, 2025, with consultation closing October 7, 2025. Finalization is expected during 2026. The interaction between Annex 22 and FDA’s domestic guidance will determine the practical operating envelope for transnational manufacturers.

Third, the ICH M15 implementation cadence. As ICH M15 finalizes, the FDA, EMA, PMDA, and other agencies will begin issuing regional implementation guidance. The cumulative effect will be a more coherent international framework for AI in regulated pharma activities — including manufacturing — than the field has had at any point in the last five years.

Fourth, the cadence of CDER’s external engagements with industry on AI in drug development. The agency’s external engagements page documents the agency’s pattern of public engagement on AI topics. These engagements often signal forthcoming guidance direction six to twelve months before formal publication, and quality leaders monitoring them can anticipate finalization patterns with reasonable confidence. The pattern of engagement through 2025 and into 2026 has consistently emphasized risk-based, lifecycle-aware approaches to AI use in regulated workflows.

Fifth, the operational pattern that emerges from sponsor pre-submission meetings on AI components. While the specifics of individual meetings are not public, the aggregate pattern of how the FDA is interpreting the credibility framework in real review contexts is increasingly discussed at industry conferences. Quality leaders engaging with PDA, ISPE, and regulatory affairs forums can extract significant operational signal from how peer sponsors describe their pre-submission experiences. The aggregate signal is more useful than any single anecdote and helps calibrate the framework’s practical edges.

The strategic implication for pharma quality leaders is straightforward. The framework is not yet final, but it is recognizable. Organizations that act on the recognizable framework now will be better positioned than organizations that wait for the final form. The cost of operationalizing the current direction is materially lower than the cost of catching up under inspection pressure.

A subtler but equally important strategic implication: the framework will continue to evolve. Quality teams that build their QMS extensions as a one-time project, calibrated to the specific December 2024 – May 2026 documents, will face mounting alignment debt as the framework matures. Teams that build their extensions with an explicit assumption that the framework will continue to evolve — and that design their documentation, governance, and training programs to accommodate that evolution — will age much better. The framework is not a destination; it is a direction of travel that will produce additional documents over the next three to five years. The organizational posture should match.

The investment in this work, finally, has compounding returns. The first Tier 3 AI use case to move through the framework is the most expensive; subsequent use cases ride on the same infrastructure, the same SOPs, the same governance, the same documentation patterns. Programs that frame the initial investment as overhead for the first use case alone systematically underinvest. Programs that frame it as foundation for the AI portfolio as a whole calibrate the investment correctly and capture the leverage that the framework was designed to produce. This portfolio framing is the right way to make the case to the board, to QA leadership, and to the cross-functional steering committee that will own the framework’s ongoing maturation.

References & Sources