AI in Pharmaceutical Manufacturing: Preparing for Two FDA Guidances on CDER’s Agenda

What CDER’s 2026 Guidance Agenda Lists

Most commentary on this topic starts with a line like “FDA is about to issue guidance on AI in manufacturing.” That sentence is half right. The best way to see which half is to read the agenda itself, not a summary of it.

CDER posts a guidance agenda each year under FDA’s good guidance practices regulation. The current version lists the new and revised draft guidances CDER plans to publish in calendar year 2026, and it is dated July 2026.¹ It replaced the February 2026 posting, which RAPS reported on February 25, 2026.³

The Two Entries That Matter for Manufacturing

Two entries in the July 2026 agenda deal directly with AI in drug manufacturing:¹

So the premise that there is “an FDA guidance on the CDER agenda” about AI in manufacturing undercounts it. There are two planned documents, and they are likely to do different jobs. The first is grouped with the pharmaceutical quality and CMC topics, next to items on stability, container closure, and distributed manufacturing. That placement suggests it will speak to how AI affects product quality, control strategy, and what goes into an application. The second carries the name of an existing validation approach, computer software assurance (CSA), and calls itself supplemental. That points toward how to establish confidence in AI-based software itself. Beyond those titles and categories, nothing about their content is public.

What the Agenda Does Not Promise

The agenda’s first footnote is plain: “CDER is not bound by this list of topics nor required to issue every guidance document on this list.”¹ A note at the end adds that agenda items reflect guidances under development as of the posting date. Nothing on the list has a target month, and topics can sit on agendas for more than one year.

The title of the agenda also matters. It lists new and revised draft guidances. When either AI document is published, it will almost certainly be a draft, marked “not for implementation,” with a comment period. The final version would come later, possibly with changes. Anyone planning a program around a date for “the FDA rules on AI in manufacturing” should plan around a draft first and a final version on an unknown timeline after that.

Published Versus Planned: A Status Check

The planned guidances do not arrive in empty space. FDA has published a string of documents on AI and on models in manufacturing since 2023. They carry very different weight. Some are regulations, some are final guidance for another product area, some are drafts, and some are discussion material the agency says is not guidance at all. Treating them as equal is the most common mistake we see in readiness plans.

The Credibility Draft Is Still a Draft

The brief for this article asked us to confirm the status of the January 2025 draft guidance on AI to support regulatory decision-making. FDA’s guidance page still labels it a draft, level 1 guidance, issued January 2025.¹² The Federal Register notice of availability ran on January 7, 2025, and set a comment deadline of April 7, 2025.¹⁴ FDA’s own strategy document on innovative manufacturing technologies lists it, together with the 211.110 draft, among the draft guidances already published to clarify areas of regulatory uncertainty.²⁶ It has not been finalized.

This matters for manufacturing because the credibility draft is not only about clinical trials. Its scope covers AI used to produce information or data supporting regulatory decisions on safety, effectiveness, or quality, across the drug product life cycle, including manufacturing.¹³ It excludes drug discovery and AI used for operational efficiency that does not affect patient safety, drug quality, or study reliability. We covered the draft’s seven steps in detail in an earlier article, linked in the reading list below, so we will focus here on the parts that speak directly to manufacturing.

The Device CSA Guidance Is Final, but It Is Not a Drug Guidance

The September 2025 CSA guidance is final. Its Federal Register notice describes it as recommendations for “computers and automated data processing systems used as part of medical device production or the quality system,” and it supplements FDA’s older General Principles of Software Validation guidance.¹⁸ The listed contacts are in CDRH and CBER. Many drug manufacturers already use its risk-based thinking in their computerized system validation programs, and that is a reasonable choice. It is not, however, a CDER guidance for drug CGMP, and it does not address AI specifically.

That gap is what the planned “supplemental” CSA guidance for AI-based systems appears designed to fill. The title names drug manufacturing and clinical investigations, and the word “supplemental” suggests it will build on existing CSA guidance. We do not know what it will say, and this article will not pretend to.

What the 2023 Discussion Paper and Its Comments Tell You

The most direct evidence of what CDER is working on is the discussion paper it published in March 2023, “Artificial Intelligence in Drug Manufacturing.” It opens with a disclaimer worth reading exactly: “This paper is for discussion purposes only and is not a draft or final guidance.”⁵ It was a request for input, not a statement of policy. But it shows which questions CDER’s scientific and policy staff considered open after reviewing the existing regulatory framework.

The Five Areas CDER Flagged

The paper identified five areas of consideration where CDER wanted public feedback:⁵

1. Cloud applications and oversight of manufacturing data and records. When model updates, diagnostics, and monitoring analytics run in the cloud or with a third party, existing quality agreements may not cover the risks of AI in monitoring and control, and inspectors may find traceability harder to establish.
2. Data volume from connected equipment. Networked sensors and equipment can generate far more data per batch. Manufacturers may need clarity on which data must be stored or reviewed, and on sampling rates and compression that still leave an accurate record.
3. Which AI applications fall under regulatory oversight. AI used for equipment maintenance, continuous improvement, scheduling, or raw material characterization may or may not touch CGMP or application requirements, and companies need to know where the line is.
4. Standards for developing and validating AI models used for process control and to support release testing. The paper noted limited industry standards and FDA guidance for models that affect product quality, and raised bias, transfer learning, and explainability.
5. Continuously learning systems. The paper asked when an AI model can be treated as an established condition of a process, what changes require notifying FDA, and how inspectors would examine a model that updates itself.

For each area, the paper listed the regulations and guidance it saw as potentially relevant. Those lists are useful on their own. They include 21 CFR Part 11 and sections 211.68, 211.110, 211.165, and 211.180 of the drug CGMP regulations, ICH Q8(R2), Q9, Q10, Q12, and Q14, and FDA’s 2011 process validation guidance.⁵ In other words, CDER did not start from the idea that AI needs a new rulebook. It started from the rules already on the books and asked where they get hard to apply.

How Long the Conversation Has Run

The Federal Register notice announcing the paper set a comment deadline of May 1, 2023.⁶ FDA reopened the comment period in September 2023 “to update comments and to receive any new information,” with a new deadline of November 27, 2023.⁷ In between, FDA and the Product Quality Research Institute held a two-day virtual workshop on September 26 and 27, 2023, on the regulatory framework for AI in pharmaceutical manufacturing.⁸

So the first planned CDER guidance on this topic will arrive at least three years after the questions were asked. That is not unusual for FDA, but it is a reminder that manufacturers who waited for guidance before starting any AI governance work have already waited a long time.

What Commenters Asked For

Individual comment letters show the range of requests. The Society of Quality Assurance, commenting on May 1, 2023, answered CDER’s question about whether guidance would help with a direct yes: “Guidance in using AI in drug manufacturing would be beneficial.”¹⁰ Its list of topics included model development, validation, and maintenance; how to classify and validate a mix of low, medium, and high impact models in one workflow; data selection and splitting; model logging and audit trail expectations; monitoring and change control; and how to manage cloud providers that store data or AI systems. SQA also recommended that generative AI outputs be verified by humans. The Biotechnology Innovation Organization filed comments the same day, stating its support for FDA’s effort to improve future guidance in this area.¹¹

The most complete view comes from a May 2025 commentary in AAPS Open, written by CDER and CBER staff, that summarizes feedback from both the discussion paper docket and the workshop.⁹ The authors are careful to say the paper reflects their summary of public feedback and should not be read as FDA’s views or policies. With that caveat, its findings are the clearest picture of what industry told FDA it needs. Interested parties:

• value good data management practices;
• want best practices for developing, validating, and maintaining AI models;
• may face uncertainty when managing AI models supplied by third parties; and
• find it hard to fit AI into the pharmaceutical quality system (PQS).

The summary also records specific points that are likely to shape any guidance. Commenters distinguished between upstream process control uses, which they saw as typically lower risk, and downstream quality assurance decisions, which they saw as higher risk. They suggested hybrid models that adapt only within set boundaries as a way to manage self-learning systems. They raised expectations for product comparability after model updates, especially for biologics. They asked for international harmonization of CGMP expectations for AI so that inspections are more predictable. And some raised the use of generative AI for manufacturing documents, such as deviation reports, with human oversight.⁹

The workshop registration data in the same paper gives a sense of scale. Approximately one-third of attendees reported they were already using AI for data analysis, process development, quality assessment, or process monitoring. Thirty-one registrants planned to submit an application for a product manufactured using AI, and about 51 participants planned to engage FDA about manufacturing a product using AI within the next five years.⁹

The Rules That Already Apply to AI on the Manufacturing Floor

The planned guidances will add FDA’s current thinking. They will not create the obligation to control AI that affects product quality. That obligation already exists, and FDA has started enforcing it.

CGMP Does Not Wait for Guidance

Section 501(a)(2)(B) of the Federal Food, Drug, and Cosmetic Act and the drug CGMP regulations in 21 CFR Parts 210 and 211 apply to the methods, facilities, and controls used to make a drug. None of them has an exception for software that learns. The draft credibility guidance says so directly in a footnote: use of AI in manufacturing, for example in production and process controls, must be carried out in line with CGMP, and the quality control unit’s responsibilities under 21 CFR 211.22 and 211.68 apply.¹³

In practice, a few sections carry most of the weight for AI:

• 21 CFR 211.22 gives the quality control unit responsibility for approving or rejecting all procedures and specifications that affect the identity, strength, quality, and purity of the drug product.²³ If AI drafts or changes such documents, the quality unit still owns the approval.
• 21 CFR 211.68 covers automatic, mechanical, and electronic equipment, including computers, and requires appropriate controls and checks. AI-based systems used in production or the quality system are computer systems under this section.
• 21 CFR 211.100 requires written procedures for production and process control, and it is the basis for FDA’s process validation expectations.
• 21 CFR 211.110 requires in-process sampling and testing to monitor output and validate the performance of manufacturing processes that may cause variability.²⁴
• 21 CFR Part 11 applies to electronic records and signatures, including those an AI system creates, modifies, or relies on.

The First Warning Letter Made the Point

On April 2, 2026, FDA issued a warning letter to Purolea Cosmetics Lab, a drug manufacturer in Livonia, Michigan, with a section headed “Inappropriate Use of Artificial Intelligence in Pharmaceutical Manufacturing.”²² The firm had told investigators it used AI agents to create drug product specifications, procedures, and master production or control records. FDA wrote: “If you use AI as an aid in document creation, you must review the AI generated documents to ensure they were accurate and actually compliant with CGMP.” The agency cited the failure as a violation of 21 CFR 211.22(c).

The letter also noted that the firm had not conducted process validation before distribution, as 21 CFR 211.100 requires, and that the firm replied it was not aware of the requirement because the AI agent never told it. FDA stated that if the firm resumes production and uses AI for CGMP activities such as developing procedures and specifications, any AI output or recommendation must be reviewed and cleared by an authorized human representative of the quality unit.²²

This was a firm with basic CGMP failures well beyond AI, and it would be wrong to read the letter as a crackdown on AI in manufacturing. The narrower lesson is more useful: FDA applied existing regulations to AI use without any AI-specific guidance in place. The planned documents will not change that.

The 211.110 Draft Already Takes a Position on Models

The January 2025 draft guidance on 21 CFR 211.110 is the closest thing FDA has published to a CDER position on models in commercial control strategies. It is a draft, but it is specific. It says process models can be part of the overall control strategy and that advanced manufacturing is generally suited to them. It then says FDA has not seen process models that show both that their underlying assumptions will stay valid during routine manufacturing and that the manufacturer can detect when an assumption is no longer valid.¹⁵ Its conclusion: “control strategies that rely solely on current process models would be insufficient to satisfy the requirements of § 211.110.”

The draft recommends pairing process models with in-process material testing or process monitoring, and it invites manufacturers who want to use alternative control strategies to contact CDER’s Emerging Technology Team early.¹⁵ The draft uses the general term “process model” rather than AI. But an AI model that predicts an in-process attribute is a process model for this purpose, and nothing in the draft suggests a machine learning model would be held to a lower standard than a mechanistic one.

Where FDA’s Published Thinking Points

Reading the published documents side by side, several themes repeat. These are the parts of a readiness plan you can build with reasonable confidence, because they appear in more than one FDA or FDA-endorsed source. We have kept each theme tied to the source that states it.

The Manufacturing Example in the Credibility Draft

The credibility draft includes one worked manufacturing example, and it is worth studying because it shows how FDA reasons about AI in a release decision. Drug B is a parenteral injectable in a multidose vial, and fill volume is a critical quality attribute. A manufacturer proposes an AI-based visual analysis system for 100% automated assessment of fill level. The question of interest is “Do vials of Drug B meet established fill volume specifications?”¹³

The draft rates the decision consequence as high, because a wrong fill volume could lead to medication errors. But the manufacturer also measures fill volume on a representative sample for each batch as part of release testing, so the AI model is not the sole determinant of release. That lowers model influence, and the draft rates the overall model risk as medium.¹³ The practical point for manufacturers: the controls around a model, not only the model, set how much evidence you need. Keeping an independent check in place is a legitimate design choice that reduces the credibility burden.

Established Conditions and Planned Changes

The discussion paper asked when an AI model becomes an established condition. The credibility draft gives a partial answer. It says detailed lifecycle maintenance plans, including performance metrics, monitoring frequency, and retesting triggers, “should be made available for review as a component of the manufacturing site’s pharmaceutical quality system,” with a summary in the marketing application for product- or process-specific models.¹³ It also says sponsors may use ICH Q12 tools, proposing model-related elements as established conditions along with a plan to manage changes to them, so FDA can indicate in advance which changes would not need a submission before they are made.

That is a draft recommendation, not a rule. But it lines up with what commenters asked for, and it uses tools (ICH Q12 established conditions and postapproval change management protocols) that are already final and in use. It is a sound basis for planning.

ICH Is Moving Toward a Process Models Guideline

The ICH Assembly endorsed a reflection paper on October 8, 2025, titled “Proposed ICH Guideline Work to Facilitate the Adoption of Advanced Pharmaceutical Manufacturing,” and FDA lists it under its FRAME initiative as a March 2026 item.^19,20 The paper names process modeling, including AI-based models, as an advanced manufacturing technology. It says there is a need for global alignment on terminology, a model risk framework, the basis for regulatory oversight, and data requirements. It proposes a new ICH guideline on process models as the preferred first step, one that would cover documentation of models and model updates over the lifecycle and could provide principles applicable to AI models.¹⁹

A reflection paper is not a guideline, and a new ICH topic has to be proposed and adopted before drafting starts. But this tells you the direction: FDA’s regional guidance on AI in manufacturing is likely to be followed, at some point, by harmonized ICH text on models generally. Manufacturers with sites in several regions should expect expectations to converge over time, not diverge.

What We Are Not Predicting

It is tempting to fill the gaps. We will not. The published record does not tell us whether the planned CDER guidance will:

• allow continuously learning models in commercial control strategies, and under what conditions;
• set specific validation or documentation deliverables for AI-based systems;
• address generative AI used for CGMP documents beyond the human review the Purolea letter already requires;
• define which AI uses in manufacturing need to be described in an application; or
• set expectations for quality agreements with AI and cloud vendors.

Each of these was raised as a question in 2023. None has a published answer from CDER yet. A readiness plan that assumes a specific answer to any of them is a bet, and it should be labeled as one.

Preparing Now Without Guessing: A Six-Step Plan

The good news is that most of the work that prepares you for the planned guidance is work CGMP already asks for. The steps below rely only on requirements in force today and on themes that appear in several published FDA sources. None depends on guessing unpublished text.

How the Six Steps Map to the Sources

Keep the Plan Proportionate

A common failure mode is to treat every AI feature the same way. The published sources point the other way. The credibility draft, the FDA and EMA principles, and the ICH reflection paper all tie the level of oversight to risk and to the model’s contribution to assuring quality.^13,17,19 An anomaly detection model that flags equipment trends for an engineer to review does not need the same evidence package as a model that decides whether a vial is released. Spend the effort where the decision consequence is.

What to Do When the Drafts Appear

Because both planned documents are listed as draft guidances, their publication will open a comment period, typically announced in the Federal Register. The credibility draft and the 211.110 draft each allowed about 90 days.^14,15 That window is short for a manufacturer that has not already done its inventory and classification.

A 90-Day Response Plan

When a draft is published, the first job is a gap assessment, not a rebuild. Compare the draft’s recommendations to the inventory and context-of-use records from steps 1 through 3. Mark each gap as one of three types: something you should do regardless because CGMP already requires it, something the draft recommends that you agree with, or something you think is unclear or unworkable. The third group is what belongs in your comments.

Comments are most useful to FDA when they are specific. The 2023 feedback shows the agency heard broad requests for clarity many times. A comment that describes a real use case, the control you have in place, and the exact recommendation that would create a problem is far more likely to shape the final text. Trade associations will comment too, but they cannot describe your process.

Do Not Rebuild Around a Draft

Drafts change. The credibility guidance has been a draft since January 2025. The 211.110 guidance has been a draft for the same period. Treat a new AI manufacturing draft as FDA’s current thinking, which is useful, but avoid large system redesigns that depend on a draft recommendation surviving to the final version unchanged. The durable investments are the ones in the six-step plan, because they rest on requirements already in force.

Watch the Other Regions, but Keep FDA Separate

Manufacturers with European sites are also tracking the EU’s draft GMP Annex 22 on AI, which remains a consultation draft with no final text. That is a separate process with separate scope, and we cover it in other articles. The ICH reflection paper is the place where these regional efforts may eventually meet.¹⁹ For now, keep one set of records (inventory, context of use, risk rating, lifecycle plan) and map it to each region’s expectations as they are published, rather than building separate programs.

Signals Worth Tracking

• The next CDER guidance agenda posting, to see whether either AI item is removed, retitled, or marked published.²
• CDER’s newly added guidance list and the Federal Register for notices of availability.²⁷
• Any final version of the credibility guidance or the 211.110 guidance, since the manufacturing AI guidance will likely be written to fit alongside both.
• FDA’s FRAME initiative page, which collects CDER’s advanced manufacturing documents, and the report on Emerging Technology Team activities that FDA’s strategy document says will be issued by December 2026.^20,26
• New ICH topic decisions on process models.
• Warning letters and Form 483 observations that cite AI, which show how investigators apply current rules in practice.

Conclusion

CDER has two AI manufacturing guidances on its 2026 agenda: one on AI and ML quality considerations in pharmaceutical manufacturing, and a supplemental computer software assurance guidance for AI-based systems added in July 2026. Neither is published, the agenda is not binding, and when they come, they will come as drafts. Meanwhile, the January 2025 credibility guidance and the 211.110 guidance remain drafts, the device CSA guidance is final but not written for drug CGMP, and the only binding rules on AI in drug manufacturing are the ones that have always applied: CGMP, Part 11, and the quality unit’s responsibility for what gets approved. The Purolea warning letter showed FDA will apply those rules to AI now.

The practical path is to prepare for FDA AI pharmaceutical manufacturing guidance by doing the work current rules already require, organized the way FDA’s published thinking points: know every model, classify it by what decision it affects, document its context of use and risk, control its changes, and tighten the data and vendor controls around it. That work holds up whatever the final text says. Sakara Digital works with pharma and biotech organizations putting this kind of AI governance into their quality systems. If you are sorting out where your manufacturing AI stands against current CGMP and want an independent view on where to start, we are happy to have that conversation.

For Further Reading