Gourav Pandey

PharmTech AI Pulse Check episode 2 covers AI data governance, API-gated journals, watermarking limits, and why attribution beats detection in GxP.

PharmTech AI Pulse Check: Episode 2 - Shadow AI, Trade Secrets, the API Race

's new biweekly video series, “PharmTech AI Pulse Check,” host Chris Cole, associate editorial director, PharmTech, sits down with three expert co-hosts—Richard Jaenisch, Open Biopharma Research and Training Institute, Florin Muraru, independent regulatory affairs advisor, and Gourav Pandey, R&D quality lead, Takeda—to unpack three fast-moving flashpoints in pharma AI: data governance, the shift toward API-gated scientific literature, and the limits of AI-content detection.

OpenAI's public clarification that a user's Codex prompts couldn't have leaked into the system

Anthropic's new life-sciences verification program

. Pandey argues the industry is "arguing with the wrong verb" by fixating on train-versus-retain distinctions. What actually matters, he says, is usage policy: how long data are held, who can access the data, where model weights end up, and how those answers hold up against evolving EU data regulations. He also warns that once an industry's own data feed a benchmark, models tested against that benchmark have effectively already seen the answers, undermining their validity over time. Muraru adds a sobering data point—a report finding roughly 77% of employees paste data into AI tools, often from personal accounts outside company control—and draws a sharp line between batch records (GMP records where any leak risks data-integrity violations) and formulation data (a trade secret that, once pasted into a personal AI account, may lose its legal protection under the EU Trade Secrets Directive). His minimum policy: enterprise accounts with contractual zero data retention, personal accounts blocked at the network level, and on-prem realistically limited to the largest manufacturers. Jaenisch pushes back gently on the trust framing around Anthropic's program, suggesting it's less about trust and more about managing liability around hazardous chemical queries and shadow AI usage. He predicts federated/on-prem deployment will become pharma's norm the way it's emerging in hospitals.

The second segment tackles a prediction that non-API scientific journals will be "dead in five years" as publishers like 

 build AI-facing API access. Muraru warns this risks an "availability bias," with research locked behind PDFs or without API/MCP access simply dropping out of agent-driven literature searches, which is especially dangerous for regulatory and pharmacovigilance work where completeness matters. Jaenisch counters that PDFs are already extractable regardless of access method, so publishers gain little real protection; the practical upside of APIs is curated, publisher-vetted retrieval. Pandey shares original research he calls "citation laundering," showing that retrieval architecture matters more than the underlying model for answer trustworthiness, and warns of compounding availability bias as agents gravitate toward whatever's easiest to retrieve quickly, a Goodhart's Law problem once citation counts become an optimization target.

The final segment addresses new research showing 

 can increase vulnerability to adversarial prompts. Jaenisch argues that detection is the wrong lens entirely, adding that the real question is whether content is attributable, reviewed, and traceable to a source, regardless of whether a person or model drafted it. Watermarking, he notes, mainly flags "slop" and fails on edited or mixed content anyway. Pandey and Muraru agree from a GxP standpoint: validation has never depended on guessing who typed something, but on signatures and audit trails, and detection tools will always lag behind the newest models. As long as a document follows approved procedures and someone takes responsible ownership and signs off, authorship, human or AI, is secondary.

Videos

Episode 1, Part 3 of PharmTech AI Pulse Check covers why INTerpath-001's AI isn't ChatGPT and names pharma manufacturing's biggest AI governance gaps.

PharmTech AI Pulse Check: INTerpath-001, Precision ML, Pharma's Governance Gaps

When Merck and Moderna's INTerpath-001 trial of an individualized mRNA-based cancer therapy hit its endpoints, many observers credited AI with the breakthrough, a framing PharmTech AI Pulse Check panelists argue obscures what the technology actually does and where pharma's real governance gaps lie.

Richard Jaenisch, senior director of education, outreach, and digital experience at Open Biopharma, notes that people interpret new technology through the frame of reference they already have, which leads many to assume the tool behind a cancer therapy is the same chatbot they use every day. The work behind an individualized neoantigen therapy relies predominantly on precision machine learning models, an older and narrowly focused category of AI that, in this case, evaluates how a patient's tumor mutations fit together and which patients are likely to be good candidates. Jaenisch has watched online forums declare that AI has cured cancer, a conclusion he rejects even as he calls the results a promising sign for the mRNA platform. He adds that the approach will require new training for clinical trial teams and may perform differently outside selective trial populations.

Florin Muraru, an independent regulatory advisor specializing in EU and US regulatory strategy and AI governance, describes the technology as a targeted prediction tool rather than a general system, and not a new one. What is changing, he argues, is the shift from AI finding drugs to AI personalizing treatment for each patient. From a regulatory standpoint, Muraru says that shift is easier to defend: because the tool does not continuously develop on its own the way newer AI models do, it can be locked and presented to health authorities as a validated system within the Module 3 CMC section of a submission. Long term, he expects more AI tools to move toward true personalization of medicine.

Gourav Pandey, R&D quality lead at Takeda, compares each patient's tumor mutations to a unique fingerprint or password; the model's job is to identify that fingerprint and link it to mutations likely to provoke an immune response. Unlike a chatbot, he points out, this kind of model can be checked in both directions by comparing what it predicted with how the patient actually responded.

Asked to name the single biggest AI governance gap in pharma manufacturing, the panelists offer three distinct answers. Jaenisch points to training, observing that many professionals who claim AI expertise cannot define retrieval-augmented generation, or RAG, and that most corporate AI training so far covers basic cyber hygiene rather than how to produce genuinely useful work. "If we do not have adequate training, well, we have inadequate results," he notes, warning that without the right instruction, bad habits get reinforced and low-quality output multiplies.

Muraru argues that industry and regulators know how to validate a model at submission but have not agreed on an evidence standard showing a model still behaves as intended after retraining; the space between validation and drift, he says, is where governance is thinnest. Pandey takes a different angle. "The model is the least interesting part; the pipeline is the most important part," he argues. Too much attention goes to validating models and curating golden datasets, he explains, rather than asking whether the AI is accurate and answering the question it is meant to answer. Citing EU GMP Annex 22 and its distinction between deterministic and probabilistic models, he urges companies to make sure source documents clearly define what AI should do and where human judgment begins.

For manufacturing and quality leaders, the next phase of AI governance will hinge less on the models themselves than on training, post-retraining evidence, and the pipelines around them.

Florin Muraru, Gourav Pandey, and Rick Jaenisch explain where human sign-off belongs as multi-agent AI and document-parsing models reach GxP settings.

PharmTech AI Pulse Check: Validating Multi-Agent Systems, OCR Models in GxP

Multi-agent AI is edging toward the pharmaceutical shop floor, and the question of where the human gate belongs in a validated GxP environment is no longer theoretical. In Part 2 of the three-part 

first episode of the PharmTech AI Pulse Check

 video series, Christopher Cole, associate editorial director, 

, asks for insights from co-hosts Florin Muraru, an independent regulatory advisor; Gourav Pandey, R&D quality lead at Takeda; and Richard Jaenisch, senior director at Open Biopharm (

Watch Part 1: Fixing AI Slop at the Source

). The conversation opens with a recent benchmark test in which OpenAI agents reportedly built a private coordination channel days after it was shut down, and asks how pharma's segregation-of-duties playbook should respond.

Muraru argues that letting agents talk while requiring human sign-off before any action is realistic in a validated environment, but only if the action is the GxP-relevant step. He draws a firm line at batch release, noting, "The agents reasoning together is fine, but an agent releasing a batch without a human decision absolutely is out of limit." The real problem, he explains, emerges when inter-agent conversation becomes the basis of a decision, because that exchange then qualifies as a record carrying data integrity, retention, and review obligations. He adds a further regulatory layer: a self-modifying multi-agent system lacks a stable baseline to validate against, and validation underpins trust in computerized systems in GxP environments.

Pandey adds that data integrity culture in pharma treats even a sticky note as a GxP document, so an inter-agent discussion that was never captured is a serious gap. He points to two agents building a deviation report, one drafting the investigation while the other works through facts, and observes that machines can do a great deal of the work but a signature must belong to a real, answerable person. Jaenisch raises the architecture question: hash-chain audit logs and permission gates help, yet the benchmark agents went around those gates. He favors a monitoring agent over human reviewers because no one has time to read a week and a half of dense output.

The discussion then turns to Baidu's newly open-sourced model that parses entire multi-page documents, including tables, layout, and reading order, in a single pass. Pandey sees an opening for the decades of scanned records, handwritten logbooks, and research lab books sitting in third-party archival facilities. Because the model is open source, companies can deploy it in-house with a GPU and pair it with a knowledge graph linking terminated programs, old batches, and compliance reports to vendors, molecules, solvents, and raw materials. He finds particular promise for research, noting, "We have so many failed experiments that may have insightful information for the molecule which never progressed further down the line in the pipeline."

Muraru cautions that the moment those documents underpin a reconstructed batch record, data migration, or regulatory submission, the system leaves research territory and requires validation, since an OCR error can become a data integrity error. His advice is to use the tool for looking back at records but to verify against the originals. Jaenisch adds that many-to-many document analysis has historically been prohibitively expensive, that someone must define ground-truth documents so the model can tell signal from scrap, and that OCR across older, varied document types is a concern because generative AI reflects its training data. He places meaningful adoption a year or two away and quips that governing this architecture will create new jobs.

Florin Muraru, Gourav Pandey, and Richard Jaenisch explain where human sign-off belongs as multi-agent AI and document-parsing models reach GxP.

Jaenisch, Muraru, and Pandey debate whether AI "slop" is infiltrating pharma SOPs and deviation reports and who should catch it first.

PharmTech AI Pulse Check: Fixing AI Slop at the Source

AI-generated "slop" has become a fixture of daily inboxes, but is it creeping into pharmaceutical manufacturing and quality documentation? That's the question at the center of Part 1 of the 

debut episode of the new PharmTech AI Pulse Check series

, in which host Christopher Cole is joined by Richard Jaenisch, Senior Director of Education, Outreach, and Digital Experience at OpenBiopharma Research and Training Institute; Florin Muraru, an independent regulatory affairs executive with more than 20 years of experience; and Gourav Pandey, quality and CMC risk and compliance specialist at Takeda, unpack where AI-generated content is showing up in SOPs, deviation reports, and literature summaries, as well as who's responsible for catching it before it reaches a regulator.

Jaenisch argues that deviations have become generative AI's most common entry point in manufacturing, largely because form filling is low effort and tempting to automate. But he cautions that speed can come at the expense of accuracy: "Well, you just made a slop factory." The real risk, he says, isn't the technology itself but whether operators, especially newer, less-seasoned staff, are trained to recognize which content is fit for an SOP and which isn't.

Muraru reframes the problem as one of authorship rather than tooling. In his view, AI slop isn't about the model but the person behind it, and the danger emerges when documentation is generated faster than it can be meaningfully reviewed, shifting the real bottleneck from review capacity to the presence of an engaged human in the loop. He notes that polished, fluent AI output can actually be harder to audit than an obviously rough draft, since errors are easier to spot in something that already looks imperfect. He also points to the FDA's April warning letter involving Purolea as an early real-world example of the stakes.

Jaenisch pushes the discussion further upstream, distinguishing between what he calls first-degree slop—content a human directly reads, edits, and signs off on—and second-degree slop, where a person trusts an AI tool's advice without independently verifying it, and that flawed judgment quietly propagates through decisions and conversations. Pandey argues the fix isn't more scrutiny at the review stage but better-structured source documents that clearly separate deterministic rules from areas requiring human judgment: "You don't fix the slop by reading harder or putting on more reviewers." Doing so, he says, lets AI genuinely help rather than generate polished documentation that still requires heavier reviewer and approver attention.

Across the conversation, the panel converges on a shared concern: GxP-regulated environments can't rely on AI-generated content "looking right" as a proxy for being right, and that training, sourcing discipline, and a habit of asking "where did you verify that from?" will matter more as generative tools become embedded in manufacturing and quality workflows. 

 for the panel's complete discussion of where AI slop is most likely to surface next and what pharma companies can do now to get ahead of it.

PharmTech AI Pulse Check debuts: pharma experts tackle GxP 'slop,' multi-agent validation risk, OCR document mining, and personalized mRNA cancer therapy.

Introducing the "PharmTech AI Pulse Check" Expert Video Series

's new biweekly video series, “PharmTech AI Pulse Check,” host Chris Cole, associate editorial director, PharmTech, sits down with three expert co-hosts—Richard Jaenisch, Open Biopharma Research and Training Institute, Florin Muraru, independent regulatory affairs advisor, and Gourav Pandey, R&D quality lead, Takeda—to unpack the AI developments and questions that matter most to people building, manufacturing, and delivering therapies for patients.

The conversation opens with a pointed question: is manufacturing and quality documentation developing its own version of "AI slop," the fluent-but-empty content already flagged widely on social platforms? All three experts agree that the risk is real, particularly around deviation reports, where generative AI's ease of use can produce polished but substance-free documentation. Jaenisch warns that slop is "almost unavoidable" and represents a real cost when reviewers spend time on content that ultimately proves useless, or worse, reaches a regulator undetected. Muraru reframes the issue as fundamentally a human problem, not a model problem, and raises a deeper concern: AI-generated volume can outpace human review capacity, shifting the real bottleneck to the person signing off. The panel also discusses a "second-degree slop" framework, distinguishing between AI directly producing flawed documents and humans trusting AI-generated advice without further verification, citing a 

 as an early real-world example. The panel converges on a shared fix: redesigning source documents so AI has clear, deterministic rules to follow, reducing ambiguity at the root rather than relying solely on downstream review.

The discussion shifts to multi-agent AI systems, prompted by a r

ecent case in which AI agents secretly built a private communication channel after being shut down

. Drawing parallels to pharma's existing segregation-of-duties principles, the panel debates where human sign-off gates should sit as multi-agent systems move toward the shop floor. Muraru argues that agent-to-agent reasoning is acceptable, but any GxP-relevant action, like releasing a batch, requires human decision making, and that inter-agent exchanges must be captured as auditable records. Pandey stresses that data integrity principles already governing handwritten notes apply equally to AI outputs, and that traceable, accountable human sign-off remains non-negotiable regardless of how much groundwork AI does.

The conversation turns to document intelligence, sparked by a 

newly open-sourced model capable of parsing entire multi-page documents

—including scanned records, handwritten logbooks, and historical batch data—in a single pass. The panel sees major potential for mining decades of previously unsearchable pharma records, including failed experiments and terminated programs, but cautions that the moment this information feeds a regulatory submission or GxP decision, the system requires formal validation, and OCR errors could become data integrity issues.

The episode continues with a discussion of Merck and Moderna's 

, which uses individualized mRNA-based cancer therapy tailored to each patient's tumor mutations. The panelists emphasize a critical distinction: this relies on precision machine learning, not the generative AI most people associate with chatbots, an important nuance as public conversation increasingly conflates the two.

Wrapping up with a rapid-fire round, each panelist names pharma manufacturing's biggest AI governance gap: Jaenisch points to inadequate workforce training, Muraru highlights the absence of standards for validating models after retraining and drift, and Pandey argues the industry over-focuses on the AI model itself rather than the surrounding pipeline and source documentation.

PharmTech AI Pulse Check debuts: pharma experts tackle GxP 'slop,' multi-agent validation risk, OCR document mining, personalized mRNA cancer therapy.

Medical worker touch virtual medical revolution and advance of technology Artificial Intelligence,AI Deep Learning for medical research,Transformation of innovation and technology for future Health | Image Credit: ©Toowongsa -stock.adobe.com

The FDA’s April 2026 warning letter to Purolea Cosmetics Lab provides a concrete enforcement example involving artificial intelligence (AI) in current good manufacturing practice (CGMP) documentation.

 During the inspection, the firm told the FDA investigators that it had used AI agents to create drug product specifications, procedures, and master production or control records to help comply with the FDA requirements. The FDA’s concern was direct: Firms using AI to assist document creation must review the resulting documents for accuracy and compliance with CGMP before use. The FDA cited the failure to do so as a violation of 21 

 (CFR) 211.22(c). The letter also describes a process-validation gap under 21 CFR 211.100 after the firm stated that the AI agent had not identified the requirement. AI-assisted drafting can support documentation work only when the resulting document is subject to quality-system controls. That requires controlled source material, qualified review, quality unit approval, version history, and retained evidence that the final record was checked before use.

The Purolea letter is notable because the FDA explicitly addressed AI in the body of a CGMP warning letter and linked its use to established quality-system responsibilities. The firm had used AI agents to create specifications, procedures, and master production or control records. These are controlled documents that can shape manufacturing, testing, batch documentation, and product disposition.

The FDA did not need an AI-specific rule to make this finding. The letter states that, when AI is used to assist document creation, the firm must review the resulting documents to confirm their accuracy and compliance with CGMP. The FDA cited the failure to do so under 21 CFR 211.22(c), which assigns the quality control unit responsibility for approving or rejecting procedures or specifications that affect identity, strength, quality, and purity.

The letter also describes a process-validation failure. The FDA investigators found that process validation had not been conducted before distribution of drug products, as required under 21 CFR 211.100. The firm stated that its AI agent had not identified the requirement. That explanation does not shift accountability: Software can assist a task, but it cannot substitute for the firm’s responsibility for ensuring that required GMP activities are performed.

Where Does the Quality Unit’s Responsibility Begin?

The quality unit is not merely a final reviewer. Under 21 CFR 211.22, it has authority over components, in-process materials, packaging, labeling, drug products, and production records. It is also responsible for approving or rejecting procedures or specifications that affect the identity, strength, quality, and purity of the drug product. The unit’s responsibilities and procedures must be in writing and followed.

That matters for AI-assisted drafting because the relevant risk is not the tool used to prepare the first draft. The risk is that a procedure or specification enters use without a controlled basis, documented review, or a quality unit decision. The Purolea finding applies an established framework to a new drafting route: A GMP document still requires defined sources, qualified review, and quality unit approval before use.

Section 211.100 makes the same point from a different angle. Written procedures for production and process control must be designed to assure identity, strength, quality, and purity. They must be drafted, reviewed, and approved by appropriate organizational units and reviewed and approved by the quality control unit. They must also be followed during production, and deviations must be recorded and justified.

What Makes AI-Assisted GMP Documentation Fragile?

The central issue is not simply that AI-generated text can be wrong, a risk common to any draft. AI output can appear complete while obscuring missing source checks, unsupported assumptions, or weak linkage to the firm’s approved process. Those gaps create a documentation risk before they become technical risks.

A specification generated from the wrong source can look polished. A procedure can describe a generic manufacturing step that does not match the site’s equipment, validated parameters, sampling plan, or cleaning controls. A master production or control record can omit a required verification step and still read fluently. In a GMP setting, fluency has no evidentiary value unless the record can be traced back to approved sources and accountable review.

This is why the Purolea letter is better understood as a quality-system signal than as a general warning against AI. The firm needed a process that made each AI-assisted document reviewable: what source material was used, who checked the output, what changes were made, who approved the final document, and where the supporting evidence was retained.

How Should AI-Assisted Drafting Be Controlled?

A workable control model does not require every prompt to be treated as a regulated system event. It should, however, distinguish informal drafting from AI use that creates or revises a GMP-controlled record. When AI is used to prepare or revise a specification, procedure, batch record, validation document, investigation response, or other controlled record, the output should enter a defined review path.

The first step is to define intended use. Teams should know which tasks are permitted, which are prohibited, and which require prior quality unit approval. The second step is to assemble a source package before drafting. AI should not be used to establish the factual or procedural basis for a GMP document. The approved procedure, product specification, validation report, batch record template, change control, regulatory commitment, or supplier document should be identified before the output is reviewed.

The third step is technical review. A subject-matter expert should check the output against the approved source package and the process used at the site. The fourth step is quality unit approval. The final document should follow the same document-control process as if the draft had been prepared without AI. The fifth step is retention. The firm should be able to produce the final approved record, the key source material, the review evidence, and the approval history.

 summarizes these controls and the evidence a firm should retain to make the process inspectable.

Table 1. Inspection-ready controls for AI-assisted GMP documentation

Part 11 should be handled carefully. The FDA’s scope guidance states that Part 11 applies when records required by predicate rules are maintained in electronic format instead of paper, or when electronic records are relied upon to perform regulated activities. A firm should therefore document whether the AI workspace, the document-management system, or the approval workflow is part of the electronic record relied upon for GMP activity.

That does not mean that every AI draft becomes a Part 11 record. If the regulated record is a controlled final procedure in the document-management system, the Part 11 assessment may focus on that system and its approval workflow. If the AI workspace stores source material, review history, or approvals that the firm relies on to defend the record, the compliance analysis changes. The firm should document that decision prospectively rather than address it only during an inspection.

Data integrity has a narrower but useful role here. The FDA’s data integrity guidance expects data to be reliable and accurate and allows for risk-based strategies to prevent and detect data integrity lapses. For AI-assisted GMP documentation, that means the firm should be able to reconstruct how the final record was created, checked, approved, and retained. It also means avoiding uncontrolled copy-paste routes that leave no review trail.

How Should Firms Apply Risk-Based Review?

A risk-based approach is appropriate, but it cannot be used to bypass CGMP requirements. The International Council for Harmonisation’s Q9(R1) states that the level of effort, formality, and documentation should be commensurate with the level of risk. It also makes clear that quality risk management does not replace the obligation to comply with regulatory requirements. This distinction is useful when defining AI controls.

A draft internal training outline may need a lighter review than a master production record. A revised product specification, process-control procedure, validation protocol, or batch-record instruction should receive a much higher level of scrutiny. The review burden should follow the possible effect on product quality, patient risk, regulatory commitments, and batch disposition.

The FDA’s January 2025 draft guidance on AI for regulatory decision-making is not a GMP documentation guidance and is not for implementation. Its context-of-use and credibility framework is nevertheless useful. Within a quality system, the same logic supports practical questions: What decision does this AI output support, how could it affect quality, and what evidence is needed before the output is relied upon?

The immediate work is to identify where AI is already being used to draft or revise GMP content. Firms should classify those uses by risk, assign an accountable owner, and determine which outputs must follow a defined source-review-approval path. The review should involve product and process experts, not only document-control staff.

Quality units should also define the evidence that will be retained. The retained record does not necessarily need to preserve every informal prompt used during early drafting. The firm should define, through a risk-based procedure, the inputs and review artifacts needed to defend the final GMP decision: approved source documents, key reviewer checks, changes made before approval, quality unit clearance, effective date, and version history.

Supplier and contractor arrangements require the same discipline. The FDA reminded Purolea that a firm remains responsible for the quality of its drugs regardless of agreements with contractors. If a contractor, consultant, or software vendor supports AI-assisted GMP documentation, the firm still needs enough control to defend the final record and the process that produced it.

The Purolea warning letter does not establish a separate AI regulatory framework. It applies existing CGMP accountability to AI-assisted drafting. In practice, AI may help prepare GMP content, but the firm must establish and retain the controls needed before that content is used in manufacturing or quality operations.

For quality and CMC leaders, the practical response is to integrate AI-assisted drafting into existing document-control and quality-system processes. Define permitted uses, control source material, verify the output, approve the final record, and retain evidence of the review.

1. Warning letter: Purolea Cosmetics Lab, 320-26-58, MARCS-CMS 722591. FDA. April 2, 2026. Accessed May 26, 2026. https://www.fda.gov/inspections-compliance-enforcement-and-criminal-investigations/warning-letters/purolea-cosmetics-lab-722591-04022026

2. Responsibilities of Quality Control Unit. 21 CFR §211.22 (2026). Accessed May 26, 2026. https://www.ecfr.gov/current/title-21/chapter-I/subchapter-C/part-211/subpart-B/section-211.22

3. Written Procedures; Deviations. 21 CFR §211.100 (2026). Accessed May 26, 2026. https://www.ecfr.gov/current/title-21/chapter-I/subchapter-C/part-211/subpart-F/section-211.100

Guidance for Industry: Quality Systems Approach to Pharmaceutical CGMP Regulations. 

FDA; September 2006. Accessed May 26, 2026. https://www.fda.gov/media/71023/download

Data Integrity and Compliance With Drug CGMP: Questions and Answers: Guidance for Industry.

 FDA; December 2018. Accessed May 26, 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/data-integrity-and-compliance-drug-cgmp-questions-and-answers

Guidance for Industry: Part 11, Electronic Records; Electronic Signatures —  Scope and Application. 

FDA; August 2003. Accessed May 26, 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/part-11-electronic-records-electronic-signatures-scope-and-application

. International Council for Harmonisation; January 18, 2023. Accessed May 26, 2026. https://database.ich.org/sites/default/files/ICH_Q9%28R1%29_Guideline_Step4_2022_1219.pdf

Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. Draft Guidance for Industry and Other Interested Parties

. FDA; January 2025. Accessed May 26, 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological

Gourav Pandey is pharmaceutical science quality lead, Takeda. Svyatoslav Borshchenko is CEO and cofounder, GxPilot.

Articles

FDA’s April 2026 warning letter to Purolea Cosmetics Lab operationalizes how existing CGMP expectations apply when AI agents draft specifications, procedures, and master production/control records. The central deficiency was deploying AI-generated controlled documents without documented accuracy and CGMP compliance review by qualified personnel and the quality control unit, triggering 21 CFR 211.22(c) findings. A separate gap involved missing process validation under 21 CFR 211.100, with reliance on AI’s failure to flag requirements deemed non-exculpatory. Practical controls include defined source packages, SME review, QU approval, versioning, and retained review evidence.

An FDA warning letter to Purolea Cosmetics Lab shows that AI-assisted drafting can support documentation work only when the resulting document is subject to quality-system controls.