Novo Nordisk Failed to Report Three Deaths on Ozempic. AI Would Have Caught Them.

The Warning Letter That Rewrites Pharmacovigilance

On March 5, 2026, the FDA issued Warning Letter 26-HFD-45-03-01 to David S. Moore, President of Novo Nordisk Inc. The letter cited serious violations of postmarketing adverse drug experience reporting under Section 505(k) of the Federal Food, Drug, and Cosmetic Act and 21 CFR 314.80. The drugs named: semaglutide (Ozempic, Wegovy), liraglutide (Victoza, Saxenda), nedosiran sodium, and estradiol.

Internal procedures at Novo Nordisk systematically filtered out valid safety reports before they could reach regulators.

Case Argus #1342548: a male patient on semaglutide died. Novo Nordisk invalidated the report, citing a "missing patient identifier." The identifier was present in the source documents. Case Argus #1171264: a lay reporter described the death of a patient receiving semaglutide. Novo closed the case because the reporter did not provide "consent" for followup. No such consent requirement exists in federal regulations. Case Argus #1079792: a physician reported that a patient on semaglutide became depressed and committed suicide. As of the date of the warning letter, there was no documentation that Novo Nordisk had investigated the case or submitted it to the FDA.

There was also Argus #1334278: a consumer reported suicidal ideation while taking semaglutide on December 9, 2024. The case entered medical review on December 16. It was not actually reviewed until February 3, 2025, nearly two months later, and only after FDA inspectors flagged it during their onsite inspection. It was submitted to the FDA on February 5, 2025.

These are not isolated failures. The violations indicate systemic deficiencies affecting Novo Nordisk's entire product portfolio.

FDA Warning Letter 26-HFD-45-03-01

The FDA identified four internal procedures that enabled the failures. Procedure Q014048 allowed the rejection of adverse event reports deemed "unrelated" to the product, directly contradicting the regulation's definition of an adverse drug experience as an event "whether or not considered drug related." Procedure Q0360683 required explicit consent from reporters before conducting followup, a requirement with no legal or regulatory basis. A third procedure allowed case rejection for "lack of patient identifiers" even when identifiers existed in source documents. And procedure Q040018, which required medical review completion by day 10, was routinely ignored, with cases sitting unreviewed for months.

This was the third FDA admonishment of Novo Nordisk in recent weeks. In November 2025, the agency issued a warning letter for the company's Bloomington, Indiana facility (a former Catalent site) after inspectors found more than 20 uninvestigated deviations linked to possible "mammalian hair" contamination in drug product vials. In 2025, Novo also received untitled letters for misleading advertising: the "There's Only One Ozempic" campaign featuring actors Justin Long and John Hodgman was cited as "false or misleading," and the "Live Lighter" campaign for Wegovy was cited for implying clinical superiority without evidence.

Novo Nordisk's share price fell 2.8% on the day the letter was made public, with declines reaching 3.6% on the NYSE in subsequent sessions. Over the broader 30 day period, the stock posted a 16.4% decline. TD Cowen downgraded the company from "Buy" to "Hold." Bernstein rated the stock "Underperform." At DKK 246.95, the shares sat approximately 25% below the consensus analyst target of DKK 330.48. The FDA required a corrective action plan within 15 business days, including a retrospective review of all adverse event reporting deviations since January 1, 2021.

A company that built $26 billion in annual semaglutide sales could not reliably report a suicide to the FDA.

The System Was Already Drowning

Novo Nordisk's failures occurred inside a pharmacovigilance system that was already overwhelmed before GLP 1 drugs became the fastest growing drug class in history.

A systematic review of 37 studies found a median adverse event underreporting rate of 94%. The interquartile range was 82% to 98%. For severe or serious adverse drug reactions specifically, underreporting runs at approximately 80%. Underreporting is greatest for psychiatric and gastrointestinal disorders, the two categories most relevant to semaglutide's safety profile. There is no significant difference between general practice and hospital settings. The problem is structural.

The adverse events that reach regulators are a sliver of what actually occurs. The cases Novo Nordisk failed to report were drawn from an already tiny fraction of actual events. When a company's internal procedures actively suppress reports from that fraction, the gap between reality and the regulatory record becomes dangerous.

The FDA's Adverse Event Reporting System (FAERS) contained 19,252,329 background patients and 57,212,790 adverse event records from Q1 2004 to Q2 2025. From 2016 to 2023, the database received 2,062,099 serious adverse event reports, approximately 257,000 per year. The system now processes more than 2 million individual case safety reports annually. On March 11, 2026, six days after the Novo warning letter, the FDA migrated FAERS to a new platform called the Adverse Event Monitoring System (AEMS), a signal that even the regulator recognizes its infrastructure needs modernization.

The human side of the equation is equally strained. Manual ICSR processing takes 2 to 4 hours per routine case. A pharmacovigilance team processing 400 cases per month consumes 1,000 specialist hours on case construction alone. Traditional case processing can consume up to two thirds of a company's entire pharmacovigilance budget. Meanwhile, 77% of community pharmacists rate their workload as "high" or "extremely high," and more than half experience burnout, with rates exceeding 80% among community pharmacists. Pharmacovigilance is specifically cited as a high pressure role where deadlines, compliance demands, and long project cycles create stress. Overworked professionals experience heightened dissatisfaction, leading to higher turnover, leading to further staffing deficits.

When Pharmacovigilance Breaks Down, People Die

Novo Nordisk is the latest in a string of major pharmaceutical companies that failed at drug safety monitoring. The pattern has cost tens of thousands of lives and billions of dollars.

Vioxx (rofecoxib), Merck, 2004. Merck voluntarily withdrew Vioxx in September 2004 after evidence emerged of increased heart attack risk. Dr. David Graham of the FDA estimated the drug caused 55,000 premature deaths from heart attacks and strokes and 88,000 to 140,000 cases of serious heart disease in the United States alone. Merck employees ghostwrote 20 scientific articles and published them under respected scientists' names. When a clinical trial showed cardiovascular risk, Merck officials claimed naproxen decreased cardiovascular risk rather than Vioxx increasing it, an inversion of causality that the data did not support. Senior FDA officials called Dr. Graham's research "junk science" and retaliated against researchers who spoke publicly.

Roche, 80,000 unreported adverse events, 2012. The UK's MHRA discovered that Roche had failed to properly assess approximately 80,000 cases of potential drug side effects, including details of 15,161 deaths. The reports came from a US patient support program for patients who could not afford their medications. The cases had never been processed within Roche's pharmacovigilance database, meaning they were never passed to EU regulatory authorities. The EMA initiated an infringement procedure. Roche avoided a fine by cooperating and correcting the issue, but the case became a reference point in pharmaceutical regulatory compliance failure.

Abbott Laboratories, Depakote, 2012. Abbott paid $1.5 billion to resolve allegations of illegal off label promotion of Depakote and failures to report adverse events tied to those uses. Earlier, in 1999, Abbott was fined $100 million for manufacturing violations at its Lake County, Illinois plant, with an additional $112 million in subsequent penalties.

GlaxoSmithKline, Avandia. GSK was accused of withholding data showing increased cardiovascular risks with the diabetes drug Avandia (rosiglitazone), leading to severe market restrictions and a $3 billion settlement.

In every case, company procedures prioritized commercial protection over safety signal escalation. Reports were filtered, delayed, or reframed. The failures were discovered by external investigators, not internal quality systems. By the time regulators intervened, the damage was measured in lives. Every one of these failures involved drug safety systems that were supposedly compliant with pharmaceutical regulations. Every one of them showed that manual pharmacovigilance processes, dependent on human judgment under organizational pressure, are vulnerable to the same forces that created the failures in the first place.

The GLP 1 Monitoring Crisis

The pharmacovigilance challenges surrounding GLP 1 drugs are enormous. The Novo Nordisk warning letter arrived at a moment when the monitoring infrastructure for semaglutide and its class was already under extraordinary strain.

As of Q2 2025, 11 million unique patients were using a GLP 1 receptor agonist across indications, with consistent year over year growth. Pharmacy records show 19.1 million patients with GLP 1 prescriptions. An estimated 16 million Americans, 6% of US adults, have used injectable GLP 1 drugs for weight loss. Prescriptions for GLP 1s in obesity indications rose 587% from 2019 to 2024. In February 2024 alone, semaglutide drugs were prescribed 2.6 million times. Combined semaglutide sales reached $26 billion in 2024, growing at 40% annually. The GLP 1 weight loss market is projected to reach $48.84 billion by 2030. An additional 1.5 million users of compounded GLP 1 drugs in the US as of January 2026 exist largely outside standard pharmacovigilance systems entirely.

The adverse event data for the class is enormous and contested. FAERS contains 250,014 adverse event reports associated with GLP 1 receptor agonists from 2005 to 2024. A 2024 pharmacovigilance study identified 22,287 adverse reaction records related to semaglutide specifically. The reporting frequency of gastrointestinal disorders for semaglutide was 4.21 times higher than for other drugs in the overall database. Pancreatitis (389 cases, reporting odds ratio of 20.27) carries the highest clinical priority score. Emerging signals include hemorrhagic diarrhea (ROR 3.69), pancreatic failure (ROR 36.34), hepatic pain (ROR 4.20), and abnormal hormone levels (ROR 6.51).

The mental health signal is the most contested and the most directly connected to the Novo warning letter. FAERS disproportionality analysis shows significant signals for depression (ROR 1.87) and suicide/self injury (ROR 1.73) with semaglutide. Real world cohort studies show the opposite: semaglutide associated with lower risk of suicidal ideation in overweight and obese patients. The EMA initiated a formal investigation in July 2023 and concluded the evidence was insufficient for a causal association. The FDA recommends close monitoring for mood changes, depression, and suicidal behavior during treatment.

The Novo Nordisk warning letter cited a completed suicide and a case of suicidal ideation in semaglutide patients that went unreported. For a drug class already under regulatory scrutiny for mental health effects, the failure to report and investigate these specific cases gutted the pharmaceutical quality control process at exactly the moment it mattered most.

The pharmacovigilance system was built for a world where drug classes had tens of thousands of users. GLP 1 drugs have tens of millions. Novo Nordisk's compliance failures exposed what happens when that gap goes unaddressed.

Semaglutide's patent expiration is approaching, which will further expand access and monitoring volume. The current system cannot scale through manual processes alone.

AI Pharmacovigilance Tools: Who's Building What

Multiple pharmaceutical companies already deploy AI tools that would have caught the Novo Nordisk failures. Novo was not among them.

Machine learning models achieve AUCs of 0.92 to 0.95 for signal detection in pharmacovigilance, outperforming traditional disproportionality methods. The WHO's vigiMatch system detects approximately 50 million report pairs per second. Natural language processing achieves 70% to 82% accuracy in extracting adverse drug reactions from unstructured data and reduces manual workload by up to two thirds. AI systems now translate adverse drug reaction calls received in any language, automatically identify the four minimum reporting elements (identifiable reporter, identifiable patient, adverse reaction, suspect medicinal product), and populate electronic ADR forms.

The numbers hold up. ArisGlobal's NavaX engine delivers 80% faster signal assessment, approximately 50% fewer false positives, and 65% efficiency gains across case processing. Its sixth Top 25 pharmaceutical company client adopted NavaX in 2025. Indegene reports 40% to 60% cost reduction from ICSR processing automation. A Pfizer pilot achieved F1 scores of 0.72 to 0.74 for extracting reporter, drug, and adverse event terms using commercial AI systems.

The FDA itself has entered this space. Project Elsa (Electronic Language System Assistant), launched June 2, 2025, is built on Anthropic's Claude model operating within AWS GovCloud. It supports automated literature review for adverse event signals, triage and prioritization of case information, and ICSR intake processing. The tool is already used in training across the agency by scientists, reviewers, and field investigators. But Elsa has its own problems: the system has experienced false citations and data hallucinations, and questions remain about data isolation between different sponsors' submissions.

The operational arithmetic is straightforward. A pharmaceutical company spending $100 million annually on pharmacovigilance that achieves the demonstrated 40% to 60% cost reduction through AI tools saves $40 million to $60 million per year. But the Novo Nordisk crisis makes a different case: the cost of unreported deaths, regulatory sanctions, and billions in lost market capitalization dwarfs the investment required to deploy these systems.

Therapeutic Drug Monitoring and the Prediction Frontier

Pharmacovigilance is reactive by design. It detects adverse events after they occur. The more valuable application of AI in drug safety is prediction: identifying which patients will experience adverse events before they happen, and intervening before the event enters the reporting system at all.

Therapeutic drug monitoring is the clinical practice of measuring specific drug concentrations in a patient's blood at designated intervals to maintain levels within the therapeutic window, high enough for efficacy, low enough to avoid toxicity. Approximately 20 to 26 drugs are routinely monitored in clinical laboratories, with an additional 25 to 30 drugs that may benefit from monitoring. The drug classes include antiepileptics (phenytoin, carbamazepine, valproic acid), immunosuppressants (cyclosporine, tacrolimus, sirolimus), cardiac drugs (digoxin), antibiotics (vancomycin, aminoglycosides), and psychiatric medications (lithium, clozapine). The therapeutic drug monitoring market reached $1.92 to $2.30 billion in 2025 and is projected to hit $3.44 billion by 2029.

AI and machine learning are changing how this works through four approaches. Model Informed Precision Dosing (MIPD) combines Bayesian estimation with ML to predict individual pharmacokinetics and optimize dosing; XGBoost models have demonstrated performance comparable to or exceeding traditional methods. Hybrid PK/ML models integrate the physiological interpretability of population pharmacokinetic frameworks with the corrective power of machine learning. ML based concentration prediction processes patient demographics, genetic information, concurrent medications, and monitoring results to predict individual drug pharmacokinetics with greater precision than conventional methods. And closed loop dosing systems monitor drug levels in real time and adjust dosages accordingly.

Drug interactions compound the challenge. In the United States, 74,000 emergency department visits and 195,000 hospitalizations annually stem from antagonistic drug interactions. Among hospitalized patients, 6.7% experience a serious adverse drug reaction, with a fatality rate of 0.32%, resulting in more than 2.2 million serious ADRs and over 106,000 deaths annually. An estimated 30% of these hospitalizations could be avoided with proper drug interaction monitoring.

The tools for drug interaction prediction are improving but inconsistent. DrugBank 6.0 now catalogs 1,413,413 drug interactions, a 300% increase from the prior version, along with 2,475 drug and food interactions. Micromedex draws from approximately 8,500 journals and 750,000 publications and serves 4,500 hospitals. But a 2025 comparative analysis of five popular drug interaction checkers found they agreed on only 16% to 24% of identified interactions for SSRIs, revealing significant inconsistency in drug testing methods across the platforms clinicians rely on daily.

The connection between therapeutic drug monitoring, drug interaction prediction, and pharmacovigilance is direct. When monitoring fails or is absent, adverse events increase, feeding the pharmacovigilance burden that already overwhelms the current system. AI monitoring can analyze patient data, including genetic information, medical history, lifestyle factors, and concurrent medications, to anticipate adverse drug reactions before they happen. That is the shift from reactive drug safety to predictive pharmaceutical risk assessment, and it is where the technology has the greatest potential to reduce harm rather than merely process the paperwork that documents it.

The Regulatory Reckoning

The Novo Nordisk warning letter landed in the middle of three converging regulatory deadlines that will reshape pharmacovigilance requirements.

ICH E2B(R3): April 1, 2026. The international standard for electronic individual case safety report transmission becomes mandatory. The FDA began implementation on January 16, 2024, with a transition period that ends April 1, 2026. E2B(R3) provides internationally standardized data formats for safety reporting, including enhanced file structures, data elements, coding structures, medical terminologies, and transmission standards. Companies that have not upgraded their safety databases face compliance risk at the exact moment when the Novo warning letter has put pharmacovigilance under a microscope.

EU AI Act high risk requirements: August 2, 2026. The strictest requirements for high risk AI systems take effect. AI tools used in drug safety are candidates for high risk classification, requiring extensive documentation, human oversight, and transparency measures. The ban on prohibited AI practices already took effect on February 2, 2025.

EMA GVP Module Updates: February 2026. The updated Good Pharmacovigilance Practice modules took legal effect in February 2026. The key change: AI pharmacovigilance tools are now expected, not optional. Modern pharmacovigilance systems must use advanced analytics and automated signal detection to remain fit for purpose. From August 2025, Marketing Authorization Holders must routinely monitor EudraVigilance alongside all other relevant data sources. It remains the company's responsibility to validate, monitor, and document AI/ML model performance and include AI operations in the pharmacovigilance system.

Enforcement is intensifying. In 2025, the FDA issued more than 50 warning letters to GLP 1 compounders and manufacturers. The Novo Nordisk case shows willingness to cite a company with $26 billion in semaglutide revenue for systemic pharmacovigilance failures. Consent decrees remain the FDA's most severe enforcement tool, and historical examples like Abbott's $100 million in fines and ongoing monitoring requirements show the agency will use them.

Manual pharmacovigilance processes that were considered adequate five years ago no longer meet the standard. Pharmaceutical regulatory compliance now requires electronic reporting, structured safety data, expedited timelines for serious adverse events, and increasingly, AI signal detection. Companies that treat AI pharmacovigilance as a 2028 initiative will find themselves explaining to regulators why their systems fall short of standards published in 2026.

What This Means for Formulation Science

The Novo Nordisk pharmacovigilance failure is a data processing problem. Safety data existed. Reports were filed. Patients, physicians, and consumers contacted the company. The system failed at the point where data needed to be evaluated, escalated, and transmitted to regulators. Every one of the four procedural failures the FDA identified (causality rejection, consent requirements, identifier invalidation, review delays) is a failure of data processing, not data collection.

Formulation development has the same problem at a different scale. Formulation scientists work with enormous quantities of experimental data: dissolution profiles, stability results, excipient compatibility studies, process analytical technology outputs, critical quality attribute measurements. Generating data is not the bottleneck. Synthesizing it into decisions that are scientifically sound, regulatory defensible, and timely is.

DeepC was built for exactly this. As an AI co scientist for formulation development, DeepC's architecture processes, synthesizes, and surfaces critical information from complex pharmaceutical datasets, with the audit trails and data provenance that regulators require.

• Regulatory Intelligence: DeepC’s Research Agent mines and synthesizes pharmaceutical literature, regulatory databases, and clinical evidence across authoritative sources including FAERS, DailyMed (154,834+ drug labeling records), ClinicalTrials.gov, and PubMed. The same data infrastructure that powers pharmacovigilance signal detection powers formulation intelligence.
• Safety Signal Integration: DeepC’s platform surfaces safety signals from adverse event databases with severity scoring and confidence levels. For formulation scientists, this means understanding the safety profile of an active pharmaceutical ingredient and its excipients before making design decisions, not after adverse events accumulate.
• Audit Trails and Data Lineage: Every recommendation DeepC generates is traceable to its regulatory source. Excipient suggestions cite the FDA Inactive Ingredient Database. Safety flags reference the adverse event record. This is the data governance and documentation that Principle 6 of the joint FDA and EMA AI principles requires, built into the system’s architecture from the start.
• Quality by Computational Design: DeepC extends traditional Quality by Design into predictive CQA identification, virtual design space exploration, and risk informed formulation screening, all aligned with ICH guidelines. The same standards that govern pharmaceutical quality control in manufacturing apply to the AI formulation decisions DeepC supports.

The lesson of the Novo Nordisk crisis extends beyond pharmacovigilance reporting. It shows what happens when pharmaceutical data processing relies on manual procedures that are vulnerable to organizational pressure, human error, and capacity constraints. Formulation development faces the same vulnerabilities. When a scientist must manually review thousands of stability data points or cross reference excipient interactions across dozens of regulatory databases, the risk of missed signals is the same risk that led to three unreported deaths on semaglutide.

The pharmacovigilance system failed because it could not process the volume, complexity, and urgency of safety data for a drug class serving 11 million patients. Formulation development faces an analogous challenge: 3.6 million possible formulation combinations for a single compound, years of stability data, thousands of excipient interactions, and regulatory requirements that demand traceability for every decision. AI is not optional for either problem. The question is whether the AI is built specifically for pharmaceutical science, with the domain knowledge, regulatory grounding, and audit infrastructure the work demands, or whether companies will try to solve pharmaceutical problems with general tools and hope regulators do not notice.