A pharmacovigilance writer in Hyderabad opens her queue on a Tuesday in May 2026, six weeks after the FDA's E2B(R3) cut-over made the structured electronic case safety report the only acceptable transmission format. Case number 2026-04471 is a serious unlisted adverse drug reaction: a fifty-eight-year-old woman on a marketed biologic developed acute interstitial nephritis requiring hospitalization. The intake fields are populated, the MedDRA terms are provisionally coded, and the clock is running, because a serious unexpected case has a fifteen-day expedited timeline. She pastes the structured intake into the enterprise pharmacovigilance model and asks for a case narrative for Section E.i. Twenty seconds later a fluent, chronological narrative appears, reading exactly like the hundreds of narratives a senior PV writer has produced. It states the patient's history, the drug exposure, the event, the dechallenge, and a causality conclusion of "probable." And in that last word lies the entire problem this lesson exists to address: the model wrote a causality assessment, and causality is precisely the judgment the writer, not the model, must own. This lesson is about drafting the narrative the model is good at, and protecting the four decisions it must never make.

What the ICSR Narrative Is, and Why E2B(R3) Raised the Bar

An Individual Case Safety Report, the ICSR, is the atomic unit of pharmacovigilance: a single patient, a single suspected adverse reaction to a medicinal product, captured in a structured format and transmitted to regulators. The case narrative, which lives in Section E.i of the E2B(R3) data element set, is the prose heart of that report. It is the clinical story that a safety physician reads to understand what happened, in what order, and whether the medicinal product plausibly caused it. Everything else in the ICSR is coded fields; the narrative is where the coded fields become a coherent clinical account that a human assessor can reason over.

E2B(R3) is the ICH standard that defines how that case data is electronically structured and transmitted, and as of 1 April 2026 it is the FDA-mandated format, replacing the older E2B(R2). The cut-over matters for AI-assisted drafting in two specific ways. First, E2B(R3) is a richer, more granular data model than its predecessor, with more structured fields, stronger expectations for the relationship between the coded data and the narrative, and tighter rules about what the narrative must contain and how it must align with the structured elements. A narrative that contradicts the coded fields, that names an event in prose that does not match the MedDRA-coded reaction term, is now a more visible defect because the structured layer makes the comparison mechanical. Second, the FDA's implementation has a phased shape: IND safety reports under the E2B(R3) standard from the 1 April 2026 date, with postmarketing ICSR submission via the FDA's Electronic Submissions Gateway NextGen pathway from 1 October 2026. The writer drafting a narrative in mid-2026 is working in a transition window where the format is mandated but the transmission rails are still being completed, which raises rather than lowers the premium on getting the narrative right the first time.

The narrative itself follows a stable clinical logic that the model has learned well: patient demographics and relevant medical history, the suspect product with dose and route and the dates of therapy, the adverse event with its onset relative to the drug, the clinical course including any interventions, the dechallenge (what happened when the drug was stopped) and rechallenge (what happened if it was restarted), concomitant medications and alternative explanations, the outcome, and the reporter's and the company's causality assessment. This is a chronological, factual structure, and it is exactly the kind of structured storytelling a language model produces fluently. That fluency is the opportunity and, for the four reserved judgments, the trap.

The Four Judgments the Human Owns and the Model Must Not Assert

There are four decisions embedded in an ICSR that determine its regulatory fate, and not one of them is a drafting task. They are clinical and regulatory judgments that carry the named qualified person's accountability, and the model's job is to assemble the facts that inform them, never to assert the conclusion. The first is seriousness. An adverse event is serious if it meets one of the ICH E2D criteria: death, life-threatening, hospitalization or prolongation of hospitalization, persistent or significant disability or incapacity, congenital anomaly, or another medically important condition. Seriousness drives the reporting timeline, and a case mis-graded as non-serious can miss its expedited window entirely. The model can surface that the patient was hospitalized; it cannot decide that the hospitalization meets the seriousness threshold, because that determination is a medical judgment with reporting consequences the writer is accountable for.

The second is listedness, also called expectedness: whether the reaction is already described in the reference safety information, the Company Core Data Sheet or the approved label. A reaction that is listed is a known risk; an unlisted reaction, like the acute interstitial nephritis in our case if it is not in the reference document, is a potential new signal and changes the reporting obligation. The model does not have authoritative access to your current reference safety information unless you load it, and even then, the determination of whether a coded reaction matches a listed term is a judgment about clinical equivalence that the model is not authorized to make. The third is causality: the assessment of whether the medicinal product caused the reaction, the "probable" the model wrote in our opening case. The fourth is the overall case validity and completeness assessment: whether the case has the four minimum criteria of a valid ICSR (an identifiable reporter, an identifiable patient, a suspect product, and a suspect reaction) and whether follow-up is required.

The discipline is to configure the model so that it drafts the factual narrative and explicitly defers these four. A well-built pharmacovigilance system prompt instructs the model to state the facts that bear on seriousness, listedness, and causality, and to flag each as a reserved human determination rather than asserting a conclusion. A narrative that ends "the company assessment of causality is probable" has had the model do the writer's job; a narrative that ends "the temporal relationship, positive dechallenge, and absence of alternative explanation are presented for the causality assessor's determination" has had the model do its own. The difference is the difference between a tool and a liability.

MedDRA Coding and the Discipline of the Lowest Level Term

MedDRA, the Medical Dictionary for Regulatory Activities, is the controlled terminology that turns a reporter's free-text description into a coded, regulator-comparable reaction term, and it is one of the genuinely strong applications of AI in pharmacovigilance, with an essential caveat about the level of coding. MedDRA is hierarchical: the Lowest Level Term (LLT) is the most granular, closest to the verbatim the reporter actually used; the Preferred Term (PT) is the level at which cases are aggregated for signal detection; above that sit High Level Terms, High Level Group Terms, and System Organ Classes. The coding rule that matters for ICSR fidelity is that the LLT should capture the reporter's verbatim as closely as possible, because the LLT preserves the original clinical nuance that aggregation to the PT can erase.

Consider the failure mode this prevents. A reporter describes "burning sensation when passing urine." The verbatim maps most faithfully to an LLT in the dysuria family. A model that over-aggregates, coding straight to a broader Preferred Term or, worse, choosing a clinically adjacent but distinct term, loses the specificity that a signal-detection algorithm needs and that a later assessor relies on to understand what was actually reported. The model is good at proposing LLT candidates from verbatim text, often better and faster than a human scanning the dictionary, but the selection of the correct LLT, especially disambiguating between clinically similar terms and respecting the MedDRA term selection points-to-consider, is a coding judgment a trained coder verifies. The pattern is the same as everywhere in this program: the model proposes, the human disposes, and the disposition is documented.

There is a specific E2B(R3) wrinkle here. Because the structured data model couples the coded reaction term tightly to the narrative, a mismatch between the MedDRA-coded reaction in the structured field and the event as described in the Section E.i narrative is now a mechanical, detectable inconsistency. If the narrative describes acute interstitial nephritis but the coded PT is a less specific renal term, the case is internally inconsistent in a way the structured layer exposes. The drafting discipline is to verify that the narrative's described event and the coded LLT and PT tell the same clinical story, and that the LLT honors the verbatim, before the case moves toward transmission.

WHO-UMC Causality: The Reasoning the Model Assembles but Does Not Conclude

The WHO-UMC causality assessment system is the framework many pharmacovigilance functions use to grade the relationship between a drug and a reaction, with categories of certain, probable or likely, possible, unlikely, conditional or unclassified, and unassessable or unclassifiable. Each category has defined criteria built around a small set of considerations: the temporal relationship between drug exposure and reaction onset, the plausibility of the drug as a cause given its known pharmacology, the dechallenge response, the rechallenge response where available, and the presence or absence of alternative explanations such as concomitant medications or the underlying disease. The structure of WHO-UMC is genuinely well-suited to AI assistance, because it is a reasoning scaffold over named facts, and the model is good at laying out the facts against the scaffold.

The legitimate, high-value use of the model is to assemble the causality reasoning transparently: to state the temporal relationship explicitly (the reaction began nine days after the first dose), to note the dechallenge result (the reaction resolved over two weeks after discontinuation), to flag the absence of rechallenge data, and to identify the alternative explanations that must be weighed (the patient was also taking a proton-pump inhibitor associated with interstitial nephritis). This assembled reasoning is exactly what a causality assessor needs in front of them, and producing it quickly and completely is real value. What the model must not do is collapse that reasoning into a category. The leap from "positive dechallenge, plausible temporal relationship, but a confounding concomitant medication associated with the same reaction" to a final grade of "possible" rather than "probable" is a weighing of evidence that the WHO-UMC framework deliberately leaves to a qualified human, because it requires judgment about the relative weight of competing considerations that no fixed rule resolves.

This is why the opening case is a problem. The model wrote "probable." But the confounding proton-pump inhibitor, itself a known cause of acute interstitial nephritis, is precisely the kind of alternative explanation that might drive a careful assessor to "possible" instead. The model did not weigh the confounder against the dechallenge; it produced the most statistically plausible causality word for a narrative with a positive dechallenge, which in the corpus is often "probable." That is pattern completion, not causality assessment, and the two look identical on the page. The named human owns the weighing, and the audit trail records that the human, not the model, made the call.

The Serious Unlisted Case and the Clock It Starts

The case in front of us is serious (hospitalization) and unlisted (acute interstitial nephritis, assumed not in the reference safety information), and that combination has a specific regulatory consequence: it is an expedited reportable case with a fifteen-day clock from the day the company became aware of the minimum valid information. The interaction between AI drafting and the clock is where the value and the danger both concentrate. The value is real: a narrative that used to take a writer forty-five minutes to compose from the intake fields can be drafted in a structured form in under a minute, and across a queue of a hundred-plus cases a day with a dozen expedited reports, that compression is the difference between meeting timelines and missing them.

The danger is that the clock creates pressure to accept the fluent draft as finished, and the four reserved judgments are exactly the judgments most tempting to let the model's confident wording stand under time pressure. The defensible workflow inverts the temptation: the model's speed is spent on the factual narrative, which buys the writer time to do the four judgments properly rather than to skip them. The writer reads the assembled narrative, verifies that every stated fact traces to the intake source (the model can transcribe a date wrong or carry a dose from a concomitant medication into the suspect product), confirms the MedDRA coding honors the verbatim, lays the causality facts against WHO-UMC and makes the grade themselves, confirms seriousness against E2D and listedness against the current reference safety information, and records that these determinations were human-made. The narrative is faster; the judgment is not delegated. That is the whole shape of defensible AI-assisted ICSR drafting.

One more point specific to the serious unlisted case: because it is a potential new signal, its narrative will be read again, by the signal-detection function, by the medical assessor preparing the periodic report, and possibly by a regulator querying the case. A narrative that quietly asserted causality the human did not endorse, or that lost the verbatim specificity in coding, does not merely risk this one report; it pollutes the signal. The fidelity of one serious unlisted narrative is leverage on the entire safety picture of the product, which is why the reserved judgments are not bureaucratic ceremony. They are the integrity of the pharmacovigilance system, expressed at the level of a single case.

Building the Narrative as a Defensible Workflow

The workflow that makes AI-assisted ICSR narrative drafting defensible separates the model's contribution from the human's at every step and records the separation, the same GxP discipline that governs every artifact in this program. The model contributes the chronological factual narrative assembled from the structured intake, the proposed MedDRA LLT candidates from the verbatim, and the transparent layout of the WHO-UMC considerations. None of these is a conclusion; all of them are assembly. The model is, in the best sense, a fast and tireless drafter of the parts of the case that are factual and structural.

The human contributes the four reserved judgments, the verification that every narrative fact traces to source, the selection and confirmation of the correct LLT and PT, the causality grade weighed against WHO-UMC, the seriousness determination against E2D, and the listedness determination against the live reference safety information. Around this, the audit trail records the model and version, the prompt and system prompt, the intake source loaded, the human edits and the human determinations, and the named writer and assessor, so that the case is defensible if a regulator asks how AI was used in its preparation. The qualified person for pharmacovigilance, the QPPV, owns the safety of the product and cannot delegate that ownership to a vendor's model, and the case record must show, on its face, that the reserved judgments were human.

The mature posture is to treat the model as the writer's fastest assistant and the strictest possible reader of its own output. It drafts the narrative in seconds; the writer then reads that narrative the way a regulator would, as a set of factual claims that must trace to the intake and four reserved judgments that must carry a human name. The twenty-second draft is the start of a defensible case, not the end of one, and the writers who internalize that distinction will clear larger queues with fewer errors than the writers who either refuse the tool or trust it past the point where the human judgment lives.

Key Takeaways

• The ICSR narrative (E2B(R3) Section E.i) is the prose heart of the case, and as of the 1 April 2026 FDA cut-over the structured data model makes narrative-to-coded-field mismatches mechanically detectable. IND safety reports use E2B(R3) from that date; postmarketing ICSRs transmit via the ESG NextGen pathway from 1 October 2026, a transition window that raises the premium on getting the narrative right.
• Four judgments are reserved for the human and must never be asserted by the model: seriousness against ICH E2D criteria, listedness against the reference safety information, causality, and overall case validity. The model assembles the facts that inform them and flags each as a reserved determination; it does not conclude.
• MedDRA coding must honor the Lowest Level Term, which preserves the reporter's verbatim specificity. The model proposes LLT candidates well, but over-aggregation to a broader Preferred Term or a clinically adjacent term erases the nuance that signal detection and later assessors rely on, so a trained coder verifies the selection.
• WHO-UMC causality is a reasoning scaffold the model assembles but does not conclude. Stating the temporal relationship, dechallenge, rechallenge, and alternative explanations is high-value assembly; collapsing them into "probable" or "possible" is a weighing of competing evidence the framework reserves for a qualified human. A confounding concomitant medication is exactly what the model's pattern-completed grade tends to ignore.
• The expedited fifteen-day clock concentrates both value and danger. AI speed should buy time to do the four judgments properly, not pressure to let the fluent draft stand. A serious unlisted narrative is read again by signal detection and the QPPV, so its fidelity is leverage on the entire product safety picture, and the QPPV's accountability cannot be delegated to a model.