A pharmacovigilance lead is eleven weeks from the data-lock date of a Periodic Safety Update Report for a marketed biologic, and the document that has to be produced is the most integrative artifact in all of pharmacovigilance: a structured, ICH E2C(R2)-aligned account of the worldwide safety experience of the product over the reporting interval, pulling together case counts, signal evaluations, study results, literature, and the cumulative exposure into a single benefit-risk evaluation that a health authority will read as the company's considered judgment on whether the product's benefits still outweigh its risks. The PSUR, which the PBRER format defines, is not a case-level document like an ICSR; it is an aggregate document, and that changes the AI problem entirely, because the failure modes here are not a fabricated dose in one narrative but an inconsistency across a sixty-page report, a signal section that contradicts the benefit-risk conclusion, or a number in Section 16 that does not match the same number in Section 6. This lesson designs the end-to-end AI workflow for PSUR and PBRER production from database lock through signal aggregation to the benefit-risk integration, with cross-version and cross-section consistency as the structural control and with the benefit-risk integration drawn explicitly as the boundary AI does not cross.

What the PBRER Actually Is, and Why Aggregation Changes the AI Problem

The Periodic Benefit-Risk Evaluation Report, the format the ICH E2C(R2) guideline defines and that the PSUR uses, is the periodic summary of the worldwide safety information for an authorized product over a defined reporting interval, structured into the standard sections that run from the introduction and worldwide marketing-authorization status through estimated exposure, the presentation of individual case and aggregate data, signal evaluation, and the integrated benefit-risk analysis that the report exists to deliver. Its defining characteristic, the one that reshapes the AI problem, is that it is an aggregate and integrative document: it does not tell the story of one case but synthesizes the entire safety experience of the interval into counts, evaluations, and a single reasoned conclusion. The reader, a health-authority assessor, reads the PBRER not to verify one fact but to follow an argument, and the argument runs across the whole document, so a number stated in the exposure section must be the number the signal section reasons from, and the signal evaluations must be the ones the benefit-risk section integrates.

This aggregate nature is exactly why the AI failure modes shift. In a case narrative, the danger is a fabricated clinical fact; in a PBRER, the danger is incoherence across sections, a signal characterized as closed in the signal section but still listed as an open concern in the benefit-risk discussion, a cumulative case count in one table that does not reconcile to the count reasoned from elsewhere, a conclusion in Section 18 that does not follow from the evidence assembled in Sections 14 through 17. The PBRER is long, it is assembled from many sources, and a model drafting its sections independently will produce locally fluent prose that does not cohere globally, which means the controlling discipline of this workflow is not per-claim citation but cross-section and cross-version consistency. The workflow exists to make the document cohere, because an assessor who finds the signal section and the benefit-risk section disagreeing reads the disagreement as a failure of the company's pharmacovigilance reasoning, not as a drafting slip.

Database Lock and the Aggregate-Data Foundation

The workflow begins at database lock, the defined cutoff at which the safety database is frozen for the reporting interval so that the aggregate data the report presents, the case counts, the cumulative and interval tabulations, the exposure estimates, derives from a single stable snapshot rather than a moving target. This is the foundation of the entire report, because every number in the PBRER traces to the locked database, and a report that mixes data from before and after the lock, or that states a count the locked data does not support, has a defect at its root that propagates into every section that reasons from the number. AI assists this stage by generating the aggregate tabulations from the locked data, the interval and cumulative case counts by seriousness and by reaction, the exposure estimates, and the line listings and summary tabulations that the report's data sections present, and this generation from structured source is a genuine acceleration over manual tabulation.

The discipline at this foundation is that every aggregate figure in the report must trace to the locked database, and the workflow generates the report's numbers from the locked source rather than allowing the model to restate them from prose, because a number the model carries from a draft or a prior version is a number that may not match the current locked data. The exposure estimate deserves particular attention, because the entire interpretation of the case counts depends on the denominator, the estimated patient exposure over the interval, and a benefit-risk argument that reasons about a reporting rate is only as sound as the exposure estimate beneath it, which is a methodological judgment the model supports but does not own. The workflow therefore anchors every count and rate to the locked database and the documented exposure methodology, so that the data foundation the rest of the report reasons from is stable, traceable, and consistent, because a PBRER whose numbers do not trace to the lock cannot defend its conclusions.

Signal Aggregation and AI-Assisted Section 16 and 17 Drafting

With the data foundation in place, the workflow aggregates the signal information for the interval, assembling the new signals, the ongoing evaluations, and the closed signals into the signal section that ICH E2C(R2) structures, and drafting the signal-evaluation and risk-characterization content that Section 16 presents. AI assists this drafting well, because the signal section is a structured synthesis of the signal-management activity over the interval, and a model that drafts a consistent account of each signal's status, evaluation, and conclusion from the underlying signal records is a real acceleration over composing the section by hand. The same applies to the benefit characterization in Section 17, where the model assembles the efficacy and effectiveness information that the benefit side of the evaluation rests on, drawing from the studies, the literature, and the established benefit profile into a structured benefit account.

The discipline that makes this drafting defensible is that each signal's representation in the report must be consistent with its representation in the signal-management system and consistent across every section of the report that mentions it, because a signal is referenced in the signal section, may be discussed in the benefit-risk integration, and may relate to a risk in the risk-management plan, and those references must agree. The workflow extracts the structured signal status, the new, ongoing, and closed determinations, from the signal-management source and generates the section content from that structured status rather than from prose, so that a signal cannot be characterized as closed in one section and open in another. The drafting of Sections 16 and 17 is where AI delivers the most acceleration in PBRER production, because the content is voluminous and structured, but the acceleration is defensible only because the signal statuses and the benefit claims are anchored to their sources and reconciled across the document, not regenerated section by section into local fluency that does not cohere.

The Benefit-Risk Integration: The Boundary AI Does Not Cross

The integrated benefit-risk evaluation is the section the entire PBRER exists to deliver, and it is the boundary this workflow draws hardest, because the benefit-risk integration is the reasoned judgment that weighs the assembled benefits against the assembled risks and concludes whether the balance remains favorable, and that judgment is a human determination the model supports but does not make. The reason is the same reason the program holds causality and comparability conclusions as human domains: the benefit-risk integration is not a summary of the evidence but a weighing of it, requiring clinical and regulatory judgment about the significance of a signal, the robustness of the benefit, the adequacy of the risk minimization, and the residual uncertainty, and a model can assemble the inputs to that judgment without being able to make the judgment itself. The workflow positions the AI as the assembler of the benefit-risk inputs, the structured benefit profile, the characterized risks, the signal conclusions, the effectiveness of risk minimization, and as the drafter of the descriptive scaffolding, but the integrated conclusion, the determination that the benefit-risk balance remains favorable or has changed, is written and owned by the qualified person who is accountable for the report.

This boundary is not a limitation to be engineered away but the defining accountability of the document, because the PBRER's benefit-risk conclusion is the company's formal position on whether the product should remain on the market under its current conditions, and that position cannot be delegated to a pattern-completer. The danger the workflow specifically guards against is the model producing a fluent, confident benefit-risk conclusion that reads like a considered judgment, because such a conclusion is the most plausible-looking and most dangerous output the model can generate here: it has the form of the judgment without the judgment, and a benefit-risk conclusion that does not follow from the integrated evidence, or that overstates the favorability of the balance, is the failure that an assessor and a signal of regulatory concern are most attuned to. The workflow therefore requires that the benefit-risk integration be authored by the accountable qualified person, that the conclusion be traceable to the integrated evidence the report assembles, and that the AI's role be limited to assembling and structuring the inputs that the human integrates, with the integration itself recorded as a human authorship event.

Cross-Version and Cross-Section Consistency as the Structural Control

The validation control that defines this workflow is consistency, checked across sections within the report and across versions of the report over time, because the PBRER is a recurring document and each new edition must be consistent with its predecessors except where the new interval's evidence justifies a change. The cross-section consistency check reconciles every quantity and every signal status that appears in more than one section, confirming that the cumulative case count in the data section matches the count reasoned from in the signal section, that a signal characterized as closed in Section 16 is not treated as open in Section 18, and that the benefit-risk conclusion follows from the benefits and risks the report assembled, flagging any quantity that appears with two values or any signal that appears with two statuses. This is the PBRER equivalent of the cross-document consistency check that runs through the entire program, scaled to the internal coherence of a long aggregate document where the argument depends on every section agreeing with every other.

The cross-version consistency check compares the current report against the prior editions, confirming that a risk characterized one way in the last PSUR is either characterized the same way or changed with a documented justification, that the benefit-risk conclusion's trajectory across editions is coherent, and that a signal closed in a prior report is not silently reopened or a prior conclusion silently contradicted without explanation. This matters because an assessor reads the current PSUR against the cumulative history of the product, and an unexplained inconsistency between editions, a risk that quietly changed characterization, a conclusion that reversed without rationale, reads as either an error or a concealment, both of which damage the report. The workflow extracts the structured claims, the counts, the signal statuses, the risk characterizations, and the conclusions, and diffs them within the report and against the prior editions, surfacing every inconsistency for resolution before the report is finalized, because the inconsistencies are invisible to a section-by-section read and visible only when the same claim is tracked across the document and across time.

The Validation Spec, the Audit Trail, and What the Assessor Asks

The Level 3 deliverable is the validated workflow and the audit trail that make the PBRER defensible to a health-authority assessment, not the report prose alone. The validation spec states the intended use, the production of an ICH E2C(R2)-aligned PBRER from a locked safety database, the fitness-for-purpose statement aligned to the FDA-EMA principle, and the acceptance criteria, which are the traceability of every aggregate figure to the locked database, the cross-section consistency of every shared quantity and signal status, the cross-version consistency against prior editions, and the human authorship of the benefit-risk integration. The audit trail captures the database-lock identifier and date, the model and version, the system prompt identity, the temperature, the timestamp, the locked-data source for every tabulation, the signal-management source for every signal status, the consistency-check logs within and across versions, the AI-assembled inputs to the benefit-risk evaluation, and the named qualified person who authored the integrated conclusion and signed the report.

The human handoff is explicit and the accountability is unambiguous. The model generates tabulations from the locked data, drafts the signal and benefit sections from their structured sources, and assembles the benefit-risk inputs, but the qualified person responsible for the report authors the integrated benefit-risk conclusion, a named scientist confirms the signal statuses and the aggregate figures against their sources, and a named author signs the report with the consistency logs and the lock reference attached. The benefit-risk judgment, the determination that the product's benefits continue to outweigh its risks under the current conditions of use, remains the company's formal pharmacovigilance position and the qualified person's accountability; the workflow accelerates the assembly and hardens the consistency, but it does not own the conclusion. When an assessor asks the company to reconcile a signal's status across the report or to show that the benefit-risk conclusion follows from the assembled evidence, the consistency logs, the locked-data traceability, and the recorded human authorship of the integration already hold the answer. The model assembles the aggregate; the named qualified person integrates the benefit and the risk and certifies the conclusion.

Key Takeaways

• The PBRER is an aggregate, integrative document, which changes the AI failure mode from a fabricated fact to incoherence across the report. An assessor follows an argument that runs across the whole document, so the danger is a signal characterized as closed in one section and open in another, or a count that does not reconcile across sections, making cross-section and cross-version consistency the controlling discipline.
• Database lock is the data foundation, and every aggregate figure must trace to it. The report's counts, tabulations, and exposure estimates derive from a single frozen snapshot, and the workflow generates the numbers from the locked source rather than restating them from prose, because a number that does not trace to the lock has a defect at the root that propagates into every section that reasons from it.
• AI delivers the most acceleration in drafting the signal and benefit sections, anchored to their structured sources. Sections 16 and 17 are voluminous and structured, so the model drafts them well, but each signal's status is generated from the signal-management source and reconciled across the document, so a signal cannot be closed in one section and open in another.
• The benefit-risk integration is the boundary AI does not cross. The integrated conclusion is a weighing of evidence requiring clinical and regulatory judgment, not a summary, so the model assembles the inputs while the accountable qualified person authors the conclusion, because a fluent benefit-risk conclusion that has the form of judgment without the judgment is the most dangerous output the model can produce here.
• Cross-version and cross-section consistency, validated and audited, make the PBRER defensible. The workflow diffs every shared count, signal status, risk characterization, and conclusion within the report and against prior editions, because an unexplained inconsistency between sections or editions reads as an error or a concealment, and the consistency logs, locked-data traceability, and recorded human authorship are the record that answers a health-authority assessor.