There is a moment in every pharmacovigilance department that the rest of the company never sees. A case lands that is serious, unexpected, and at least possibly related to the product, and a clock starts that the organization cannot stop, slow, or negotiate: fifteen calendar days from the day day-zero awareness was established to a validated, submitted expedited report. The expedited 15-day report is the fastest-turn artifact in all of drug safety, faster than a 483 response, faster than an information request, faster than anything in regulatory affairs, because the obligation is to the patient population currently taking the drug. When a writer reaches for AI on a 15-day case, the temptation and the trap arrive together: the model can compress the drafting from forty-five minutes to ninety seconds, which is exactly the relief a stretched queue needs, and the clock is exactly the pressure that makes a writer accept a draft they should have interrogated. This lesson is about using AI to win back time on the 15-day report without surrendering the one decision the qualified person cannot delegate, and without letting speed become the thing that breaks the case.

What Triggers the Fifteen-Day Clock and Why It Cannot Slip

The 15-day expedited report exists because some adverse reactions are too consequential to wait for a periodic report. Three conditions must combine to trigger it, and the combination, not any single condition, is what starts the clock. The reaction must be serious, meeting one of the ICH E2D and regulatory seriousness criteria: death, life-threatening, hospitalization or its prolongation, persistent or significant disability, congenital anomaly, or another medically important condition. The reaction must be unexpected, meaning unlisted in the reference safety information, the approved label or Company Core Data Sheet, so the event is not already a characterized risk of the product. And there must be a reasonable possibility of a causal relationship to the product. When all three are present, the case is expedited.

The regulatory anchors differ by jurisdiction but converge on the same fifteen-day window. In the United States, postmarketing expedited reporting for marketed drugs sits under 21 CFR 314.80, which requires the applicant to report serious and unexpected adverse experiences within fifteen calendar days. In the investigational space, 21 CFR 312.32 governs IND safety reports, with a tighter seven-calendar-day window for unexpected fatal or life-threatening suspected adverse reactions and fifteen calendar days for other serious unexpected reactions. In the European Union, EU Good Pharmacovigilance Practices Module VI governs the management and reporting of individual cases, with the same fifteen-day expedited timeline for serious cases and a tighter window for the most severe. The writer must know which regime governs the case in front of them, because the clock length and the exact triggering definitions are regime-specific, and a 15-day report submitted on day sixteen is a compliance failure regardless of how good the narrative is.

The defining property of this clock is that it does not slip for any reason internal to the company. The TLF is late, the assessor is on leave, the system is down, the case is complex: none of it moves the date. Day-zero is the day any employee or agent of the company first received the minimum information that constitutes a valid case, and the count runs in calendar days, not business days. This unforgiving structure is why AI is attractive here and why it is dangerous: the time AI saves is real and immediately useful against a clock that cannot be extended, and the same clock is what tempts the writer to treat the model's fluent draft as a finished report rather than a fast first draft.

Where AI Genuinely Saves Time on a 15-Day Report

The honest case for AI on the expedited report is specific and substantial, and it concentrates in the assembly and drafting work that is mechanical rather than judgmental. The model is fast and reliable at converting the structured intake into a chronological case narrative: taking the demographics, the suspect product with dose and dates, the reaction with onset timing, the clinical course, the dechallenge, the concomitant medications, and the outcome, and rendering them into the ordered prose that the Section E.i narrative requires. This is the single largest time sink in manual 15-day drafting, and it is exactly the structured-storytelling task the model does well.

Beyond the core narrative, the model accelerates several adjacent tasks. It proposes MedDRA Lowest Level Term candidates from the reporter's verbatim, speeding the coding step that a trained coder then confirms. It assembles the WHO-UMC causality considerations transparently, laying out the temporal relationship, the dechallenge and rechallenge status, and the alternative explanations, so that the human assessor has the reasoning in front of them rather than having to reconstruct it. It drafts the structured-field content for the E2B(R3) data elements consistently with the narrative, reducing the manual transcription that introduces narrative-to-coded-field mismatches. And it can run a completeness check, flagging which of the four minimum valid-case criteria are present and which intake fields are missing and may require follow-up. Across a queue where a dozen of a hundred-plus daily cases are expedited, this compression is the difference between a department that meets its 15-day on-time rate and one that does not.

The mental model is that AI converts the writer's scarcest resource on an expedited case, the minutes between day-zero and the submission deadline, from drafting minutes into judgment minutes. The work that has to be fast is the assembly; the work that has to be right is the judgment. A well-built workflow uses the model to finish the assembly in seconds so the human has the full fifteen days, minus processing, to spend on the parts of the case that determine whether it is reportable, what it should say, and who is accountable for the conclusions. Used this way, AI does not shortcut the 15-day report. It funds the judgment the 15-day report demands.

The QPPV Decision That Cannot Be Delegated

At the center of the expedited report is a determination that no model may make and no vendor may absorb: the assessment of whether the case meets the reportability criteria, and the causality and expectedness judgments that drive it. This determination ultimately sits within the accountability of the Qualified Person for Pharmacovigilance, the QPPV in the EU regime and the equivalent responsible safety officer elsewhere, whose name attaches to the company's pharmacovigilance system and to the integrity of its reporting. The QPPV's accountability is a regulatory construct precisely because someone must be answerable for the judgment that a case is or is not an expedited report, and that someone is a named human, not a configured tool.

The reason this cannot be delegated is not bureaucratic caution; it is the structure of the judgment itself. Reportability turns on three assessments that each require weighing. Seriousness requires deciding whether a clinical event meets a threshold, which can be genuinely ambiguous, a hospitalization that was precautionary, a condition that is medically important but not on the standard list. Expectedness requires comparing the reaction to the reference safety information and judging clinical equivalence, deciding whether a reported event is or is not the same as a listed term. Causality requires weighing the temporal relationship and dechallenge against confounders and alternative explanations. Each of these is a judgment under uncertainty, and the model, which produces the most statistically plausible answer rather than a weighed one, will generate a confident determination that has the form of a judgment without the substance of one. A model that outputs "this case is expedited-reportable, causality possible, reaction unexpected" has produced three reserved judgments in the costume of analysis, and a writer who lets that stand under deadline has allowed the model to make the QPPV's call.

The discipline is to configure the workflow so the model never states the reportability conclusion. It presents the seriousness facts against the E2D criteria and defers the determination. It presents the reaction alongside the relevant reference safety information and defers the expectedness call. It assembles the causality considerations and defers the grade. The human then makes each determination, and the case record shows, attributably, that the human made it. The 15-day clock makes this discipline harder, because the pressure to accept the model's confident conclusion is highest when time is shortest, which is exactly why the workflow must enforce the deferral structurally rather than relying on the writer to remember it at minute fourteen of day fourteen.

The Failure Mode: When Speed Becomes the Thing That Breaks the Case

The characteristic failure of AI-assisted expedited reporting is not a dramatic hallucination. It is a quiet acceptance, under time pressure, of a fluent draft whose reserved judgments were never actually made by a human. The case is serious, the clock is at day twelve, the queue has eleven other cases, and the model's draft reads cleanly and ends with a tidy causality grade and an expectedness call. The writer, trusting the fluency and pressed by the deadline, validates the case and submits. The report goes out on time and is wrong in a way that on-time submission cannot fix, because the determination that made it reportable, or that should have made it reportable and did not, was a pattern-completed default rather than a human judgment.

Consider the two directions this fails. In one direction, the model under-calls: it grades a borderline case as expectedness "listed" or causality "unlikely," the writer accepts it under pressure, and a case that should have been an expedited report is instead handled routinely, so a potential new signal is submitted late or not as expedited at all. This is the more dangerous direction, because the entire purpose of the expedited mechanism is to surface new serious risks fast, and an under-called case defeats it silently. In the other direction, the model over-calls or mis-states a fact, the report is submitted with an error, and the company has put an inaccurate serious-case assessment into the regulatory record under its own name. Both directions share a root cause: the human treated the model's speed as a reason to skip the judgment rather than as the means to afford it.

There is a second, subtler failure specific to the deadline. Because the model can transcribe the narrative incorrectly, swap a date, carry a concomitant medication's dose into the suspect product, mis-state the dechallenge, a writer racing the clock may submit a narrative whose facts do not match the intake. On a non-expedited case there is time to catch this in a later review pass; on a 15-day case the later pass may not exist, so the verification that every narrative fact traces to the intake source has to happen inside the compressed window. The discipline that protects the case is to spend the AI-saved minutes on exactly this reconciliation and on the reserved judgments, not on clearing the next case in the queue. The time AI gives back is meant to be reinvested in the case, not extracted from it.

Building the Expedited Workflow Against the Clock

A defensible expedited workflow is designed around the clock rather than in spite of it, and it sequences the human and model contributions so that judgment is never the thing that gets compressed. The opening move, on day-zero, is the human establishment of validity: confirming the four minimum criteria and starting the clock with an attributable timestamp, because the clock's start is itself a determination the record must show. The model then performs the assembly in seconds: the chronological narrative from intake, the LLT candidates, the WHO-UMC fact layout, the structured E2B(R3) fields, and a completeness flag for missing data that drives the follow-up request.

The human then spends the bought time on the three reserved determinations and the fact reconciliation. Every fact in the narrative is checked against the intake source. The MedDRA coding is confirmed to honor the verbatim and to match the narrative. Seriousness is decided against E2D, expectedness against the current reference safety information, and causality against the assembled WHO-UMC considerations, each weighed by the named human. The reportability conclusion is made by the human and recorded as human-made. Around all of this, the audit trail captures the model and version, the prompt and system prompt, the intake loaded, the human edits and determinations, the day-zero timestamp, and the names of the writer and the assessor, so that the case is defensible if a regulator asks how AI was used in a report submitted under the company's name against a non-negotiable clock.

The strategic point is that the 15-day report is the case that most rewards getting the human-AI division exactly right, because it is where the clock makes the temptation to over-delegate strongest and the cost of over-delegating most immediate. A department that uses AI to assemble fast and to fund the judgment will clear its expedited queue on time with the reserved judgments intact. A department that uses AI to skip the judgment will clear the queue on time and discover, at the next inspection or the next signal review, that on-time submission of an un-judged case is not compliance, it is a faster path to a finding. The clock cannot be slowed. The judgment cannot be delegated. AI, used with discipline, is how a writer honors both at once.

The 15-Day Report as the Stress Test of the Whole Workflow

The expedited report is worth dwelling on beyond its own importance, because it is the stress test that reveals whether a pharmacovigilance function's AI practice is real or performative. Any workflow looks disciplined on a routine case with three weeks of slack. The discipline is only proven under the 15-day clock, where every shortcut is tempting and every corner cut is invisible until it is found. A function that maintains the human reserved judgments, the fact reconciliation, and the audit trail under expedited pressure has a practice that will hold everywhere; a function whose discipline dissolves under the clock has a practice that was never load-bearing, only decorative.

This is why the expedited report deserves its own lesson rather than a paragraph in the narrative-drafting lesson. The narrative skills are the same; the conditions are not. The 15-day report adds the unforgiving clock, the immediate patient-safety stakes, and the maximum temptation to let the model's speed substitute for the human's judgment. Mastering it means building a workflow whose deferral of the reserved judgments is structural and whose fact reconciliation happens inside the window, so that the discipline does not depend on a tired writer's willpower at the end of a long queue. The writer who can run a defensible 15-day report under AI assistance can run anything in pharmacovigilance, because they have proven the discipline holds where it is hardest to hold.

Key Takeaways

• The expedited 15-day report is the fastest-turn artifact in pharmacovigilance, triggered by the combination of serious, unexpected, and reasonably possibly causal. It sits under 21 CFR 314.80 for marketed drugs, 21 CFR 312.32 for IND safety reports (with a seven-day window for fatal/life-threatening), and EU GVP Module VI, all converging on a fifteen-calendar-day clock from day-zero awareness that cannot slip for any internal reason.
• AI genuinely saves time on the assembly, not the judgment. The model drafts the chronological narrative, proposes MedDRA LLT candidates, assembles the WHO-UMC considerations, and runs completeness checks, converting the writer's scarcest resource from drafting minutes into judgment minutes. Used well, AI funds the judgment the 15-day report demands rather than shortcutting it.
• The reportability determination cannot be delegated to a model or a vendor; it sits in the QPPV's accountability. Seriousness, expectedness, and causality are each judgments under uncertainty, and the model produces a confident default with the form of a judgment but not the substance. The workflow must enforce deferral structurally so the human makes and records each determination.
• The characteristic failure is quiet acceptance under deadline, not dramatic hallucination. An under-called case submitted on time defeats the entire purpose of the expedited mechanism by hiding a potential new signal, and a mis-transcribed fact enters the regulatory record under the company's name. On-time submission of an un-judged case is not compliance; it is a faster path to a finding.
• The 15-day report is the stress test of the whole PV AI practice. Discipline that holds under the unforgiving clock holds everywhere; discipline that dissolves under it was decorative. Spend the AI-saved minutes on fact reconciliation and the reserved judgments inside the window, capture day-zero and every determination in the audit trail, and the writer who masters this can run anything in pharmacovigilance.