Parallel Dossier Architecture
The sequential model of NDA preparation — build clinical data package, write study reports, write ISS/ISE, write clinical overview, assemble Module 2 — made sense when writing was the bottleneck. When AI can generate a first draft from a data summary in hours, the bottleneck shifts to coordination and review sequencing.
The parallel model
In a parallel dossier architecture, multiple document streams run simultaneously:
Stream 1 — Module 5 (Study Reports): Individual CSRs start generation as soon as statistical analysis packages are locked. AI drafts the narrative sections from the SAP, TFLs, and protocol. Biostatisticians review statistical sections; medical writers review narrative.
Stream 2 — Module 2.7 (Clinical Summary): Clinical Summary sections begin drafting from available data, with placeholders for not-yet-locked datasets. Updated automatically as new datasets come in.
Stream 3 — Module 2.5 (Clinical Overview): Benefit-risk narrative drafted in parallel with summary sections, updated as key data becomes available.
Stream 4 — Module 3 (Quality/CMC): CMC documentation proceeds on its own track and feeds into the integrated dossier at defined checkpoints.
The enabling technology
Parallel architecture requires two things AI makes possible: (1) first-draft generation from partial information, so documents don't need to wait for all upstream documents to be complete before they start, and (2) automated consistency checking across documents, so when a key number or finding changes in one document, it gets flagged and updated across all others.
Without AI, running these streams in parallel requires an enormous coordination burden — writers have to manually track every cross-reference. With AI handling consistency verification, the human burden drops to reviewing the exceptions, not managing the network.
Implementation requirements
Parallel architecture requires investment in three areas: a shared data repository where analysis outputs are deposited as they're produced, a document management system that can track cross-references across the dossier, and a workflow governance structure that defines who can unblock each document stream.
The first two are technical. The third is organizational, and it's usually the harder problem.