Somewhere in every regulatory publishing team, a person spends a meaningful part of their week dragging files into folders. Source documents arrive — from authors, from partners, from a document management system — and someone who knows the eCTD structure by heart places each one into its section: this stability report to 3.2.P.8.3, this batch record to 3.2.P.5.4, this CV to 1.3. It is skilled work, it is repetitive, and at dossier scale it is a bottleneck measured in days.

The obvious modern answer is “use AI to classify the documents.” We built that — and then we deliberately demoted it to the last resort. This post explains why DnXT’s auto-assign engine is deterministic first and AI last, and why that ordering matters more than it might seem.

What the Data Said

Before building the engine, we mined the placement decisions embedded in real published dossiers: 3,081 actual eCTD sequences containing 51,029 placed document leaves. Every leaf in a published sequence is a ground-truth answer to the question “where does a document like this go?” — a placement decision made by a professional and accepted by an agency.

The mining produced 1,384 reusable placement patterns connecting observable document features — source folder path conventions and filename structures — to eCTD target sections. The headline finding: the overwhelming majority of real-world placement decisions are not judgment calls at all. They are conventions, applied consistently for years, sitting unexploited in every company’s submission archive. A file that arrives in a folder called stability with a filename matching a long-tail pattern your organization has used 200 times before does not need a neural network to find its home. It needs the pattern looked up.

The Three-Layer Engine

The auto-assign engine classifies each incoming document through three layers, cheapest and most reliable first:

  • Layer 1 — Metadata. If the document arrives with structured metadata (a document type from a connected DMS, an explicit section tag), placement follows directly. Confidence is highest, cost is zero.
  • Layer 2 — Mined patterns. The 1,384-pattern library evaluates the document’s source path and filename. A pattern hit yields a placement with a quantified confidence score derived from how consistently that pattern mapped to that section across the mining corpus.
  • Layer 3 — AI. Only documents that fall through both deterministic layers reach content-based AI classification, which reads the document and proposes a section — clearly marked as an AI proposal, with lower default confidence.

Every placement, from any layer, carries two things into the review UI: a confidence score and a human-readable “why” — “matched pattern: filenames like *-stab-summary from folder /CMC/stability/ → 3.2.P.8.3 (seen 214 times)”. The publisher reviewing the assignment sees the reasoning, not just the verdict.

Why Deterministic-First Beats AI-Only

The AI-only architecture — send every document to a model, accept its classification — demos beautifully and degrades badly. Four reasons we inverted it:

  • Repeatability. The same file must classify identically today and next month. Pattern lookups are perfectly repeatable; model outputs drift with model versions, prompts and sampling. In a validated process, “it usually says the same thing” is not a property you want load-bearing.
  • Explainability. “Seen 214 times in your own submission history” is an explanation a QC reviewer and an auditor both accept instantly. A model’s rationale is, at best, a plausible narrative.
  • Failure economics. When a pattern misfires, the fix is visible and permanent: correct the pattern, and the error class disappears. When a model misfires, you have a prompt-engineering project.
  • Cost and latency. Deterministic layers are effectively free and instant at any scale. Reserving AI for the genuinely ambiguous tail means the expensive layer runs on the fraction of documents that actually need it.

None of this is anti-AI — Layer 3 exists because the tail is real: novel document types, partner files with alien naming conventions, scanned legacy content. The design principle is narrower: use AI for the residual, not the routine. It is the same philosophy behind the rest of DnXT’s AI document intelligence — AI assists, deterministic systems anchor, and humans decide.

The Human Stays in the Loop

Auto-assignment produces proposals, not fait accompli. The publishing workflow presents assignments sorted by confidence: high-confidence placements are bulk-confirmable in one pass, low-confidence and AI-layer placements are flagged for individual attention. The person who used to drag every file now reviews exceptions — which is both faster and a better use of the expertise that made them good at the dragging.

Confirmed placements flow into the dossier through the same bulk placement machinery the publishing engine exposes, and downstream eCTD validation still checks everything — placement automation upstream never substitutes for validation downstream.

The Compounding Asset

Because Layer 2 is mined from submission history, the engine improves with use in a way that is auditable rather than mystical: each newly published sequence adds placement decisions to the corpus, the miner refreshes the pattern library, and coverage of your organization’s conventions deepens. Two companies running the same engine end up with different pattern libraries — each fitted to its own naming habits and document flows. The knowledge that used to live only in your senior publisher’s head becomes an inspectable, versioned asset.

Frequently Asked Questions

What happens with a document the engine has never seen anything like?

It falls through to AI classification, arrives flagged with low confidence, and a human places it. That decision then becomes part of the history the next mining pass learns from.

Can we inspect or edit the pattern library?

Patterns are readable rules — path and filename structures mapped to sections with occurrence counts. They can be reviewed, and corrections propagate to all future assignments.

Does this replace document classification in our DMS?

No — it consumes DMS metadata as its highest-confidence layer. Good upstream classification makes auto-assignment stronger, not redundant.

Watch a Dossier Assemble Itself

The engine is best understood watching a folder of real documents land in their sections with reasons attached. Book a demo — bring a sample document set, and we will show you the three layers, the confidence scores, and the review pass live.