Draft Module 3 from what you have already filed and what is true now
An agent that retrieves the structure and language of your prior submissions, pulls the current data from the systems that hold it, and returns a drafted section where every number is cited to the record it came from.
The content already exists. Assembling it is the job.
A supplement needs a 3.2.S.2.2 section. Nothing about it is novel: you have filed adjacent sections before, the process description exists in a development report, the validation data is in the LIMS and the MES, and the specifications are approved and current. The work is a person rereading three prior filings to match the established language, transcribing current numbers into a document by hand, and then a second person checking every transcribed figure against its source.
That takes two to four weeks for a first draft, and a meaningful share of it is transcription and numeric QC rather than regulatory judgment. It is also where the errors live: a number typed from a table that was superseded, a specification quoted from a version that is no longer effective, a batch list that is missing the lot manufactured last month.
The scarce resource on your team is regulatory judgment. Very little of that fortnight uses any.
A section drafted in your own established language.
The agent is given a section to draft. It retrieves the corresponding sections from your prior submissions to establish structure and phrasing, pulls the current data from the systems of record, and returns a draft that follows your own conventions, accompanied by a citation table mapping every figure in the text back to the record it was queried from.
Prior filings set the form. Live systems set the numbers.
The two halves of this are deliberately separated. Language and structure come from documents you have already filed, so the draft reads like your submissions rather than like generic regulatory prose. Numbers come from queries against the systems of record, so nothing in the draft is a figure the model produced.
- Prior submission corpus for the equivalent sections in earlier filings, plus the development and validation reports the section needs to draw on, retrieved to establish structure, depth and the phrasing conventions your filings already use.
- LIMS and MES for the current data the section requires: batch analysis results, validation outcomes, in-process controls, and the batches in scope for the section as filed.
- Effective specifications taken from the current approved version rather than from whatever version appeared in the prior filing, with the version identifier carried into the citation.
- Every figure is queried, never generated. The model assembles prose around values returned by tool calls. It has no capability to produce a number of its own, which removes the failure mode that would matter most in a submission.
- A citation table delivered with the draft, mapping each figure in the text to its source system, record identifier and retrieval timestamp, so the numeric QC step becomes a check against a list rather than a hunt.
This is a first draft. A regulatory professional owns the content, the argument and every word that reaches an agency. The agent does not file anything, does not decide what is reportable, and does not touch the submission system. What it removes is the transcription and the hunting, not the judgment.
The judgment stays. The transcription goes.
- To a first draft, much of it transcription and format-matching
- Three prior filings reread to re-establish the established language
- A second person checks every transcribed figure against its source
- Errors concentrate in superseded tables and out-of-date specification versions
- A structured first draft following your own prior sections
- Every figure queried from the system of record and cited to it
- Numeric QC becomes a check against a citation table
- Real regulatory editing still follows, and it is still the majority of the value
These figures come from my demo environment running on synthetic CMC data. They are not client results. The "today" column is what I watched teams actually do across HGS, GSK and Novavax; the "with an agent" column is measured on the demo, against a data model built to be representative rather than against your systems.
It needs your filings to be retrievable, not just archived.
The distinctive requirement here is the document side. Prior submissions have to be reachable as text rather than as a published PDF archive nobody can query, and the sections need enough structure that the right prior section can be found reliably. On the data side the usual question applies: whether the batches, results and specification versions a section cites can be pulled by query rather than assembled by hand.
The Data Diagnostic covers both, and for regulatory work it usually starts with the document corpus rather than the systems. Fixed fee, two to three weeks, and it tells you which sections are within reach first.
The Data Diagnostic, including the price
Your IT and quality systems group will want the architecture, the read and write paths, the Part 11 position and the validation approach. That is all written down on the governance and validation page, in a form you can forward or print.
Send me a workflow.
Tell me the process that eats your team's week. I'll record an agent running it on your data model and send it back. No call required.
kyle@kylelangham.comNo form, no gate, no email capture.
Would rather talk it through? 30 minutes on your workflow, no pitch.