CTD Module 2 authoring, explained.

CTD Module 2 explained: the summary a regulator reads before the study reports

One application, from final study reports to a regulator's first read

1. Studies finish and reports go finalVeltrimab's clinical study report (CSR) VLT-302, 2,140 pages, goes final on August 18. The integrated safety report follows on September 9.
2. Owners are named for each sectionSix people take the 14 separately managed parts, with December 14 as the one date.
3. Writers draft Module 2 and check every number4,812 statements, each cited to a page. This is where 1,935 and 1,947 are caught, or carried forward.
4. Each owner signs, and the sections go into the electronic dossier, the eCTDSix of 14 parts are signed by October 6. The rest are due by November 27.
5. A regulator reads Module 2 firstAnything the writers missed reaches FDA's filing review and its letter at day 74, or EMA's questions at Day 120.

Step 3 is the subject of this guide. The product, people and figures in this example are fictional.

Why Module 2 may add nothing new, and three more surprises

CTD, data cut, analysis set, source of truth

Fourteen terms are defined in the one-page summary.

The seven sections of CTD Module 2, and how many pages ICH expects of each

In shortICH lists seven sections in a fixed order. Two are administrative and five are the writing. Added up, ICH's expectations give about 250 to 690 pages of text, and most of that range comes from the Clinical Summary. ICH writes them as what a section should normally stay within, with tables and figures left out. They are guidance, and no regulation sets a ceiling.

"Module 2 should contain 7 sections in the following order", the guideline says, and it lists them.1 The order looks odd at first, because the two overviews come before the detailed summaries they rest on. An assessor, the regulator's reviewer, reads the overview first, and turns to the summary when a claim needs checking.

An overview argues; a summary reports

Module 2 holds two kinds of writing. ICH says the Clinical Overview "should primarily present the conclusions and implications of those data, and should not recapitulate them".6 The comparisons in the Clinical Summary "should focus on factual observations". In the overview, ICH expects interpretation and "discussion of the place of the test drug in the armamentarium", meaning among the treatments doctors already have.6

When the two blur, they start to compete: an overview that re-summarizes and a summary that interprets, each with its own version of the same table. Keeping them apart also limits how many places a number has to live. ICH encourages cross-references over repetition: "It is not intended that material presented fully elsewhere be repeated in the Clinical Overview".6

Seven sections on paper, a dozen or more files in the eCTD

ICH's granularity annex sets out how finely a dossier may divide into separate files. The Quality Overall Summary may be one document or split into drug substance and drug product parts, with one document for each drug substance. For 2.7.3, "One document for each indication should be submitted", "although closely related indications can be within a single document".1 An indication is the disease a medicine is filed to treat.

The annex's tables for 2.3 and Module 3 also depend on the version of the eCTD standard in use. Whatever you choose stays: "the same approach should be used throughout the life of the dossier", because a later change replaces a file.1

ICH M4, FDA and EMA rules for Module 2, and what changed by November 2026

In shortThe format comes from ICH M4 and its three companion guidelines, unchanged for a decade on the clinical side. What is moving: the quality guideline is being rewritten, with adoption expected in June 2027; eCTD v4.0 is arriving on different clocks at FDA and EMA; FDA's drug center, CDER, revised its refuse-to-file manual in October 2025; and AI rules are still forming.

Which Module 3, 4 and 5 reports feed each Module 2 section, and where the ISS belongs

In shortEach Module 2 section draws mainly on one module, but clinical facts cross from section to section, and quality facts can reach the clinical overview. FDA calls one mix-up a common problem: the integrated summaries of efficacy and safety (ISE and ISS) are long analyses for Module 5, and 2.7.3 and 2.7.4 are short summaries of them.3

The documents a Module 2 writer reads

The ISS and ISE sit in Module 5, and 2.7.4 summarizes them

FDA wrote a whole guidance on this mix-up. NDAs must contain both integrated summaries under 21 CFR 314.50(d)(5)(v) and (vi)(a). There are "no corresponding regulations" for biologics license applications (BLAs), though applicants "are encouraged to provide these analyses".3

The clinical summary sections are "subject to space limitations in the eCTD of about 400 pages, whereas a typical ISS alone often can be substantially larger". FDA names the trap: "applicants incorrectly assume that the clinical summary sections satisfy the regulatory requirement for the ISE and ISS", which "may result in a refusal-to-file action".3

A narrow exception exists for small applications, where 2.7.3 or 2.7.4 can carry the narrative of the integrated summary and the appendices and datasets sit in 5.3.5.3, each location pointing to the other.3

How a patient count, a data cut and a specification limit go wrong in Module 2

In shortThe errors in our fictional dossier are real numbers in the wrong place: a count from another analysis set, a table from an older data cut, a limit from a development document. Each reads well and looks right next to a citation. Three worked cases from our fictional dossier show the arithmetic behind them and what each one changes.

818 or 824: one trial, two right patient counts

A trial randomizes patients, assigning each by chance to the drug or a placebo, and then some never receive a dose. ICH E9 describes the full analysis set as "as close as possible to the ideal implied by the intention-to-treat principle", "derived from the set of all randomised subjects by minimal and justified elimination of subjects".19 In plain words, that is nearly everyone randomized, with few exclusions and a reason for each.

In the fictional study VLT-302 the plan excludes anyone who never received a dose, so the table prints two rows.

The draft of 2.7.3.2 read: "In VLT-302, 818 patients were randomized 2:1 to veltrimab (n=546) or placebo (n=272)." Every number is printed in the report, and the citation points to the right table. The word "randomized" is what is wrong. The corrected sentence keeps both figures, each with its population: "In VLT-302, 824 patients were randomized 2:1 to veltrimab (n=549) or placebo (n=275); 818 received at least one dose and formed the full analysis set."

1,935 or 1,947: what twelve patients change

Exposure is the number that moves most. 2.5.5 cites the June 30 draft: 1,935 patients and 2,207 patient-years. A patient-year is one patient treated for one year. 2.7.4.1 cites the September 9 final: 1,947 patients and 2,316 patient-years.

The gap of 12 patients comes with a gap of 109 patient-years. Twelve new patients can add only a little time in a few weeks, so most of that gain must be patients already counted staying on treatment. That is our reading of the example, and it shows why both figures in a sentence have to change together. Clinical-trial tables often state rates per 100 patient-years, so every such rate has this figure underneath it.

100 ± 10 or 95 to 105: a limit that doubled in width

Section 2.3.P.5 of the fictional dossier states a protein concentration of 100 ± 10 mg/mL, which works out to 90 to 110 mg/mL. The registered release limit in 3.2.P.5.1, specification version 2.1, is 95 to 105 mg/mL. The summary's range is twice as wide, and it claims a limit the application does not support. ICH's instruction reads "Specification(s) from 3.2.P.5.1 should be provided", which names the place to copy from.2

Six kinds of Module 2 error, and which check catches each

In shortIn our judgment, reading a sentence catches few of these errors, because the sentence reads well. Opening the cited page catches more, yet it passes a stale data cut, since the old page does print the old number. A consistency check, comparing the same fact across sections and against the source's version, catches that one.

What published research says about summaries and about checking by eye

No regulator publishes how often Module 2 sections disagree. The nearest published evidence concerns a similar document: a short summary of a longer report written by the same people. A scoping review of 17 studies comparing journal abstracts with their full reports found a median inconsistency of 39%, with individual studies ranging from 4% to 78%, and a median of 19% for major inconsistencies where those were separated.20 Those are journals, with no filing review, so treat this as an analogy and measure your own.

Reading is also a weak way to find errors. In a study of data checking by 195 students, visual checking produced 2,958% more errors than double entry, typing the data twice and comparing, and did no significantly better than entering the data once.21 A quality reviewer re-reading 4,812 statements against their pages in the last weeks before filing is doing the kind of task the study found weak. A study of students on data sheets cannot say how far that applies to a trained reviewer.

Refuse to file, the 74-day letter and review questions: where a Module 2 problem surfaces at FDA

In shortFDA decides within 60 days whether an application can be filed. Four general entries on its deficiency list concern finding data. On a refusal FDA keeps 25% of the fee, $1.15 million in fiscal 2027, and the full fee is due again when the application comes back. A smaller Module 2 problem can surface as a review issue in a 74-day letter or later questions.

The four general deficiencies about finding data

CDER's review manual was revised on October 23, 2025. Under 21 CFR 314.101(d)(3) FDA may refuse to file an application that is incomplete on its face, and CDER reads that to include "when required content is presented in a form that makes it inaccessible".4 Appendix B lists complex deficiencies that may support a refusal, alone or in combination. Its general entries are:

• "Application is unreasonably disorganized"
• "Data tabulations (line listings) and/or graphical displays are not interpretable, are inadequately labeled, or do not indicate the sources of the data"
• "Inadequate annotation in final reports or summaries of where individual studies/clinical investigations or individual data and records can be found"
• "Problems with hypertext links"4

Those entries concern whether a reviewer can find things, and a filing reviewer meets them first in the summaries. Disagreement between two sections has no entry of its own, though it shows a reviewer that the summaries cannot be relied on. FDA's draft guidance says an application may be incomplete when it is "unreviewable".23 A written reason does not rescue a gap: "the mere presence of an explanation is not adequate".4

What a refusal costs in fees

We found no regular public count of refusals. FDA's own longitudinal paper excludes refused applications from its analysis.26 What FDA does publish is how often filed applications fail their first review: a complete response, the letter that says an application cannot be approved as it stands.

FDA published more than 200 complete response letters on July 10, 2025, on applications submitted from 2020 to 2024 and since approved.27 They are public reading for any team that wants to see how FDA words its deficiencies.

The EMA Day 120 questions, clock-stops and Health Canada's view of a poor quality summary

In shortModules 2 to 5 are common across ICH regions, so one Module 2 serves several agencies. The EU clock allows up to 210 active days, stopped while the applicant answers questions, and the first stop usually lasts three to six months. Health Canada says plainly what an inaccurate quality summary costs: more assessor time.

A number that disagrees between sections gives an assessor one more question for that list. Each question has to be answered inside a clock-stop that the applicant has agreed, while the people who wrote the section have often moved on to the next filing.

Health Canada asks for 40 to 100 pages, and spells out why accuracy matters

Health Canada's quality guidance recommends limiting the summary to "the minimum number of pages required to summarize key information (e.g. 40-100 pages)", and offers an optional chemical-entity template. It adds: "The filing of an inaccurate or incomplete QOS will result in greater expenditure of an assessor's time in retrieving, assessing and summarizing data."29

For a synthetic drug filed in both regions, a summary within ICH's 40 pages also sits inside Health Canada's range. This guidance excludes biologics, so a biologic such as our fictional veltrimab needs a separate check.

One Module 2, many labels

Module 2 also feeds the label. A European Medical Writers Association overview notes that the Clinical Summary and Clinical Overview "provide the supporting information for the Summary of Product Characteristics (SmPC) or the product label", "so it is important these documents are consistent".30 A figure that drifts in 2.5 can drift into the prescribing information.

How many applications need a Module 2 each year, and what the market data leave out

In shortPublished counts give the scale: FDA's drug center approved 46 novel drugs in 2025, EMA recommended 104 medicines, and about 65 fee-paying full-application equivalents reach FDA in an average year. Supplements and other regions add many more Module 2 updates. Nobody publishes what writing them costs.

What a day of delay is worth: two Tufts figures from one dataset

A day of delay has two published values, both from the Tufts Center for the Study of Drug Development and one dataset of 645 drugs: about $500,000 in lost prescription sales in its journal article,35 and about $800,000 in its August 2024 white paper.36 We found no source that reconciles them. Both put direct trial cost at about $40,000 a day for Phase II and III. The white paper presents these as updates to an older, much-repeated $4 million to $5 million figure.36

The sales numbers are averages across products, measure sales and leave out costs, and were built for trial delay, so use them as a starting point for your own product's figure.

Can AI write a regulatory submission? What FDA, EMA and the GMP rules say as of 2026

In shortNothing bans AI medical writing for submissions. FDA's draft guidance leaves drafting out of scope where it does not affect patient safety, drug quality or the reliability of results. EMA's paper, in its section on product information, expects close human supervision and a quality review of all model-generated text before submission. EU Annex 22, on AI in good manufacturing practice (GMP), is still a draft, and it concerns manufacturing. The applicant stays answerable for every word.

Reviewers' own tools are changing too: FDA announced an agency-wide generative AI tool on June 2, 2025.44

The published evidence on fabricated detail in clinical text

EMA's word "plausible" has data behind it. Researchers planted one fabricated detail in each of 300 clinical vignettes and tested six large language models. Under the default prompt, the models built on the fake detail about 50% to 82% of the time. A prompt written to reduce the problem, the mitigation prompt, cut the rate without removing it.45 The test used clinical decision support cases, and no Module 2 drafting was involved. The lesson that transfers is that a wrong figure reads as fluently as a right one.

On the time side, one large drugmaker reported cutting a fully human-reviewed clinical study report first draft from an average of 180 hours to 80, with errors down by 50%.46 That is one company's own account of its own tool, on study reports. We found no independent published test of AI drafting on Module 2.

Which Module 2 tasks suit software

Templates, structured content authoring or statement-level drafting: what each reaches in Module 2

In shortRegulatory medical writing automation covers three kinds of product. Templates fix structure. Structured content authoring reuses approved text. Statement-level drafting keeps each fact with its page. They solve different problems, and many teams use more than one. Only the third can tell you that a number has stopped matching its page.

A tool's citation coverage says little about its citation accuracy

The common mistake in judging these tools is counting citations. Coverage asks what share of statements carry a reference. Accuracy asks what share of those references, when opened, say what the statement says. The fictional 818 sentence carried a citation to the right table: full marks on coverage, wrong on accuracy.

For the same reason, a validated system can reliably store a wrong sentence. Part 11 validation looks at "accuracy, reliability, consistent intended performance" of the system and its records.15 Whether a sentence agrees with its source is a separate measurement.

Who writes and signs CTD Module 2 today, and where the working days go

In shortOwnership follows the module a section summarizes, so five functions work inside one document against one date. Writers draft, scientific owners approve, a quality reviewer re-reads numbers near the end, and a submission lead builds the eCTD. In a fictional sponsor's estimate, half the working days of a first draft go to reading, checking and reconciling.

CMC lead to submission lead: who touches Module 2

Six hand-offs between a final study report and the eCTD build

Where 110 working days go on a first Module 2 draft

In-house writers, a contract research organization or AI tools

From a changed exposure figure to a signed 2.7.3: one morning in the Module 2 binder

In shortThis is the example from the top of the page, inside the product. Sam Patel changes 2.7.4.1. Dr. Maya Chen opens it, and the consistency agent checks its facts live. It finds the 12-patient gap, shows both pages and drafts the fix, which goes to Dr. Omar Haddad because he owns 2.5. Then Maya fixes the disagreement in her own section and signs.

1

Module 2 as a binder, with the evidence beside every section

Tuesday, October 6. The home screen is the ICH section tree: 14 sections, 4,812 statements each cited to a page, 312 facts tracked across modules, two conflicts open, six of 14 signed. Each section shows the studies it cites, a ring for how much of it is cited and verified, and who signs.

Maya owns 2.7.3, in owner review at version 0.6 and due October 8. A conflict already sits on it: 2.7.3.2 says 818 patients randomized in VLT-302, and 2.5.4 says 824. On the right, the sign-off steps show 213 of 214 pages opened and matched.

2

Sam updated 2.7.4, and exposure turns amber

At 10:05 a.m. Sam Patel, the clinical scientist, updated 2.7.4.1 Exposure to the drug. Maya opens 2.7.4, and the consistency agent checks its 29 facts against 2.5, 2.7 and Module 5: 28 match, one conflicts.

The card states it in one line: 2.5.5 says 1,935, 2.7.4.1 says 1,947, 12 patients apart. The final Integrated Summary of Safety and the June draft sit side by side, each with its version and date, and the screen names who decides: Dr. Omar Haddad owns 2.5.

3

One click opens the superseded page

Maya opens what 2.5.5 cites: the Integrated Summary of Safety, June data cut, draft version 0.9, page 22 of 296, marked superseded. The passage is highlighted: 1,935 subjects, 2,207 patient-years, data cut June 30, 2026.

This page does say 1,935, so a check that only opens the cited page passes this sentence. The conflict comes from comparing facts across sections and knowing which version is the source of truth.

4

The agent finds why they differ and drafts the fix

The consistency agent opens both versions of Table 1.2, finds that 12 patients enrolled in VLT-303 after the June 30 cut, and proposes a tracked change to 2.5.5, cited to the final report, page 22: "Across the program, 1,947 patients received at least one dose of veltrimab, for 2,316 patient-years of exposure."

Maya does not own 2.5, so she cannot apply it. She sends it to Omar at 10:18 a.m. The conflict stays open on the section until he applies it.

5

2.7.3 cannot be signed while one statement disagrees

Back in her own section, 213 of 214 statements are verified. One is red. The citation verifier opened CSR VLT-302, page 112, Table 14.1.1.1, and found 824 randomized and 818 in the full analysis set.

On Facts across modules she chooses Fix with agent. It compares the populations and proposes wording that keeps both figures, then re-checks 2.5.4 and 2.7.6, which already say 824. Maya owns 2.7.3, so she applies it. Now 214 of 214 are verified, and she signs at 10:26 a.m. as Dr. Maya Chen, Clinical efficacy lead, with the meaning "Approved as section owner" and her password. The section locks as version 1.0.

Ten more screens: the draft, the blocked signature, all 312 facts, the sources and the rules

In shortOne morning is the story. These screens show the rest: the section as a reviewer reads it, the block on signing, the facts board and its fix, the signature dialog, the sent fix, the binder afterward, every source with its version, progress against the design target, the citation rules, and a second application.

Six agents read, draft and check Module 2; the named section owner applies changes and signs

In shortOne agent reads the sources, three draft by discipline, one opens every cited page, and one compares facts across modules. None of them applies a change or signs. Each section has a named owner, a fix to another person's section goes to that person, and a signed section locks as a version.

Source reader to consistency agent: the six Module 2 agents

The section owner signs; the submission lead releases Module 2 to the build

Agent steps and human actions are kept apart on the section's history. A signed section is locked as a version, and a later change opens a new version with a reason. These are controls that support your compliance with 21 CFR Part 11; your own validation and procedures complete it. Botminds is SOC 2 Type II audited, with information security managed to ISO 27001.

The six people who own Veltrimab's Module 2 in the screens

Modules 3 to 5 go in; cited Module 2 sections, a citation map and signed versions come out

In shortThe solution reads what you already have: study reports, integrated summaries, Module 3 and 4 documents, your template and style guide. It returns sections ready for your eCTD build, with every citation mapped and every conflict listed. Its outcome figures here are design targets, measured in the live solution on your own dossier.

Weeks to a first draft, section by section, in the fictional BLA

Your template, document system and eCTD tool stay in place

Your writers stay too. Their work moves toward the argument and the review. The agents cannot approve a draft, so every section still needs an owner.

Cited to a page, compared across modules, signed by its owner

What a late Module 2 and a mismatched number cost a sponsor

In shortStart from your own costs. Nobody publishes the weeks Module 2 takes or how often it is wrong, so the numbers that matter are yours. A few published figures frame them: the value of a day, the fee a refusal keeps, and how often filed applications need a second cycle.

Finding your own Module 2 baseline in a week

Reviewer trust, the label, your writers' time and the next filing: costs no calculator counts

What faster Module 2 drafts return when you count writer hours against the tool's yearly cost

In shortPut in your own numbers. The calculator values the hours released on each Module 2, sets them against what the tool and your testing cost, and keeps any value of filing sooner on a separate line, because that value is the most tempting and the least certain. Every default marked illustrative is an assumption to replace.

Reading, checking, reconciling: where Module 2 saves

Module 2s a year, hours by section and your hourly cost

A sponsor that files one application every two or three years may be better served by a template and a disciplined quality review. A sponsor or contract research organization with several filings and supplements a year passes the 1.3 Module 2s on the chart's top line, provided a pilot confirms the hours.

How ready your regulatory writing team is: filings, source control, owners, findings and sign-off

In shortPick the answer closest to your team today. Eight questions score 0 to 2; the last is for reflection. The meter adds up your score as you go, and each question says why it matters and what a strong answer sounds like.

A low score means naming section owners and final sources first

What to ask vendors of regulatory writing software about citations, superseded data cuts and who signs

In shortAsk every vendor the same questions, us included. The answers show how a statement is built, whether its citation has been opened, what happens when a source changes, who may apply a change, and how you would measure all of it on your own dossier.

We found no published independent test of any Module 2 tool's accuracy, as of November 2026. A planted-fault test on your own closed submission is the fairest comparison, and the same test measures your current quality review.

Test Module 2 software on a closed submission with planted faults, then decide on four numbers

In shortMeasure first, then test on a closed submission whose answers are settled, with faults you planted and a superseded report left in. Then run alongside one live section. None of these steps commits you to a vendor, and the test measures your current process as well.

1. Trace ten facts through a filed Module 2: patients randomized, patients exposed, a key result, a specification limit. That is your baseline for escaped conflicts.
2. Choose a closed submission and two sections: one numeric, such as 2.7.3, and one argument, such as 2.5. Keep the data in the environment your agreements require.
3. Plant 40 faults from your own quality history: wrong population, superseded source, rounding, development limit, overstated wording. Lock the key, and keep planted text away from any live dossier.
4. Run and grade: two graders who did not choose the tool, against the page, with a third to settle disputes. Sample 60 unflagged statements too, to see what was missed.
5. Decide with four numbers: citation accuracy, catch rate by fault type, false flags on clean text, and owner minutes to a section they are willing to sign.

Citation accuracy, catch rate and false flags per planted fault

CTD Module 2 questions answered: sections, page limits, QOS, ISS, refuse to file and AI

In shortThe questions writers and buyers search for most, each answered on its own.