In This Article
- Executive Summary
- What AI Is Doing in CSR Drafting Today
- What Regulators Have Said About AI-Drafted Submission Text
- A Three-Zone Map of the CSR Under ICH E3
- Source-to-Text Traceability for Every Number
- QC of Numbers Against the TLFs
- Reviewer Accountability and Version Control
- Data Confidentiality With AI Vendors
- What to Document for Inspection
- Conclusion
- For Further Reading
- References & Sources
Executive Summary
Sponsors and their writing vendors are now using generative AI to draft clinical study reports (CSRs) from the protocol, the statistical analysis plan (SAP), and the tables, listings, and figures (TLFs). The published experience is consistent: AI drafts the protocol-derived sections well, struggles with results sections that depend on tables, and is least reliable when it is asked to produce numbers itself. That pattern should shape the controls.
This article offers a three-zone map of the CSR. Zone 1 is text carried over from the protocol and SAP. Zone 2 is the number-bearing results text. Zone 3 is patient narratives and appendices. Each zone gets a different AI role and a different control. The central rule for Zone 2 is simple: numbers are inserted from validated TLF outputs, never generated, and every number in the text carries a link back to the output it came from.
The article then covers the controls a quality unit will ask about: source-to-text traceability, QC of numbers against the TLFs, reviewer accountability, version control across a moving set of sources, data confidentiality with AI vendors, and the file you should be able to hand an inspector. It draws on published industry pilots, peer-reviewed and preprint research, and what FDA, EMA, and ICH have said so far. One point up front: FDA’s January 2025 draft AI guidance says it does not address AI used for drafting a submission when that use does not affect patient safety, drug quality, or the reliability of study results. That is a conditional exclusion, and a wrong number in a CSR is exactly the kind of error that can make the condition fail.
What AI Is Doing in CSR Drafting Today
A clinical study report is the integrated account of one clinical study. It combines the clinical and statistical description of the study, the results, and appendices that hold the protocol, sample case report forms, listings, and supporting documents. It is also one of the longest documents a medical writing team produces, and much of its content is restated from documents that already exist. That combination is why CSRs were among the first regulated documents that sponsors tried to draft with large language models.
The published record on this is still thin, and much of what circulates comes from vendors. The sources below are the ones that describe a method, not only a result.
An In-House Pilot at a Japanese Sponsor
Shionogi described its program in a 2025 PharmaSUG paper. The company set a goal of a 15 to 25 percent improvement in drafting efficiency, using an in-house generative AI system on a secure cloud platform that draws on organizational templates and past trial documents.1 Three design choices in that paper are worth noting because they are controls, even though the paper does not call them that:
- Section-by-section generation. The system drafts one section at a time rather than the whole report in one pass.
- Template-to-source mapping. The CSR template is mapped ahead of time to the protocols and figures that each section should draw from, and that mapping is stored in a database.
- Logging. Each generation request and its output are recorded, so issues can be investigated later.
The pilot results are the most useful part. In a small test with medical writers, sections that mostly restate the protocol (Sections 1 to 8) could be drafted in minutes to a quality suitable for further editing. Chapters that depend on figures and tables showed inaccuracies in data extraction at first. Section 11, the efficacy evaluation, varied so much from study to study that writers still tuned prompts or wrote the text themselves. Even Section 9, the study design, needed fixes for tense and wordiness.1 The authors also state that AI-generated text is a preliminary draft only and is never used verbatim in a submission.
A Consultancy Architecture for the CSR
Deloitte has published a CSR case study that breaks the report into components and assigns each to a different tool: generative AI for summaries and narrative text, conventional engineering for building analysis datasets, and integration with content management and structured content authoring platforms.2 Its design includes a review agent that highlights low-confidence passages for the writer, and it lists data privacy, regulatory compliance, and human oversight as considerations. Deloitte estimates a 20 to 30 percent reduction in medical writer effort across documents, and for a top-10 biopharma averaging 50 annual submissions, an estimated 25 percent efficiency gain from using generative AI for CSRs. Those are estimates from a consultancy, not measured outcomes, and should be read that way.
What the Research Says About Tables
Two studies from industry statisticians address the part of the CSR where errors matter most. A Novartis-led team used GPT-4o with public CDISC pilot data in ADaM format, generating 16 tables and one Kaplan-Meier plot.3 Their method had the model write Python code, run it against the datasets, and show both the code and the result. That approach matched the published reference results with 100 percent accuracy, except for analyses involving statistical tests, which needed more prompt work and statistician input. The more important finding is what did not work: when the team had the model fill table shells directly, without the code step, they found the accuracy of the table results too low to pursue.
A 2026 PharmaSUG paper from Merck looked at a related task, generating TLF shells and the R code behind them. Over 1,999 bootstrap experiments across three LLM providers, a hybrid retrieval approach with reranking scored better than direct prompting (a mean quality score of 85.7 against 81.7). Iterative, LLM-guided debugging got the generated code to run in 70 percent of cases within three to five rounds.4 The same paper notes that commercial CSR automation products often lack publicly available performance metrics or independent academic validation.
The Pattern Across the Evidence
Read together, these sources point the same way. Language models do well at restating and reorganizing text that already exists in a controlled source. They are weaker when they must read structured tables, and they are not reliable when asked to produce a number by reading or reasoning rather than by running validated code. The controls in the rest of this article follow from that pattern. They are not general AI governance. They are specific to how a CSR is built.
One scoping note. Testing whether a drafting tool is fit for use before you deploy it (an acceptance test of the tool) is a separate activity from controlling each CSR the tool helps produce. This article is about the second: the controls that apply to every report, every time.
What Regulators Have Said About AI-Drafted Submission Text
No regulator has issued guidance specific to AI-assisted CSR authoring. What exists is a set of broader statements, each of which touches the CSR at a different angle. It helps to read them precisely, because the commentary around them is often looser than the text.
FDA: A Conditional Scope Exclusion
FDA’s January 2025 draft guidance, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, proposes a risk-based credibility framework for AI models that produce information used to support regulatory decisions. It is still a draft. Its scope section states that it does not address AI used “for operational efficiencies (e.g., internal workflows, resource allocation, drafting/writing a regulatory submission) that do not impact patient safety, drug quality, or the reliability of results from a nonclinical or clinical study.”5
Some teams read that sentence as “FDA has said AI drafting is out of scope.” That reading drops the second half. The exclusion applies only when the drafting does not affect the reliability of results. A CSR that misstates a hazard ratio, a responder rate, or the number of serious adverse events affects how the results are understood by the reviewer. So the exclusion is best read as an expectation: keep the AI’s role in drafting such that it cannot change the results, and be able to show that. If an AI tool starts to analyze data or produce results, it is no longer doing drafting, and the credibility framework becomes relevant. The same paragraph encourages sponsors to talk with FDA early if they are unsure whether a use is in scope.
EMA: Close Human Supervision of Generated Text
EMA’s Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle, adopted in September 2024, has one passage on drafting. It addresses AI used for drafting, compiling, editing, translating, tailoring, or reviewing product information documents, and says such use should happen under close human supervision, with quality review in place to make sure all model-generated text is factually and syntactically correct before submission.6 That passage is about product information (labeling), not CSRs, and it should not be cited as if it named CSRs. But the reasoning it gives, that generative models are prone to plausible but wrong or incomplete output, applies to any submission text. The same paper places responsibility on the sponsor or applicant to make sure that all models, datasets, and data processing pipelines used are fit for purpose.
FDA and EMA Together: Traceable Data and Whole-System Testing
In January 2026, FDA and EMA published ten Guiding Principles of Good AI Practice in Drug Development. They are principles, not guidance, and they are framed around AI used to generate or analyze evidence. Two of them translate directly to CSR work. Principle 6 asks that “Data source provenance, processing steps, and analytical decisions are documented in a detailed, traceable, and verifiable manner, in line with GxP requirements.” Principle 8 asks that performance assessment cover the complete system, including human-AI interactions.7 For a CSR, the complete system is the tool plus the writer, statistician, and reviewers who use it. Testing the model alone does not tell you whether the process produces a correct report.
ICH E6(R3): Nothing New, Which Is the Point
The most useful regulatory text for AI-assisted CSRs does not mention AI at all. ICH E6(R3), the revised good clinical practice guideline, asks the sponsor to make sure clinical trial reports in marketing applications meet the standards of ICH E3. It asks for appropriate and documented quality control of statistical programming and data analysis, naming outputs for the clinical trial report as an example, and for traceability of data transformations and derivations. And it states that when the sponsor transfers trial-related activities to a service provider, ultimate responsibility for those activities, including the reliability of trial data, stays with the sponsor.8
Taken together, these documents do not create a new rulebook for AI in medical writing. They confirm that the existing obligations apply in full. The sponsor owns the accuracy of the report, regardless of who or what drafted it.
What none of these documents say. None of the regulator documents reviewed here require a sponsor to disclose in the CSR that AI helped draft it, and none prohibit it. That leaves a decision for each company to make and write down. Whatever you decide, decide it in a procedure, not report by report.
A Three-Zone Map of the CSR Under ICH E3
ICH E3 sets out the structure and content of a CSR. It was finalized in November 1995 and describes a single core report acceptable to the regulatory authorities of the ICH regions.9 A 2012 question-and-answer document confirms that “ICH E3 is a Guideline, not a set of rigid requirements or a template, and flexibility is inherent in its use.”10 Sections can be reordered, renamed, or added when the study calls for it, but if a report does not address an aspect of E3 that is relevant to the study, the report should say so and give the reason.
That flexibility matters for AI drafting in two ways. First, a tool cannot be trusted to decide on its own which E3 topics apply and which can be dropped. The decision to omit or restructure a section is an authoring decision that needs a named owner. Second, the more a company standardizes its CSR template, the easier it is to map each section to its sources. The CORE Reference, developed by regulatory writers and statisticians and published in 2016, gives section-level content guidance built on E3 and the 2012 Q&A and was designed to help authors write CSRs that need minimum redaction before public disclosure.11 TransCelerate’s common CSR template, released in November 2018, draws on both E3 and CORE. The CORE authors’ 2019 review of that template is a reminder that “A single CSR Template cannot address all study designs without some degree of inherent structural flexibility.”12
Three Zones, Three Different Risks
The practical step is to classify every section of your CSR template into one of three zones, based on where its content comes from and what kind of error it can carry. The Shionogi pilot results line up closely with this split.1
Carried-Over Text
Sections restated from the protocol, protocol amendments, and SAP: ethics, investigators and administrative structure, introduction, objectives, and most of the investigational plan (E3 Sections 5 to 9). Main risk: the text reflects the wrong protocol version or misstates a planned method.
Number-Bearing Results Text
The synopsis, study patients, efficacy evaluation, safety evaluation, and discussion and conclusions (E3 Sections 2 and 10 to 13). Main risk: a wrong, rounded, or misattributed number, or an interpretation the data do not support.
Narratives and Appendices
Narratives of deaths, other serious adverse events, and other significant adverse events (E3 12.3.2), plus appendix documents in Sections 14 and 16. Main risk: a patient-level fact drawn from the wrong source, or personal data that should not be there.
Human-Only Decisions
Which E3 topics apply, how sections are restructured, the benefit-risk framing in the discussion, and the final conclusions. These stay with named people and are recorded as decisions.
The table below turns the zones into rules. It is a starting point to adapt, not a standard.
| Zone | Typical E3 Sections | Allowed AI Role | Primary Control |
|---|---|---|---|
| 1. Carried-over text | 5 Ethics; 6 Investigators; 7 Introduction; 8 Objectives; 9 Investigational plan | Draft by restating and converting tense from the approved protocol, amendments, and SAP | Comparison against the named source version; check that every protocol amendment is reflected in 9.8 |
| 2. Results text | 2 Synopsis; 10 Study patients; 11 Efficacy; 12 Safety; 13 Discussion | Draft sentence structure around numbers that are inserted from validated outputs; never compute or transcribe numbers | Every number tagged to a TLF output ID; 100 percent numeric QC; statistician review of interpretation |
| 3. Narratives and appendices | 12.3.2 Narratives; 14 Tables and figures; 16 Appendices | Draft narrative prose around data fields filled by code; assemble appendix lists | Field-level source check for any manually added fact; medical review of every narrative; personal data check |
| Human-only | Structure decisions; 13 benefit-risk framing; conclusions | Suggest, never decide | Named decision owner; decision recorded in the review log |
Why the Synopsis Belongs in Zone 2
The synopsis is short, and teams often treat it as a summary to be written last. In practice it is the most-read part of the report and the place where a number most often disagrees with the body, because it is edited separately and late. Put it in Zone 2 and apply the same rule: its numbers are inserted from the same TLF outputs as the body, not copied from the body text, and a consistency check compares the two.
Locking the Template
An AI tool that produces its own headings will drift from your template. Structure is a control in its own right. The drafting tool should write into a locked template whose section headings, numbering, and required sections are fixed by the medical writing group, with any restructuring done by a writer and recorded. That also keeps the report aligned with the CTD placement your regulatory operations team expects.
Source-to-Text Traceability for Every Number
The most important control in an AI-assisted CSR is one that good medical writing teams already practice informally: every number in the text can be traced to the output it came from. With AI drafting, that practice has to become a formal, checkable record. The reason is the finding from the Novartis study: a model that reads a table and writes the number into a sentence is doing exactly the task the researchers found too inaccurate to pursue.3
The Number Chain
Every number in a CSR has a chain behind it. Data are collected in the case report form, mapped to SDTM, derived into ADaM analysis datasets, and summarized in TLF outputs by validated programs. ICH E6(R3) already asks the sponsor to ensure traceability of data transformations and derivations and documented QC of programming that produces outputs for the report.8 The last link in the chain, from TLF output to CSR sentence, is the one most often left to a person’s care rather than a record. It is also the link AI drafting changes.
The rule for Zone 2: numbers are inserted, not generated. The AI may draft the sentence “A total of [N] participants were randomized, of whom [n] ([%]) completed the study.” The values in brackets are filled from the validated output by a deterministic step, and each one carries the output identifier it came from. The model never reads a value out of a table image or a PDF and types it into a sentence.
What a Source Map Looks Like
The practical form of this rule is a source map: a structured record, kept alongside the draft, that lists every number in the CSR text with its location and its source. The source map can be generated by the same tooling that inserts the values. At minimum, each entry should hold:
- The CSR location (section and paragraph, or a stable placeholder ID)
- The value as it appears in the text, including rounding and units
- The TLF output ID, the table row and column or statistic, and the analysis population
- The version or run date of that output
- Whether the value was inserted automatically or entered by hand, and if by hand, by whom
Values entered by hand are allowed. Sometimes a writer needs a number that no table gives directly, such as a difference between two arms that the SAP did not specify. Those values should be visible as manual entries, sent to the statistician for confirmation, and preferably replaced by an output from a validated program. A source map that shows every manual entry is far easier to defend than a draft that looks automatic but is not.
Building the Source Map Into the Workflow
Tag the Template
Before drafting begins, mark every Zone 2 placeholder in the CSR template with the TLF output that should feed it, using the TLF shell numbering from the SAP. This is where Shionogi’s template-to-source mapping becomes a control rather than a convenience.
Insert Values From Machine-Readable Outputs
Have the statistical programming team deliver outputs in a machine-readable form (datasets or structured files) alongside the formatted tables, so values are read by code, not by a model reading a page.
Let the Model Draft Around the Values
The model writes the surrounding sentence structure, comparisons, and transitions, with the values already fixed. Instructions should forbid the model from adding, changing, or removing a number.
Detect Any Untagged Number
After each drafting pass, a script scans the text for any numeral or spelled-out number that is not in the source map. Each one is either tagged to a source, confirmed as a non-data number (a section reference, a visit day defined in the protocol), or removed.
Freeze the Map With the Draft
Each draft version is saved with its source map. When a draft goes to review, the reviewers see the same map the writer used.
Words That Carry Numbers
Traceability is not only about numerals. A sentence that says one arm had “a higher rate” of an adverse event, or that a result was “consistent across subgroups,” is a quantitative claim. Language models are prone to this kind of smoothing, because it reads well. Treat comparative and directional words in Zone 2 as claims that need a source: the specific output that shows the difference, and the SAP-defined analysis that supports calling it a difference at all. Where the SAP does not support a comparison, the text should not make one.
The same applies to population labels. “Patients” in the efficacy section must mean the analysis set named in the SAP, and the source map should record which population each value comes from. A correct number from the wrong analysis set is still an error.
QC of Numbers Against the TLFs
Medical writing teams have always run a quality control check of CSR numbers against the TLFs, usually by a second writer or a dedicated QC reviewer comparing each value by hand. AI drafting does not remove that check. It changes what the check is for and where it should be strongest.
Insertion by Code Changes the Check
Writers who automate patient narratives have already worked through this. Writing in EMWA’s Medical Writing journal, two CRO medical writers explained that when data are inserted into narratives programmatically, the medical writer does not need to QC each data item, while any text added by hand from secondary sources, such as safety database forms, still needs a QC check against the output generated by code.13 The same authors note that a medical review of each narrative is still necessary, and that AI text tools work from patterns rather than facts and often produce factual errors.
That is the right model for the whole CSR. Where numbers are inserted by a validated, deterministic step, the QC effort moves from re-checking every value to verifying that the insertion step worked: that each tag points to the right output, row, and population, and that the output version is current. Where numbers are entered by hand, or where the model has touched a sentence that contains a number, the full manual check stays.
Keep Arithmetic in Code
A 2026 preprint by statisticians from Daiichi Sankyo and Bristol Myers Squibb makes the same split for TLF review itself. The PROVE framework uses a language model only to interpret varied table wording and match labels, and leaves every numerical and logical decision to coded validators.14 In synthetic oncology reporting packages with 15 injected discrepancies each, the coded checks classified packages correctly when table labels matched the validator vocabulary exactly. When labels varied in wording, adding the model for label matching raised recall from 0.588 to 0.993. The authors’ conclusion is that language models help most with interpreting variation in TLF wording, while executable checks should stay responsible for final numerical validation. The data are synthetic and the paper is a preprint, so the specific figures should not be relied on, but the division of labor is sound.
The same paper points out that arithmetic, censoring, denominator, and subgroup discrepancies have still reached final regulatory-facing documents, citing errata in FDA oncology advisory committee materials.14 The existing human QC process is not perfect either. That is an argument for adding coded checks, not for trusting the model more.
Checks Worth Automating
| Check | What It Catches | How to Run It |
|---|---|---|
| Value match | Text value differs from the tagged output | Script compares each source map entry with the current output file |
| Untagged number | A number the model or a writer added without a source | Script lists every number in Zone 2 text not present in the source map |
| Percentage arithmetic | n and % that do not agree with the stated denominator | Recompute from n and N; flag differences beyond the SAP rounding rule |
| Population consistency | Value drawn from a different analysis set than the paragraph states | Compare the population in the source map with the population named in the text |
| Synopsis to body | Synopsis and body disagree on the same result | Match values with the same output ID across Sections 2 and 10 to 12 |
| Output version | Text built from a superseded TLF run | Compare output run dates in the source map with the final delivery manifest |
What Stays Manual
Some checks cannot be scripted, and they are the ones that need experienced people. Does the efficacy text describe the primary endpoint the way the SAP defines it, including the estimand? Does the safety text give appropriate weight to a signal that appears in a listing but not in a summary table? Does the discussion claim more than the results show? Was a sensitivity analysis that disagrees with the primary analysis reported plainly? These are interpretation checks for the statistician and the clinical lead, and AI drafting tends to make them harder, because the prose reads smoothly whether or not it is right.
Reduce QC only on evidence. If insertion by code is working, you may reduce the manual value-by-value check for inserted numbers. Base that change on evidence from your own process, such as a period of full parallel checking with zero insertion errors, and write the change into your procedure. Never reduce the manual check for any sentence where the model edited text that contains a number.
Reviewer Accountability and Version Control
An AI tool does not sign a CSR. ICH E3 lists, among the appendices, the signatures of the principal or coordinating investigator or the sponsor’s responsible medical officer, depending on the regulatory authority’s requirement.9 Inside the company, each section also has a person who answers for it. AI drafting tends to blur that, because a draft that appears fully formed can feel as though it has no author. The fix is to name the author anyway.
Who Owns What
| Role | Accountable For | Record They Leave |
|---|---|---|
| Lead medical writer | The draft as a whole, including every AI-drafted section; template structure decisions; resolution of review comments | Draft history, decision log, comment resolution record |
| Study statistician | Correct use of outputs and populations; interpretation of statistical results; confirmation of any manually entered value | Statistical review sign-off; confirmations on the source map |
| Clinical or medical lead | Clinical interpretation, safety emphasis, narratives, discussion and conclusions | Medical review sign-off, narrative review record |
| QC reviewer | Numeric QC and consistency checks per procedure | QC checklist, findings log, script outputs |
| System owner | The drafting tool’s configuration, version, access, and logs | Configuration record, change records, access reviews |
Design the Review So It Can Catch Errors
A reviewer who receives a polished draft is more likely to approve it than one who receives rough text. This tendency to over-rely on automated output, usually called automation bias, appears in the list of risks that EMA’s guiding principles on large language models draw from published research on language model harms.15 Three practices help. First, show reviewers which passages were AI-drafted and which numbers were entered by hand, so attention goes where the risk is. Second, give reviewers the source documents and the source map side by side with the draft, not only the draft. Third, ask reviewers to record what they checked, not only that they approved. “Reviewed Section 12.2 against T14.3.1.1 to T14.3.1.4 and L16.2.7” is evidence. A checkbox is not.
Version Control Across a Moving Source Set
A CSR is drafted against sources that keep changing. The protocol may have several amendments. The SAP may be finalized just before database lock. TLFs are often produced first from dry-run or blinded data, then rerun after lock, and sometimes rerun again after a data correction. A human writer working over months tracks those changes by habit. An AI drafting step does not, unless the process makes it.
The control is to treat the set of source documents as a versioned input. Every AI drafting run should record which version of each source it used: protocol and amendment number, SAP version, and the TLF delivery identifier or run date. When a source changes, the affected sections are identified from the source map, regenerated or edited, and rechecked. A draft that mixes Zone 2 text from a dry run with text from the final run is one of the easiest errors to make and one of the hardest to see.
The tool itself is also a versioned input. Record the model and version, the prompt or instruction set, and any retrieval configuration used for each run, along with the output. The Shionogi team logged every generation request and output, and stored each update so earlier versions were not lost.1 That is the minimum. If the vendor updates the underlying model in the middle of a report, you want to know which sections were drafted before and after.
Where Part 11 Applies
The final CSR is an electronic record submitted to FDA, so its creation and approval in your document management system fall under 21 CFR Part 11. Section 11.10 calls for, among other things, validated systems, secure computer-generated time-stamped audit trails that record the creation, modification, and deletion of records without obscuring earlier entries, and revision and change control procedures for systems documentation.16 Most companies already meet this through their validated document management system. The open question is the AI drafting environment: if drafts are created and edited there before they enter the document system, decide whether those drafts are records you need to keep, and if they are, keep them with an audit trail. A simple rule that works for many teams is that the drafting environment produces proposals, and nothing becomes part of the CSR until a named writer accepts it into the controlled document, where the audit trail starts.
Data Confidentiality With AI Vendors
The inputs to an AI-drafted CSR are among the most sensitive documents a sponsor holds. They include unblinded efficacy and safety results before they are public, patient-level listings, narratives built from safety database records, and the protocol and SAP, which may contain commercially confidential information. For a public company, top-line results may also be material information that has not yet been disclosed. All of this is sent to a model when a CSR is drafted, often through a vendor and sometimes through a CRO using a tool the sponsor did not select.
What Regulators Say About Inputs
EMA’s guiding principles on large language models were written for staff at EU regulatory agencies, not for industry, but they show how regulators think about the same tools. They warn that prompts may be stored and further processed, which creates risks for confidential information, data protection, and privacy. They advise users to understand whether a model is deployed under the organization’s control or is an open online model that may reuse prompt data, and to avoid putting into a prompt personal data, trade secrets, protected intellectual property, or data whose sharing is restricted by contract.15 EMA’s reflection paper adds that the applicant or marketing authorization holder is responsible for making sure all personal data, including data held indirectly within AI models, are processed in line with EU data protection law.6
Contract Terms That Matter for CSR Drafting
Ordinary vendor qualification covers much of this. The terms below are the ones specific to AI drafting that are most often missing from a standard agreement.
| Term | What to Require |
|---|---|
| Use of your data | No use of prompts, source documents, or outputs to train or improve models, for the vendor or any model provider, without written consent |
| Retention and deletion | A stated retention period for prompts, retrieved content, and outputs, deletion at the end of the engagement, and a way to confirm it |
| Sub-processors | A named list of model providers and hosting providers, with notice before changes |
| Location | Where data are processed and stored, to support your data protection obligations |
| Model change notice | Advance notice of changes to the underlying model or its version, so you can assess impact on reports in progress |
| Logs and access | Access to generation logs and audit trails for your studies during the engagement and for inspection afterward |
| CRO use of AI | Disclosure and sponsor approval before a CRO or writing vendor uses an AI tool on your documents, with the same terms passed down |
The last row is the one most likely to be overlooked. ICH E6(R3) says that when a sponsor transfers activities to a service provider, ultimate responsibility, including for the reliability of trial data, stays with the sponsor, and agreements should define roles and responsibilities clearly.8 If a writing vendor uses an AI tool to draft your CSR, that tool is part of your process whether or not you chose it. Ask the question in the statement of work.
Minimize What Goes In
The simplest confidentiality control is to send the model less. Zone 1 drafting needs the protocol and SAP, not patient data. Zone 2 drafting, under the insertion rule above, needs sentence templates and the names of outputs, not the full listings. Zone 3 narratives are the part most likely to need patient-level data, and that work is often better done by narrative templates filled by code, with the model limited to smoothing prose around fields that are already filled.
Write With Publication in Mind
Confidentiality does not end at submission. Under its clinical data publication policy, EMA publishes clinical study reports submitted in centralized marketing authorization applications, after anonymization; personal data and justified commercially confidential information are redacted.17 The EU Clinical Trials Regulation separately provides for publishing results of trials conducted in the EU. The CORE Reference was built partly to help writers produce reports that need minimum redaction before disclosure.11 An AI drafting tool that adds unnecessary patient detail to a narrative, or pulls commercially confidential text from an older document into a new report, creates redaction work later and a disclosure risk if it is missed. Add a check for personal data and unneeded confidential content to Zone 3 review.
What to Document for Inspection
An inspector or auditor reviewing an AI-assisted CSR will ask the same questions they ask about any report: who wrote it, from what sources, how it was checked, and who approved it. The difference is that “who wrote it” now has a two-part answer, and the evidence for the checking has to show that the AI step could not have introduced an undetected error into the results. Plan the file so that a person who was not on the study team can follow it.
The inspection file for an AI-assisted CSR
- Intended use statement for the drafting tool, naming the CSR zones it may draft and the tasks it may not do (computing or transcribing numbers, deciding structure, writing conclusions)
- Tool qualification evidence appropriate to that intended use, including the acceptance testing done before first use
- Procedure for AI-assisted CSR drafting, including the zone map, the insertion rule, and the QC approach
- Source version record for the report: protocol and amendments, SAP version, TLF delivery identifiers
- Tool version record: model, version, instruction set, and retrieval configuration for each drafting run
- Source map for the final report, with manual entries identified and confirmed
- QC records: script outputs, findings, and resolutions for every check in the procedure
- Review records: what each reviewer checked, their comments, and how comments were resolved
- Decision log: structure changes, omitted E3 topics with rationale, and any deviations from the procedure
- Vendor records: agreements covering data use, retention, sub-processors, and model change notice, plus the vendor’s logs for the study
- Training records for writers and reviewers on the tool and the procedure
Be Able to Explain the Process in Two Minutes
Documents are necessary but not sufficient. The lead writer and the system owner should be able to explain, without notes, what the tool does, what it is not allowed to do, how a number gets into the report, and how the team would know if the tool got something wrong. If that explanation depends on “the reviewer would catch it,” the process needs work. If it rests on “numbers are inserted from validated outputs, every number is mapped, and a script flags anything untagged,” the process is sound.
Treat Errors as Process Data
When QC or review finds an error in an AI-drafted section, record which zone it was in, whether it came from the model, a manual entry, or a source change, and how it was caught. Over a few reports, that record tells you whether the zone rules are working and where to tighten them. It also gives you the evidence base for any later change to QC intensity. Errors found late, such as in a regulatory query after submission, should go through your normal deviation and CAPA (corrective and preventive action) process, with the AI step assessed as a possible cause like any other.
Connect the CSR to the Rest of the Submission
CSR text is reused. It feeds the clinical summaries and overview in Module 2, public results postings, and publications. An error that passes into the CSR spreads into those documents too, and a later correction has to be made in each of them. A source map that follows the CSR into downstream documents, at least for key results, makes those corrections traceable. It is also one of the more practical reasons to invest in structured content and component reuse, because the same output-to-text links can then be maintained once.
Conclusion
The published experience with AI in CSR drafting is consistent enough to act on. Models restate protocol and SAP content well and save real time on those sections. They are weakest where a CSR is most sensitive: reading tables and putting numbers into sentences. The controls that work follow that split. Classify the report into zones. Let the model draft carried-over text against named source versions. In the results text, insert every number from validated outputs and keep a source map that ties each value to its origin. Keep arithmetic and comparison checks in code, keep interpretation with the statistician and clinical lead, and make reviewers record what they checked. Treat both the sources and the tool as versioned inputs, keep patient and confidential data out of the model wherever the task allows, and assemble the inspection file as you go rather than after the fact.
None of this requires new regulation, and none of it is waiting on one. FDA, EMA, and ICH have already said that the sponsor owns the accuracy of the report and the traceability of its data, whoever drafted it. Sakara Digital works with pharma and biotech organizations putting AI into regulated writing processes like this one, from the zone map and source map design to vendor terms and inspection readiness. If you are planning AI-assisted CSR drafting, or reviewing a vendor’s approach, and want an independent view on where the controls need to be, we are happy to have that conversation.
For Further Reading
For Further Reading
- Validating GenAI for Regulatory Document Authoring: A Practical Playbook
- Human Review Sampling Plans for AI Outputs in GxP Work
- Tracing Analysis Datasets Back to Source Under ICH E6(R3)
- Structured Content Authoring for Regulatory Submissions: A Readiness Checklist
- Supplier Qualification for AI Vendors in GxP: Questions That Separate Evidence From Marketing
References & Sources
- Matsuno T, Yoshida Y, Kitanishi Y (Shionogi & Co., Ltd.). “Integrating Generative AI into Medical Writing: Building an Interactive Drafting and Search Framework.” PharmaSUG 2025, Paper AI-209. https://pharmasug.org/proceedings/2025/AI/PharmaSUG-2025-AI-209.pdf
- Pushchinsky G, Parthasarathy S, Larson J (Deloitte Consulting LLP). “Transforming Medical Writing with Generative Artificial Intelligence (GenAI): Placing No Regret Bets on Medical Writing.” Deloitte. https://www2.deloitte.com/content/dam/Deloitte/us/Documents/ls-genaI-in-medical-writing.pdf
- Yang Y, Krusche P, Pantoja K, Shi C, Ludmir E, Roberts K, Zhu G. “Using Large Language Models to Generate Clinical Trial Tables and Figures.” arXiv:2409.12046, 2024. https://arxiv.org/abs/2409.12046
- Yan J (Merck & Co., Inc.). “Schema-Preserving Generation of Clinical TLF Templates and Executable R Code via Iterative LLM-Guided Debugging.” PharmaSUG 2026, Paper AP-211. https://pharmasug.org/proceedings/2026/AP/PharmaSUG-2026-AP-211.pdf
- US Food and Drug Administration. “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products.” Draft guidance, January 2025. https://www.fda.gov/media/184830/download
- European Medicines Agency. “Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle.” EMA/CHMP/CVMP/83833/2023, adopted September 2024. https://www.ema.europa.eu/en/documents/scientific-guideline/reflection-paper-use-artificial-intelligence-ai-medicinal-product-lifecycle_en.pdf
- US Food and Drug Administration and European Medicines Agency. “Guiding Principles of Good AI Practice in Drug Development.” January 2026. https://www.fda.gov/media/189581/download
- International Council for Harmonisation. “ICH E6(R3) Guideline for Good Clinical Practice.” Step 4, January 6, 2025. Sections 3.6.6, 3.16.2, 3.17.2, and Principle 10. https://database.ich.org/sites/default/files/ICH_E6%28R3%29_Step4_FinalGuideline_2025_0106.pdf
- International Conference on Harmonisation. “Structure and Content of Clinical Study Reports E3.” Step 4, November 30, 1995. https://database.ich.org/sites/default/files/E3_Guideline.pdf
- ICH E3 Implementation Working Group. “ICH E3 Guideline: Structure and Content of Clinical Study Reports, Questions & Answers (R1).” July 6, 2012. https://database.ich.org/sites/default/files/E3_Q%26As_R1_Q%26As.pdf
- Hamilton S, Bernstein AB, Blakey G, et al. “Developing the Clarity and Openness in Reporting: E3-based (CORE) Reference user manual for creation of clinical study reports in the era of clinical trial transparency.” Research Integrity and Peer Review, 2016;1:4. https://researchintegrityjournal.biomedcentral.com/articles/10.1186/s41073-016-0009-4
- Hamilton S, Bernstein AB, Blakey G, et al., on behalf of the Budapest Working Group. “Critical review of the TransCelerate Template for clinical study reports (CSRs) and publication of Version 2 of the CORE Reference (Clarity and Openness in Reporting: E3-based) Terminology Table.” Research Integrity and Peer Review, 2019;4:16. https://pmc.ncbi.nlm.nih.gov/articles/PMC6683477/
- Cobb L, Haycock N. “Narratives for a clinical study report: The evolution of automation and artificial intelligence.” Medical Writing, 2023;32(3):28–31. https://doi.org/10.56012/kcca1211
- Lu Z, Zheng C, Han Y. “Bridging Probabilistic LLMs and Deterministic Statistical Validation: The PROVE Multi-Agent Framework for Clinical Trial Reporting.” arXiv:2607.28218 (preprint), July 2026. https://arxiv.org/abs/2607.28218
- European Medicines Agency. “Guiding principles on the use of large language models in regulatory science and for medicines regulatory activities.” August 29, 2024. https://www.ema.europa.eu/en/documents/other/guiding-principles-use-large-language-models-regulatory-science-medicines-regulatory-activities_en.pdf
- Code of Federal Regulations. 21 CFR 11.10, “Controls for closed systems.” https://www.ecfr.gov/current/title-21/chapter-I/subchapter-A/part-11/subpart-B/section-11.10
- European Medicines Agency. “Clinical data publication.” https://www.ema.europa.eu/en/human-regulatory-overview/marketing-authorisation/clinical-data-publication








Your perspective matters—join the conversation.