Drafting regulatory documents with AI authoring tools
An AI authoring tool changes how a regulatory document is drafted. Some of the work passes to the tool and some stays with the writer: tables, listings and data checks can be automated, the text has to serve a machine reader as well as a human one, and drafting runs in stages around database lock. The writer sets down the rules the tool follows, including its thresholds, and guides the tool before it runs. The tool has to state its criteria and report what it cannot do, and its output still has to be checked.
What the tool does and what stays with the writer
In 2023 leaders of medical writing departments across the industry held that advanced technologies should focus on giving a good first draft, drawn from 100% verified data and the company's formatting requirements, so that writers are more efficient and concentrate on content rather than on data verification, formatting and style-guide adjustments. They wanted tools to include an option that discloses where extracted information came from, a transparency that may reduce the need for extensive quality control and for routine tasks such as emailing preidentified section reviewers. In their view it is unacceptable for a tool to be creative, adding fictitious references or producing output inconsistent with its input, and it is of no advantage to writers for the technology to recalculate or run additional analyses. Conclusions, the nuances of regulatory interactions and contextualisation are, in their view, to be left to the writer. In the leaders' view, advanced technologies are excellent at summarising and condensing both data and information. Harmonising messaging across sections, documents or submissions, as distinct from copying identical phrasing, needs in their view a human-level understanding that language models are unlikely to attain.
According to a journal editor's summary of a 2025 article by two employees of a vendor of AI tools for regulatory writing, AI-driven tools let writers move past routine tasks such as formatting and repetitive drafting and focus on developing clear, compliant and well-structured key messages for regulatory authorities.
Automated tables, listings and data checks
Employees of a generative AI company held in 2024 that automating tables, listings and figures shortens CSR authoring and improves quality assurance by flagging potential errors, inconsistencies and missing data in the databases used. On their account, documents submitted to the FDA will then inherently have higher data quality and be devoid of errors, with faster approval times and a significantly smaller chance of rejection for errors.
Authors at another generative AI company held in 2024 that generative AI can also be used to better manage the processes around a document's data inputs:
- ingesting, reconciling and synthesising large amounts of data in a fraction of the time a human would take
- cleaning, transforming and organising data so that it is ready for analysis, instead of writers sifting through it manually, checking quality on several screens and reconciling divergent formats and terminologies
- automating tasks such as data entry and validation, which reduces the risk of human error, and detecting inputs that deviate from the source documents, which can reduce the time spent on quality control
Text for a machine reader
Lean authoring, the two vendor employees wrote in 2025, was originally developed with manual writing and human reading in mind. It was optimised at the level of the individual document to save writing, quality control and review time, and it deliberately minimised repeating table data in the text, because human readers interpret patterns in tables easily and numbers in text are costly to create and check. As regulatory writing shifts towards automation and global dossier strategies, lean authoring is adapting: long text with accurate data points can now be authored in seconds, though getting AI to write in a tightly controlled, concise format takes additional development effort.
To support an intelligent content cascade across the dossier, the aim shifts from minimalism to fitness for purpose: the text must stay concise but carry enough context to be understood on its own. This matters all the more as health authorities begin to use AI systems to assist in their review. Human reviewers interpret data in tables easily, but large language models may struggle to tell which parts of a table matter most, how schemas should be interpreted or how to follow contextual links, so including key data points and core messages in the text may make an accurate AI summary more likely.
A drafting workflow around database lock
The two vendor employees give an example, in two stages, of how AI-enabled regulatory writers already generate documents in a fraction of the time.
Before database lock, the writer leads the first draft:
- Gather the source documents, dry-run data outputs or shells and the Word template, and connect them to the draft.
- If required, make the data machine-readable by linking each data point to the terms text generation needs, so that a mean age of 65 years in the placebo group is linked to years, age, mean age and placebo.
- Review the AI configuration template so that the right data and sources feed each section and the content plan fits the document's purpose, adjusting it as required.
- Populate data-independent sections such as the introduction and study design with a click, and add context or enhance the text with the tool's options.
- Review and lock those sections with the team; quality control in the tool shows the reviewer the source text and any changes.
- Review the data-to-text plan, such as describe, compare groups or create an in-text table, with the team, and adjust it as needed.
After final data are available, the writer completes the document:
- If required, convert the final data into a machine-readable format.
- Populate most data sections, such as disposition and safety, with a click, and enhance the text.
- Match efficacy tables to predefined endpoints and generate interpretive and descriptive text with the statistical analysis plan as context.
- Review and lock the data sections with the team; quality control in the tool shows the source table and any edits made to the generated text.
The rules a tool drafts by
Every medical writer follows some rules, whether or not a standard operating procedure writes them down, in the view of both a founder adviser of a vendor of a rules-based drafting platform and a leader in a pharmaceutical company's global medical writing function. In effect the profession already has such rules, the adviser argues, and does not want to acknowledge them. He agrees that statistical analysis plans do less in this direction than they could, and puts it down to people not wanting to freeze the interpretation in stone the way they do for primary efficacy. So in his view the rules need flexibility, including the flexibility to say that a case is not clear and needs a physician to look at it.
In the writing leader's experience, discussing the rules with the team up front, before the data exist, is extremely valuable and saves time. The rules may have to change once the actual data arrive. She holds that the rules-based approach makes a team think up front rather than wait, and that much then comes down to how easily the rules can be changed and how much work that is, which may depend on whether an organisation runs the same kind of study over and over, as a large pharmaceutical company may, or runs one study once. Asked how much programming a change takes, the adviser answered that his platform had, for several settings, a default definition of clinical relevance that the user could toggle, so that a user who ran it and was not satisfied could easily go back and change it, and he holds that crucial for the reasons the writing leader gave.
Thresholds and the middle ground
In one attendee's experience, rules that define clinical meaningfulness the same way across studies of different phases, drugs and diseases do not work; another agreed, adding that depending on the platform some allow for this while a pure language model likely will not. A third pointed out that non-significant results can still be clinically relevant where there is a clear dose response in a very small population, guessed that this could be programmed in a rules-based deterministic model, and wondered how often formulating the rules takes less time than going through the data, which probably depends on how much data there is.
The adviser agreed that one-size-fits-all rules are never going to work. A clinically relevant difference in adverse events between two treatment groups, even once defined, differs in small studies: if it is set at 5%, then with twenty participants a difference of one person is 5%. It also depends on where the difference lies: 97% against 92% may not be relevant where 5% against 0% would be. So the rules are more complex than one might think, and rules alone would still leave trouble. What a writer wants are rules that are generally useful but in the end conditional: if this, then this, and if that, then that. This is why, in his view, the tool needs guiding: one of its first questions is what kind of study this is, such as a phase I study in a small number of healthy volunteers, and the answer tells it what to look for.
In the adviser's account, a physician at a German motor-vehicle authority, assessing drivers whose licences had been taken away for drink-driving, looked at their liver values: with one abnormal value the driver might be telling the truth, with four or five they were lying, and two or three fell in between. An AI writing a sentence from that kind of evidence should, in his view, be expected not just to say black or white but sometimes to say that a case is grey, and the grey has to be part of the rules.
He recalls US companies that held a data deep dive, in which writers were locked in a hotel for a week or so and went painstakingly through every table with a yellow highlighter. A rules-based system, in his view, should be set a bit conservatively, because it should find everything: rather than treating only a 10% difference as interesting, it should give a short printout with all the interesting material, the small differences marked as not that interesting and the big ones standing out. The threshold he tended to see over his career was 5%, with occasions when something different was used.
Questions the tool asks before it runs
A tool that asks the writer nothing before it starts is not a good system, in the adviser's view, because parameters have to be set for the AI to understand and follow, and it cannot be let loose on data and expected to know what everything means. In a dose-limiting toxicity study with ten treatment groups and a larger group compared against an active comparator, the tables could easily have fifteen columns, and an AI simply pointed at them would compare everything with everything and produce complete chaos. So the writer has to guide it to what should be looked at and compared and to the kinds of difference to look for.
The adviser finds a stepwise agent, which first determines the endpoint and then applies endpoint-specific interpretation from a large library of prior interpretations, an interesting approach, but holds that finding enough generalised findings for it to work from will be harder than one might think. Efficacy is the exception: p-values and 95% confidence intervals are used in much the same way by everybody, so a generative system could probably just apply 0.05 and 95% and report significance against them. Asking the writer questions before the tool runs is, in his view, itself in a way such a stepwise process.
Writers who had used such rules-based tools told the adviser that what they liked was being made to stop and think about the data and the study before they started. He was surprised they had not done so before, and holds this key.
A writer at a company that had built its own rules-based drafting tool reported the same from its users in May 2024. The tool opened with a series of questions about the study, the comparative groups, the endpoints and whether to change the adverse event cut-offs. The writer answered these instead of writing prompts, so no writer had to become a prompt engineer. Unexpectedly for the company, its writers liked being guided to think about the report differently: because they knew there would be time savings, they had time to think before writing, which was a luxury. Because the writers were all guided in the same way, the output was more consistent even after a human had worked on it, and the tool prevents key things being forgotten.
Stated criteria
A tool cannot simply report findings as clinically relevant, in the adviser's view, because the authorities dislike that and ask how it was known and what the definition was. A good tool therefore has to state the criterion, clinically relevant meaning greater than or equal to whatever figure was used, so that the writer can decide whether they agree, see what the tool did, and change the figure, to 3% for instance, if they think another would be better.
The company that built its own tool designed it to show more rather than less, on the reasoning that deleting a sentence is easier than crafting one from scratch. Its standard settings were conservative, but a writer who disagreed with a rule or a parameter could change it, and it stated its rules and cut-offs so that the writer knows what is happening.
What the tool reports it cannot do
In September 2023 two authors whose company sells a rules-based drafting tool held that from a medical writing perspective a conservative approach is always best: in their experience it is better for a tool to highlight where something is missing or where interpretations cannot be made, flagging data points for the writer to investigate, than to produce a complete but misleading draft.
The key, in the adviser's view, is not to make the system arrogant in what it infers. As an example of his own arrogance, he once wrote a rule that if the tables, figures and listings have no death table, the only reason is that there were no deaths, which he thought would hold 99.9% of the time. The next client's output had no death table, the demonstration duly said there were no deaths, and the client said there were: it put deaths in listings. The rule was changed so that the tool reports that no death table could be identified in the output, rather than drawing a conclusion from the absence, so that the writer knows what is going on and nothing is guessed. Key to a good system, in his view, is not to train it simply to conclude that something with a beak and webbed feet that lays eggs is probably a duck: the exceptions, such as the platypus, have to be kept in mind, because the 80% guess is tempting but 20% is not nothing.
A client once brought the adviser a method of analysing dose-finding studies that neither he nor a statistics professor had heard of, and AI cannot be expected to have heard of it either. What he would have hoped a good rules-based system would do is say that it is unfamiliar with the statistical analysis used and cannot interpret it; a tool that covers the gap by hallucinating puts the writer in big trouble.
Medical writers like to know what is in the document and where it came from, in the view of the writer whose company built its own tool, so that tool was designed to say so: where it could not populate a table or did not know what to do with something or where it came from, it said so and why.
That tool met an unknown method too. The company asked a client for some old data to show what the tool could do, and one statistics table used a method the system did not recognise, so it reported that it did not know the rule. Not even a former statistician on the company's staff recognised the method, a very old one, and the client's answer was that it had stopped using it a decade before but had been asked for old data. The tool knows only what its builders put into it, will not go onto the internet to find out, and will not make something up to fill a gap: it leaves the section blank, populates the rest, and leaves the writer to ask why.
Checking the output
Medical writing leaders at a 2023 forum discussed concerns that companies using assistive authoring to generate content take risks with the reproducibility of content or with verifying the accuracy of outputs. The concerns they raised were mainly with generative tools, whose outputs need verification and can be subject to hallucinations in regulatory documents. In their view, quality control becomes imperative with advanced technologies, and covers not only grammar and syntax but verifying that generated content is free of inaccuracies and hallucinations. Because manual errors disappear from sections a human did not write, validated tables can reduce the need for data verification, and quality control moves towards content and context: in the not-too-distant future, reviewers may take responsibility for evaluating context and meaning, assessing the accuracy, consistency and relevance of the content.
In AMWA's 2023 survey of people who lead, train or develop medical writing teams, human oversight was the most mentioned mechanism for validating the technology, including comparing documents produced with AI against manually produced ones and developing procedures from ongoing experience with the tools. Yet 15% reported no requirements or best practices for validation in their organisation, and 25% that these were unknown.