Preparing a document process for automation

Buying an AI tool is the last step of adopting one, and a department that takes it first automates whatever it already does badly. The work that comes before the tool and around it is knowing why automation is pressed now and how far adoption has got, reviewing the process honestly, deciding which work is worth automating, standardising content, templates and data, aligning the functions upstream and downstream, choosing whether to build, buy or adapt a tool, defining it with its vendor, managing the change for the writers, proving it small before scaling it, and getting it validated and into a release cycle.

Pressure to automate

In May 2024 one head of medical writing at a contract research organisation held that a highly educated, rigorously trained writer may often find themselves doing menial technical work, and that there must be some technological solution that assigns people tasks making best use of their abilities, in keeping with her motto, think first and write later. ^{1, 2} In a 2022 panel's view, technology used well shifts a writer's focus from what AI, machine learning and other functions can accomplish to the creation of content. ³

Regulatory writers have spent great effort turning complex data into clear, compliant narratives, and in 2025 AI tools were changing that workflow: writers were expected to work where content has to be modular, reusable and fit digital workflows, and with global submissions, multiple indications and pressure on time to market, teams were exploring automation tools that generate and reuse content with less tedious intervention. ⁴ This means learning not only to create fit-for-purpose content but to configure it for automated workflows, structuring it so that it can be parsed, analysed and reused across the whole dossier. ⁴

In a 2024 account, new targets and treatment modalities were driving more trials — by some estimates the number running at any one time had grown by more than 200% since 2013 — and trials were more complex, with more data and endpoints in some areas and sparse data that must still support adequate power in others, such as rare diseases, while the workforce had not grown at the same pace. ^{5, 6} Regulatory writers remained underserved and overburdened. ⁷ Writers were caught between meticulous complex deliverables and fast turnarounds, which brought burnout and dissatisfaction, and even at small biotechnology companies with one or two assets the timelines were challenging and burnout was still an issue; sponsors increasingly outsourced to contract research organisations or other agencies, which brought its own misaligned expectations, extra review rounds, cost and time; and the automation tools of the day perpetuated inefficiencies because they were limited to very specific repetitive tasks. ⁸

In one vendor's 2024 account, the demand comes from a rapidly evolving trial landscape, with more biotechs, more studies, more complexity and tighter timelines, in which writers' expertise will always be needed while new technologies have an opportunity to help with tedious, redundant tasks. ^{9, 10}

In one consultant's 2025 account, the demand in regulated industries such as medical devices and drug–device combinations comes from too few resources for a large amount of work, in the industry, in medical writing and at the agencies and notified bodies alike, and from the need to follow what competitors are doing. ^{11, 12}

The state of adoption

In 2022 automation in writing had spotty adoption, depending on a company's size and digital sophistication. ¹³ In 2023 writing leaders assumed that company size explained the large differences in the rate and extent of adoption of advanced technology. ^{14, 15} Uptake was rapid and ongoing particularly within larger organisations, while smaller companies might find immediate adoption harder because of budget or other resource limits and might have to work on platforms their clients provide, and managing several iterations of technologies and processes could pose difficulties for smaller companies and contract research organisations. ^{16, 15}

Twenty-five people completed AMWA's 2023 survey of people who lead, train or develop medical writing teams, most (88%) from pharmaceutical, biotechnology or device companies or from contract research and writing organisations and the rest (13%) from consultancies, 96% focused on regulatory writing (4% on publications and other health communication) and 72% in departments of at least 20 writers. ^{17, 18} Sixty-five per cent worked where some form of technology was used to author content, and only 10% said there were no plans to implement any within two years; authoring technology was cited by 95% as in use or planned, formatting technology by 53%, review or editing by 47%, quality review by 42%, intelligent searching by 21% and electronic protocol builders by 16%. ^{17, 19} Information technology was named the heaviest user of technology for automation (53%), followed by biostatistics and programming (42%), data management (32%), other functions (21%), and data analysis, submissions and clinical (16% each), and structured content and assistive authoring were chosen as the most significant applications for writers in the next three to five years (65%), ahead of automated quality review (40%) and intelligent searching or digital data flow (30% each), respondents being able to select several. ²⁰

In 2023 writing groups already used structured content authoring and management, automated patient narratives, quality-checking tools such as PerfectIt, plagiarism checkers, review platforms such as PleaseReview, regulatory information management systems such as Veeva Vault, and Synchrowrite, a web-based tool that imitates Word without being affected by Microsoft's upgrade pushes; they were considering AI-assisted authoring of narratives, protocols and clinical study reports and internal GPT bots. ¹⁶ Writing leaders were interested in tools tested and validated for content reuse and structured content management. ¹⁶ Writing groups mostly used the same tools across specialties, the main difference being literature search-and-summary tools, adopted mostly in medical communication rather than regulatory writing. ¹⁴

In 2024, although GPT-3 in 2020 and then ChatGPT had drawn wide interest in AI for health care, its potential in automating regulatory submissions such as clinical study reports remained underrepresented. ²¹ Only a handful of providers offered AI-driven regulatory automation, those with working, proven solutions fewer still, and use in regulatory submissions, cross-reporting and the related areas was still marginal, even in the literature. ²² Another 2024 account took the FDA's receipt of more than 100 product registration submissions relying heavily on AI or machine learning in 2021 as a sign that AI's potential in clinical and regulatory writing was already being realised to some degree. ²³

In September 2023 many healthcare companies were employing AI, free-to-access AI tools such as ChatGPT were growing in popularity, and the use of AI tools to generate text within the industry was developing rapidly. ²⁴

The process before the tool

In May 2024, in the view of a medical writer who had used such tools for more than ten years, writers were excited about them and under pressure from senior management to be early adopters and to write documents faster, on the expectation that with ChatGPT a clinical study report is done within a click. ²⁵ There is no such magic button and no 99% perfect study report, and there is more to automation than going shopping for a tool: reaching for the tool starts with the solution instead of the problem underneath, and nothing is so useless as doing efficiently what should not be done at all. ^{26, 25, 27}

In the same writer's image, a tool is a pair of fancy running shoes: they do not get a runner over the finish line, or make them faster, without the training. ²⁷ The training is a process deep dive, a health check that every part of the process is appropriately handled, followed by a map of the process. ^{27, 28} Processes were designed for legacy systems, so the deep dive asks whether the SOPs have had a major revision and whether the team's whole approach to writing and reviewing documents has been challenged. ²⁸ A process that is not streamlined can generate problems from the start to the end product, and an AI tool laid over one with too many manual touchpoints keeps them; the time won on the first draft is carried forward only if the reviewers downstream do not undo it. ^{28, 29}

The reality check has to be honest. ³⁰ Listening to the noise shows how the process actually works today, and a process owner writing the SOP should listen carefully to what the end users say, which may show more challenges than expected. ³⁰ The brainstorming should take in all the subject-matter experts, not only the medical writers, and learn their struggles. ³⁰ Then one or two documents are walked through the process, with notes taken, to confirm the pain points. ^{30, 31} Where more is broken than can be fixed at once, what is customer facing comes first, because the customer should not have to share the pain, and the customers are the regulatory authorities and the patients, not senior management. ^{30, 31}

The fix should aim for the cure rather than another bandaid: AI is going to change the whole document development world, and a bandaid on a legacy process leaves a department flat on its face. ^{31, 32}

A technology company that ran automation as a service for more than a hundred clients put the same condition first for automation in safety in 2024: nothing can be automated while the basics are wrong, and the process has to be streamlined. ^{33, 34} Its second condition concerns data rather than documents. ³⁴ AI is only as good as the data behind it, garbage in and garbage out, and for every dollar companies spend on AI they spend seven on data, data cleansing and automation. ³⁴

In one May 2024 speaker's slogan for the data condition, there is no artificial intelligence without information architecture. ³⁵ The hard job that has to come first is harmonising and standardising the data and adopting ontologies and a common data model. ³⁵ The data-integrity problems run across the industry and inside individual manufacturers alike, there is no substitute for doing that hard job before harnessing the technology, and it takes the regulators and the manufacturers, proactively, working together. ^{35, 36}

Asked in May 2024 where the quick wins lie that give the most benefit for the least effort, one panellist repeated the slogan: get the information architecture in order before starting with AI, because workflows that are right and a use case fully known, including one's own strengths and weaknesses, are the starting point for saving time, with AI and without it. ^{37, 38}

Work worth automating

In the experienced writer's May 2024 view, a document or process suits automation when the document needs little customisation and its requirements are the same across therapeutic areas, or at least within one. ^{39, 40} The scoping should not stop at clinical study reports: the concepts apply to other document types in the group and beyond it, such as the publication team's and the case report form team's, so the tool to look for is one that automates a type of work rather than a type of document. ⁴⁰

The steps come next: a first-draft-only solution, a whole document, or only the back-end steps if the struggle is really the data import. ^{40, 41} The answer is specific to a company, its setup and its functions' needs, and the delivery options run from a closed-loop in-house system to a bought one to an in-house build, with the customisation a large sponsor needs as a further variable. ⁴¹

In 2024 one team had been trying approaches, some of them successful, to automating the protocol, the clinical study report, the investigator's brochure, safety narratives, the plain language summary, redaction or anonymisation and the informed consent form, with annual reports built from the development safety update report or the periodic safety update report as possible further candidates. ^{42, 43} A large share of documents is constructed from source documents, such as the statistical analysis plan, the tables and figures and the safety narratives, so technology can save a great deal of repetitive manual work and let writers concentrate on the science rather than on copying, pasting and quality control. ^{42, 43}

In the same team's view none of it is a button to click: what the technology assists is the time and the speed at which writers work. ⁴⁴

One 2024 proposal for implementing generative AI plans by kind of work rather than by deliverable. ^{45, 46} First, everyday AI, the efficiency gains of the working day such as writing an email or tidying a text, is separated from game-changing AI, which works at a different scale. ⁴⁵ Then, rather than taking one deliverable such as a patient narrative and doing everything for it until the solution is perfect, the plan identifies the fifteen or so things a clinical study report or a narrative needs, finds the ones common across deliverables, as text-to-text content generation applies to perhaps ten of them, and optimises that one horizontally so that it serves the whole cycle of deliverables. ^{46, 47} The content in view spans medical writing, publications, medical information, medical education, commercial content and quality-related content, which share a great deal and differ in nuance. ⁴⁸ Priorities are set on a two-by-two grid of feasibility against impact, starting with the front runners at high impact and high feasibility, which also makes the business case quick to approve. ⁴⁷

The same proposal treats generative AI as an engine rather than a car, around which something has to be built. ^{49, 50} Its limitations in 2024, bias, copyright questions, domain sensitivity, knowledge cut-offs and no training on the organisation's own data, belong to a technology dated from a 2017 Google paper, Attention Is All You Need, and changing fast, so an investment made now may be regretted when the same function arrives free of cost. ^{45, 51, 49} Four levers work around the limits: prompt optimisation, a basic skill quickly learned for minimal investment; retrieval-augmented generation, which retrieves information like a search before the prompt reaches the model and fortifies the prompt with it; fine-tuning an available small or medium model on the organisation's data, which takes time and technical skill; and building a medical writing model of one's own, at a cost of hundreds of millions of dollars and years of effort. ^{50, 52} The proposal recommends the first two, which its author felt was the right direction, in line with how the technology is developing. ^{52, 46}

Another panellist in the same May 2024 session held instead that the general models were not created for any specific kind of document or for medical writing, and that systems built and fine-tuned for the use case, though more work for the developers at the start, give good results for the specific task much faster than a general tool that can also generate rubbish. ^{53, 38}

In one former regulatory writing director's 2022 account, automation had helped writing in templates whose sections are prepopulated from other documents, such as a protocol from the investigator's brochure or the protocol concept document; in Word's own underused features, such as tagging text for reuse elsewhere in a document; in routine checks against the applicable style manual applied digitally; in review tools such as PleaseReview, which let a team see, track and incorporate all comments instead of circulating sequential versions; and in linking the integrated summaries of safety and effectiveness to the individual reports. ¹³ Registry results summaries and the FDA's study snapshots of safety by demographic characteristics could be populated as a clinical study report is written. ¹³ Automation seemed more useful in early phase 1 studies whose results are mostly data with straightforward assessments, in wearable devices recording results where tracking matters and participants can be unreliable, as in cardiac and diabetes studies, and in patient diaries, now possible on smartphones. ⁵⁴ Another gap it filled was routine text such as the safety narratives in clinical study reports, which are onerous to write where there are many variables or many participants with adverse events, as in oncology. ¹³

In 2022 aspects needing expert scientific knowledge and assessment might be producible by automation but still needed human evaluation and judgement, the medical assessment in a safety narrative was better written by a qualified medical writer, and machine-written text could take on a repetitive quality and be read as obviously machine-written and not properly evaluated. ⁵⁵ In another panellist's view, the sentience a human brings can make all the difference, especially in lay summaries and informed consent forms, and interpreting findings often means looking across data points, possibly in different reports, which algorithms may not be configured to do. ⁵⁵

Lean documents and standard content

In the experienced writer's May 2024 view, what writers can change themselves starts with lean writing: documents that minimise waste, giving exactly the information that is needed, at the right time, to the right audience, and nothing more. ³² Her advice is to create standards for every step of the process, starting with the templates, which have to comply with the regulatory guidelines and should also be checked for fitness for a new way of working: whether the structure can be streamlined, and whether effective cross-referencing can remove duplicated information. ^{32, 56} The TransCelerate suite of templates, from protocols to clinical study reports, was up to date in 2024 and lean, with a standard structure that can be implemented across therapeutic areas; adopting them or revising a company's own templates accordingly is a business decision, but they show the lean and content-reuse concepts. ⁵⁶

In 2018 writing executives observed that companies used to create document templates from their own interpretation of regulatory guidelines or simply from internal processes and needs, and had more recently begun collaborating to revise existing templates, such as the clinical study report, or to develop new ones where no industry standard had existed, such as protocols and lay-language patient summaries. ⁵⁷ Standardised templates would allow automation across related documents within a company and even across the industry, and AI is the step beyond automation. ⁵⁷

The standard text comes next. Anything usable across phases, therapeutic areas and study designs needs no AI and no reinventing, and goes straight into the template. ⁵⁸ Where a full standard would go too far, libraries serve instead, for endpoints, inclusion and exclusion criteria and contraception language, each holding what the group judges the best fit for all studies, from which a writer picks. ⁵⁸

A standard has to be policed, or teams override it. ⁵⁸ The discipline reaches the writer too: accept that a standard is a standard, stop customising it, advocate for it, be the role model for its use and limit the deviations. ⁵⁹ A cross-functional team that reviews the standards regularly keeps the wording aligned so that every function is content with it. ⁵⁹

The review process needs the same treatment, because fixing only the beginning of the process wins nothing. ⁶⁰ Everybody wants to work to a one-voice principle, and it should be enforced strictly, with senior management asked to back the writers and advocate for the change, since they are the ones asking for AI in the process. ⁶⁰ The end of the process needs a focused review by functional subject-matter experts; otherwise a good first draft from a tool is spoiled in the second and third rounds of review by personal preference. ^{60, 29}

Structured content authoring

In 2011 one pharmaceutical company described a programme that pairs structured authoring, the organisation of content into a standard hierarchy, with content reuse, the breaking of content into chunks so that the same chunk can be used in different places across documents, to reduce the effort and shorten the timelines of producing documents. ^{61, 62} In its view existing processes for authoring, reviewing and approving clinical documents create unnecessary workload and delays and perpetuate errors, inconsistency and submission-readiness compliance issues, and it aimed for time and cost savings of up to 30%. ^{62, 63}

The programme was a system of technology solutions and processes meant to enable four capabilities: concurrent authoring of clinical documents, reuse of content across documents, a high level of consistency and accuracy in the information presented, and progressive review of content at either the component or the document level. ⁶³

Its method has five activities:

• Identify the bottlenecks and redundancies in the current workflows.
• Establish a library of searchable components for reuse or repurposing.
• Create document maps that list the components in the order appropriate to each document type, and identify their dependencies and locations across documents.
• Provide a user-friendly authoring environment that lets authors concentrate on scientific content rather than format.
• Design outputs of the final documents specific to their end use. ⁶⁴

Its criteria for success are these:

• Full management support for the big picture and the required budget, and IT support for the technology requirements.
• Comprehensive training and support, with an intuitive authoring environment, to foster enthusiastic user acceptance of and confidence in the new way of writing documents.
• Incremental rollouts of technology and process that give targeted user populations quick wins.
• A logical content and document architecture.
• Processes for governing the creation and modification of reusable content components, to optimise the structured authoring and content-reuse methodology and to ensure adherence to company and health-authority standards. ⁶⁵

In one former regulatory writing director's 2022 account, structured authoring had been hard to adopt because the applications initially did not support Word documents and formatting was a problem, staff needed training, and an authoring tool often did not integrate well with older systems; it is easier for a new operation, or a company at its initial public offering, to start with it than for a large organisation to scrap its old systems. ⁵⁴ Cost counts, and so does technical competency: Word itself is not used to its full capacity, document experts are often not writers, and many writers are not sophisticated technology users. ⁵⁴

In 2025 the industry was moving from manual, unstructured document development to a content-based approach using structured content management (SCM) tools, aiming to streamline workflows and improve consistency and efficiency, and structured content authoring (SCA) was emerging as a key enabler for integrating AI into regulatory and medical writing. ⁶⁶ SCA can incorporate standard content selected by metadata into configurable templates. ⁶⁷ Parexel implemented SCA for clinical-study documents and periodic safety reports in 2022 and then added preconfigured prompts and the incorporation of user-generated AI text; its qualitative lessons, reported in 2025, claim shorter production time, better first-time quality and improved content strategy, and describe programmable prompts as reducing adoption barriers and allowing targeted control of how prompts are used while offering flexibility to generate or modify content. ^{68, 67} It concludes that combining SCA with generative AI streamlines workflows, improves summarisation and quality control, and may accelerate time to market while maintaining compliance. ⁶⁹

Asked in May 2024 where automation in nonclinical study report writing stood, one panellist said the industry was still pondering and evaluating its options, and the trend across industries was to look at structured content authoring first and then use it as the base from which to see whether automation or AI could be moved into. ^{70, 71} In the panellist's words the sequence was first we try, then we trust; AI was a buzzword and being on some kind of AI initiative had become very important, but the panellist did not know whether AI was ready for medical writing at that point. ⁷¹

In one May 2024 speaker's account, structured content authoring organises content into predefined structures or ontologies, and AI tools need clear structures in place to function properly; a structured content management system also accumulates content in a database from which, in future, approved, quality-checked content could be pulled to train AI, a stage not yet reached. ^{72, 73} The other points of contact run along the life of a document: an AI summarising for wherever the content has to be published once the outline has been produced through structured authoring, ingestion of legacy content into the platform at the start, and content or language quality checks further down the line. ⁷³

Template and data readiness

In a 2025 account, the state of an organisation's content standards is one of the most important factors in successful adoption; outdated templates, inconsistent formatting and siloed styles, without a shared approach to how documents such as clinical study reports present data and messages, make automation much harder. ⁷⁴ A clear, consistently implemented template supports automation and reuse. ⁷⁵ When study-level protocol and report content is structured for reuse it can cascade into summaries, investigator's brochures and briefing packages with little rework, an "intelligent content cascade" that lets writers shift from redundant authoring to messaging and interpretation. ⁷⁵ Large organisations whose teams are divided by function or therapeutic area tend to diverge in templates and standards, and readiness means realigning to common templates, style guides and standards. ⁷⁵ ICH M11's structured protocol guideline, technical specification and template show what readiness looks like, providing a common structure and a standard for exchanging protocol metadata electronically, and at a minimum consistent, descriptive headings let AI tools find content to reuse. ⁷⁵ Integrating AI into the writing process needs scale and standardisation to save real time. ⁴ Standardisation also limits individual preferences in style and formatting, and consistent content improves quality control, streamlines review and supports dossier assembly and approval; writers focus on what to say, the key message and the significance of the data, while technology manages how it is said. ⁴

In one former regulatory writing director's 2022 view, all such uses of automation require standardised terms and definitions, and a concept that helps in reusing information is that text must also be considered data, a document being a compilation of data elements. ¹³

Data readiness is the other half. ⁷⁶ Clinical study output tables are designed for human readers, with merged headers and indentation that machines struggle to interpret, and many AI tools work better with machine-readable formats in which each data point's relation to its descriptors is explicit. ⁷⁶ A technology team may ask what formats the data come in — SAS files, RTF tables in Word, CSV in Excel, or structured JSON, HTML and XML — and some tools can use the CDISC data sponsors already submit, writing patient narratives from individual data in SDTM or ADaM format. ⁷⁶ CDISC released the Analysis Results Standard in 2024 to make summary statistics and endpoint analyses more standardised and machine-readable, and the push for standardisation supports such output standards across sponsors. ⁷⁶ At a minimum, a consistent format for each type of table, such as an overview of adverse events, across studies lets a tool turn tables into text consistently and easily. ⁷⁶

In one vendor's 2024 account, its software has to solve an end-to-end data integration process defined from before database lock, covering data acquisition, data processing and data mapping, for which the industry is short of standards; then instant and accurate text generation; and then integration with the authoring and review tools, so that the creation chain is efficient. ^{77, 78}

What standards cost and save

In the experienced writer's May 2024 view, standards have a ledger with entries on both sides. On the gain side, what is standardised up front does not need automating, which saves time in authoring, review and quality control, and significantly reduces the risk of errors and inconsistencies and probably of inspection findings. ⁷⁹

On the cost side the first entry is stakeholder management, probably the biggest ask of anyone going back to change a process for AI. ⁷⁹ Acceptance of the standard has to be created: teams and subject-matter experts need to hear that it is not there to be a killjoy, to kill their creativity or to overrule their personal preferences, but to get the standardisable material out of the way and free up headspace. ^{79, 80} That headspace is for critical thinking, which AI does not do and which is the job of the human in the loop, so the point of leaning up the documents is time to think about the results and the messages instead of commenting that an abbreviation was not defined at first use. ⁸⁰

Lean writing is a fine balance between including too little and too much, an art that needs highly skilled medical writers, so a move to it is also a question for development plans: training, chaperoning, skilling up. ⁸¹ A standard needs a maintainer, as a prompt library does: someone with the mindset to advocate for standards and for continuity, and a firm enough stand to say no to customisation requests, without whom the discussion about customising the standards never ends. ^{81, 82} And the tool carries a price tag, whether as a licence, a cost per year or a cost per study. ⁸² Time spent exploring and piloting the tools is fine, but time is very well spent on the process deep dive, which makes processes created for legacy systems AI-ready. ^{82, 83, 84}

Getting reviewers and stakeholders behind a lean approach rather than the status quo has many answers, and two worked for the same writer. ⁸⁵ Divide and isolate the stakeholders and the reviewers, and convert them one after another. ⁸⁵ And put the concepts into short training material watched before the conversation starts: five- to seven-minute videos on lean writing and strategic review, in a comic style, helped a great deal because they were fun and people remembered them. ^{85, 86}

Alignment with neighbouring functions

In the same writer's view, a leaner process is not the writing department's to build alone. ²⁹ The growth mindset it needs runs across departments, not only among the medical writers: collaborators have to come too, aligned on the same goal, so that the functions further down the line do not hinder the improvement. ²⁹

The conversation to have is with data management and biostatistics, about standardising the presentation of the tables, listings and figures, a standard set always needed for every study. ⁸⁷ In May 2024 writers were still receiving final outputs that were not ready to copy and paste; AI will probably solve part of that, but it is in the writers' own interest that what arrives is closer to what they need. ⁸⁷

Behind that sits a question about the information flow, put as an aspiration rather than as practice. ⁸⁸ The standard in the SOPs is to start the report from the final tables, figures and listings, while in reality several iterations of them are delivered. ^{88, 89} If a tool can exchange the outputs or the dataset behind an in-text table at the click of a button, the question is whether the report has to wait for the final set, or could start from an earlier version or connect to the database directly. ⁸⁸ Writers who have learned to work from draft outputs have already developed the skills to handle the uncertainty and the potential change, which is the argument for starting earlier. ⁸⁹

Front loading applies the same idea to the parts of a report that do not depend on the data. ⁸⁹ Sections standardised in the template need no attention when the data arrive, and a report written to the TransCelerate template largely disposes of the question of future and past tense, because the introduction reduces to about two sentences. ⁸⁹ A protocol written with the report in mind leaves paragraphs that can be carried straight across with a few verbs exchanged, which automation can also do, without anyone having to massage the text. ^{89, 90}

Writers have to own their whole process to be in control. ⁹⁰ Close contact with the biostatistics team and the data managers keeps the tools and the processes aligned, so that there are no surprises. ⁹⁰ The same holds for the deliverables: the process should be checked for whether one global package can suit all instead of customised local and regional variants, so that automation is not asked to produce many scenarios. ³⁹

Disconnected systems, data and expertise

In one consultant's 2025 account, the obstacle a medium to large pharmaceutical organisation meets when it moves into combination devices is how little its systems are integrated and its people collaborate, and it has three layers. ^{91, 92} The first is systems. ⁹¹ Such a company runs an electronic quality management system, a product lifecycle system, and business functions using whatever they use, one team in a word processor and another in a shared drive, and most of those systems do not talk to each other; nor does the regulatory department talk to medical writing or to the people preparing the submissions, although everyone works towards the same output, a report to a notified body or an agency. ^{91, 92}

The second layer is data: the same work done repeatedly by functions that do not talk to each other. ⁹² The regulatory function, the vigilance function, clinical affairs and medical affairs each go separately to the same literature and authority databases, PubMed and the German and Swiss authorities' databases among them, and download and monitor continuously through the year. ⁹² The inputs are the same, and the crosstalk that would ease the final assembly most often does not happen. ^{92, 93}

The third layer is expertise, and it shows in combination products: one person in the organisation knows the drug side and another the device side, and questions such as when a notified body opinion is needed, and when to talk to the notified body rather than the medicines agency, fall between them. ^{93, 94} The aim is to bring that interdisciplinary work together and let technology take some of its load. ⁹⁴

Drafting the report before the final data

In 2024 one pharmaceutical company's global medical writing team had adopted a rules-based technology for clinical study reports, which is deterministic: the writer configures the application up front and it gives exactly what was configured. ^{95, 96} In the account of the team's digital innovation lead, the clinical study report is the example, but the principle applies to any data-driven results document that interprets data as its source. ^{95, 97} The desire to go faster is the one the pandemic taught, and manual work can take a team only so far. ⁹⁷

The method moves the writer's responsibility to the left rather than removing it. ⁹⁸ Dry-run datasets and tables, listings and figures arrive months before database lock, depending on the company and its timelines, and with the rules-based application the whole report is drafted from them in advance, the data sections as well as the protocol-dependent sections that the old shell covered. ⁹⁹ When the final statistical analysis is available the document is refreshed: the new outputs and the interpretations tied to them flow in, numbers change and text readjusts programmatically, and the writer defines the clinical messages on the new data. ^{99, 98, 100} Before the analysis, the writer and the study team agree the content, the structure of the text, which outputs matter for interpretation, and the key messages. ⁹⁸

Tested on ten study reports, the approach let writers go more than 50% faster after the final data were available in 70% of them: a first draft that had taken ten days took three to five with the pre-work and the technology. ¹⁰⁰ The other 30% were outliers because of the study teams, whose members could not always stop themselves commenting on the text, a well-meant habit whose change will not happen overnight. ¹⁰¹ The same principle was expected to hold for other results documents, the clinical summaries among them. ¹⁰¹

Writers cannot deploy the technology by themselves: it takes every stakeholder on the study and report review team, whose insight into the business rules improves the application. ¹⁰² The technology is not a silver bullet but an enabler that works only alongside the other acceleration enablers, which are robust up-front planning, simplified processes, data flow and system integration, and data and document standards. ^{102, 103} Two teams show the difference. ^{103, 104} With one, change-management sessions at the start set the expectation that the writer's text would not be revised unless it was incomplete or inaccurate, and the team, with a growth mindset, accelerated significantly after the final data. ^{103, 104} With another, everything else was in place as well, but the study team rewrote and paraphrased the earlier drafts, and the pre-work could not be used. ¹⁰⁴

The challenges still to solve before the approach becomes business as usual are these:

• Capability building and upskilling: the systems are new, and at first they probably save no time but take an investment of time in learning a new interface.
• Creating and adopting data and document standards, because once the way data are generated is standardised, standardising the output becomes easier.
• System integration: each application on the market serves one use case, and a writer should not need one application for automated redaction, another for style checking, another for inserting outputs and another for interpretation to draft one document, hence application programming interfaces and connections.
• Evolving the system on feedback from across the functions.
• Mitigation plans for the unforeseen, since the issues that arose were more about process than about the tool.
• Change management, for the writers and for the contributing authors and reviewers, while a sudden explosion of new applications makes choosing between them hard. ^{105, 106, 107}

In the lead's view the writer's role changes shape. ¹⁰⁸ The primary skills had been authoring and scientific expertise, with some project and stakeholder management; with technology the writer drives accelerated submissions through strategic authoring, and the role becomes more holistic, driving standardisation, influencing, and helping teams adopt the application, with routine content left to the technology and clinical interpretation and strategic authoring, the work writers like most, left to them. ^{108, 96} The tools need patience: a first experience falls short of expectations, and they evolve only with experience, more business rules and more trials, while regulatory requirements and guidelines keep changing. ¹⁰⁹ Asked where the quick wins lie, she pointed to reuse: a system configuration built for one therapy area could be reused and repurposed for others, which is why the writer embeds business rules meant for every study design and therapy area rather than for the one report in hand, though in 2024 there was still a long way to go before it scales. ^{37, 110}

Building, buying or adapting a tool

In a 2025 account, organisations choose whether to build a custom tool, buy a specialised platform or adapt a general-purpose tool such as ChatGPT. ¹¹¹ Building offers the most control and integration but demands substantial investment, development resources and maintenance for the tool's life, complicated by how fast AI changes. ¹¹¹ Adapting a general tool is accessible and flexible, appealing for experimenting, but carries data security risks, limited automation, no version control and no safeguards against hallucination; prompting section by section may not save much time, and long regulatory documents often exceed the context window. ¹¹¹ Buying a specialised platform brings security, compliance support, automation, maintenance and user communities, and regular enhancement, but may require workflow changes and data-sharing agreements. ¹¹¹ The right choice depends on more than feature comparison: on the organisation's goals, data-security policies and long-term scalability. ¹¹¹

In a 2024 account, a general model falls short at a clinical study report because large clinical studies yield complex and copious data about participants, and general models such as GPT-3 and GPT-4 were neither designed nor optimised to understand how those data sets relate to each other, are not clinically trained and are prone to errors and hallucinations, most evidently in complex areas such as oncology and in large studies. ²¹ Another 2024 account held that generative platforms must be fit for purpose, because general models trained on all data do not suit so regulated an industry, and that platforms built for biopharmaceutical use cases and trained on biopharmaceutical data can produce properly formatted, compliant documents and a consistent narrative across a project, and can be built around data protection, privacy and platform validation. ¹¹² To meet the data security and privacy concern, platforms should be built on strict data protection, cryptographic methods and minimal collection and retention. ²³

In 2023 writing leaders suggested ways AMWA could help on the buying side: decision trees for choosing technology; an analysis of the tools available; a standard questionnaire for vendors to complete about their tools, with vendor sandboxes showing writing use cases; a portal on the association's website where vendors update their information as their tools improve and the community posts reviews; a selection toolkit for managers with tool profiles and comparisons; a benchmarking exercise on key regulatory considerations, as a metric for evaluating tools; and successful business cases published in its journal. ^{113, 15}

Writers' part in defining the tool

In one panellist's 2026 view, medical writers are the ones using the tool and should not have to deal with something simply handed to them; they need to be involved in creating and defining it and in saying what they want it to do, rather than accepting what it does. ¹¹⁴

In one regulatory writer's 2026 experience, very enthusiastic technical people have produced outputs without understanding what the output needed to be, or what is important and critical for a particular document. ^{115, 116} The writer's role is to orchestrate the content so that in the end it is what is needed. ¹¹⁵ What programmers may not get from a template and ICH guidance is the field's institutional knowledge of what the health authorities need and how a document supports their decisions. ^{116, 117}

Many writers feel they cannot build an AI, agentic or generative, and in another regulatory writer's 2026 view the discussion has to start from them. ^{118, 119} AI is coming with its pros and cons, and when a writer sits with the IT person it is the writer on whom it depends how the AI gets built. ¹¹⁹ Because hallucinations are still present, human intervention and human quality control will always be required. ¹¹⁹

In 2018 writing executives cautioned that technology cannot be the driver, and that medical communicators should be empowered to guide the development of AI and automated systems. ⁵⁷

In one panellist's 2026 view, involvement continues after delivery. Having used a tool, the writer should go back to whoever built it and say which parts they liked and which they did not, because those are what the developers and vendors have to address; it should not be a passive arrangement in which a tool is handed over and does things. ¹²⁰ For the same panellist, involved in an AI tool, one request, for certain sections to be produced in pink, called for patience, and a degree of flexibility has to be built in anyway, because no one size fits all. ¹²⁰

The same panellist's advice is to specify first and shop second. ^{121, 122} A writer works out what their dream AI tool would do and takes those aspects to the vendors to ask whether theirs does them, rather than seeing what a tool happens to do. ^{121, 122} The ambition can be large, as in no longer spending three days on a wretched table, while AI is not magic and there may be things it cannot do. ¹²² A fundamental paradigm change in an occupation is challenging, and for that reason the profession cannot sit by and let it wash over it, but has to take hold of it and make it do what it wants. ¹²³ If writers do not put more input into these tools, what they end up with will be unsatisfying and unhelpful and will make the job harder than it already is. ¹²³

In May 2024 one panellist who had worked for several developers of electronic data capture and real-time safety data collection systems held that nobody has a better perspective than the medical writer on the flow of data from the collection point to the final product, so that perspective is imperative in developing software that collects, analyses, interprets and presents information; without it, software is developed by "a bunch of propeller heads" who "don't have a clue". ^{124, 125}

At one firm that spends a great deal on AI, the most repetitive reason a build sometimes turns out badly is that the person who runs the work was not part of it, and the speaker from that firm credits his own start in medical writing with letting him explain to heads of research and development where all the data in a submission come from. ^{125, 126} At a software company the point is symmetrical: without a mix of technology and domain expertise the result is either a tool with no use, because the use case was never explained to the technologists, or nothing at all, because the subject-matter experts are not technologists. ¹²⁶

In the technology company's 2024 view, medical writers have to be part of building automation and AI for medical writing, and the aim is to reach a point where the effort-intensive, error-prone, time-consuming and repetitive tasks are taken over and writers spend their time reviewing and ensuring that the right conclusions are drawn. ¹²⁷ In that company, AI specialists sit side by side with the medical writer, see what the writer does day to day, and check what can and cannot be automated before redesigning the process. ¹²⁸

In a former lead medical writer's 2023 account of one AI literature-review platform in development, the medical writing team took part in AI sanity verification, conceptualised decision trees and validated their logic, and gathered high-quality clinical data on the safety and performance of medical devices, which machine-learning engineers then used for training, in a human-in-the-loop approach. ¹²⁹ Once the models were running, the team created literature review projects to assess their accuracy and sensitivity. ¹²⁹ In the same account such work should discredit a prevailing misconception about the origin and quality of training data: given the stringent regulatory domain, such tools go through repeated assessment and validation to meet regulatory and legal expectations, and are built jointly by regulatory and medical writing professionals and software and machine-learning engineers. ^{129, 130}

Working with the vendor

In a 2025 account, implementation means structured collaboration with the vendor through evaluation, configuration and deployment, beginning with research, demonstrations and hands-on evaluation. ¹¹¹ Writers on such projects should expect to take part in pilots and user acceptance testing and to supply real data and templates for evaluation; sales and customer success teams usually lead, but internal writing leads review content quality, usability and integration with existing workflows. ¹¹¹ Configuration and deployment mean adapting to software development practice: development, staging and production environments, and feedback through tickets and feature requests. ¹¹¹

Change management for writers

In a 2025 account, it is tempting to apply new technology to existing ways of working, but new approaches can speed the move to AI-assisted submissions: revise templates for standardisation, consistency and fit-for-purpose lean writing; train teams on the benefits of automated generation and standardised templates; and update workflows, for example reviewing, locking and quality-checking data-independent content before database lock. ¹³¹ For high-achieving professionals, changes to long-standing practice can be unsettling, and fears of losing value, career prospects or standing can turn into scepticism or resistance; successful initiatives prioritise change management: clear communication of how roles will change, affirmation of each contributor's value, small steps to achievable goals, and a defined process for everyone to follow. ^{131, 132}

In 2023 writing leaders were expected to manage expectations about the money and time needed to configure and maintain tools, which may be significant, to communicate with corporate leadership about return on investment, risk and alignment with organisational objectives, and to be mindful of the needs of their writing and formatting teams, with change management strategies easing the transition. ^{16, 15} In the 2023 survey, management was introducing the technology through town halls, training sessions and electronic communication, some respondents wanted hands-on training, and no novel approaches were proposed. ¹³³

In 2018 writing executives saw the challenge as incorporating AI tools into work processes and preparing writing teams to use them, with communicating the tools' benefits essential and the security of AI a concern. ⁵⁷

Proofs of concept and scaling

The technology company's route in to automation for safety is a proof of concept, on a scientist's reasoning: prove something works before making it big, test it in a sandbox or test environment, and take it to production if it works. ¹³⁴ By 2024 the smaller individual use cases it had delivered with clients included the triaging of cases, the resolution of queries and intelligent quality control, where effort starts to fall and quality to rise. ^{134, 135} Aggregate report authoring was on the roadmap rather than done, with the case processing in pharmacovigilance in which people go through mountains of source documents to put cases into the database that the aggregate report writer later analyses. ¹³⁵

The direction is real-time safety, with data, analytics and insights arriving continuously rather than on a six-monthly, annual or two-yearly cycle; reaching it was expected to take five to ten years, and what was being done in 2024 were baby steps. ^{136, 135} The end state is about data rather than documents: one clinical study's data can feed a clinical study report, a summary document, a safety document, a publication and a commercial poster, and the data are the same throughout, so the full breadth of automation is a set of tools and assets that produce all those views from them. ¹³⁷ An automation of 2012 pulled the tables and listings out of the system into a Word document, then the most painful part of the work, and that is where a programme starts: with a small process improvement whose benefit can be seen, moving on from there towards the vision. ^{135, 137}

In a 2019 account, implementers should be realistic about what artificial intelligence can do and follow a graded plan in five levels. ¹³⁸ Level 1 is the manual generation of documents in silos; level 2 puts structures and processes in place to generate and reuse content manually as efficiently as possible, and most medical writing teams were at one of those two levels; level 3 adds early automation, guided by prudence and significant human oversight; level 4 automates content generation across some silos and document types; and level 5 has automation take over most content generation across all silos, with minimal expert oversight. ^{139, 138, 140} Iterative processes that need minimal critical thinking could be automated first, complex tasks could be divided into smaller iterative ones to widen what can be automated, and the journey should begin with the content pieces that are most reused before moving up the value chain. ¹³⁸ Artificial intelligence has to stay coupled with human intelligence throughout, with sensible automations rather than broad generalised approaches and a long-term strategic outlook taken in measured steps. ^{138, 141}

In one industry analyst's 2022 view, the industry was stuck in "pilotitis", running one pilot to see whether a technology works, which is unsurprising in so regulated an industry, when the question was no longer why to automate but how to implement at scale; returns on any technology come when it is implemented at scale, so a company that judges a solution on a pilot's returns many times rejects it, which should not happen. ^{142, 54} The other barriers were change management; writers' concern about being replaced, answered by their moving up the value chain, with routine work taken by AI and machine learning and the writer ensuring that data are represented and interpreted correctly and the messaging is right; skill development for staff who are not all technically confident; and transparency and regulatory compliance. ⁵⁴ In the former regulatory writing director's view, writers had been reluctant to embrace automation for fear it would replace them, but not everything they write deserves their full attention, and offloading what can be offloaded leaves them free for the critical sections. ⁵⁴

Validation and updates

In one panellist's May 2024 experience, a tool deployed for work supporting a GxP submission has to follow all the standard procedures, with the audit documents in place and the system auditable at any point. ^{143, 144} In the same panellist's view, nobody had a hundred percent automated report going to a health authority: the tools help, and the tools that are used are validated. ¹⁴⁴ A pharmaceutical company should look for IT validation of any tool it uses; that is outside the medical writer's scope, the company's IT department will be all over it, and a tool that is not validated gets no further. ¹⁴⁴

In one vendor's 2024 view, data privacy and compliance are strong requirements from the regulator and must be enforced at any cost, whatever the back end for text generation, so that whatever is sent to an application programming interface is never disclosed to anyone and the audit trail requirements are met. ¹⁴⁵ No one can promise a hundred percent accuracy from generation with large language models, by design, but with the end user in full control of the validation process such tools can still be useful assistants. ¹⁴⁶

One vendor's update cycle was quarterly minor updates, major versions once or twice a year, and bug fixes and patches within a couple of weeks of an issue being found, all handled by the IT team through transports under the support and maintenance agreements rather than by the business, and meant to leave the writers blinded to it. ¹⁴⁷ In another vendor's view, release management and software versioning are not enough on their own: a strong quality engineering process, which that vendor makes contractually mandatory by design, makes sure that what is delivered does not break what was delivered before. ¹⁴⁸