The problem
A clinical study report (CSR) is the document that sets out how a trial was run and what it found, and it sits on the critical path to a regulatory submission. Merck describes CSRs as traditionally labor-intensive, requiring teams of medical writers to sift through thousands of pages of clinical data over a period of months, resolve conflicting data, verify results and write in highly specific regulatory language (Merck).
Every week spent drafting is a week before a medicine can reach patients and start earning revenue. Merck also notes that companies interested in applying generative AI to this work had faced challenges building a platform that generates full CSRs of regulatory quality.
What they did
Merck framed the platform as part of a large-scale rebuild of its technology architecture to remove complexity from the clinical trial process. A joint Merck and McKinsey team of more than 80 data scientists, AI and life science technologists and medical experts, working across three continents, built a proprietary platform that uses large language models to draft CSRs. In the first half of 2025 the team produced and submitted the first live CSRs with it (Merck).
The design pairs table pre-processing with generative authoring: the platform maps tables, extracts data, applies styling and runs validation, and Merck says it can deliver a first draft in as little as five minutes. It is designed to operate under rigorous oversight by qualified medical writers. To adopt it, Merck says it revamped its operations and trained teams in prompt engineering, data pipelines and other skills needed to oversee the platform. At the time, the company said it was scaling the platform across its late-phase pipeline by the end of 2025.
In April 2026, Merck announced a partnership with Google Cloud under which it expects to invest as much as $1 billion over several years in AI infrastructure, engineers and Gemini Enterprise licences. Dave Williams, its chief information and digital officer, told Reuters that Merck had used AI to help prepare sections of CSRs for roughly two years and now plans to expand that approach (Reuters via BNN Bloomberg).
Results
All figures are self-reported by Merck and have not been independently audited. According to the company’s June 2025 announcement, in use across multiple studies the platform:
- cut the time to create a fully human reviewed CSR first draft from an average of 180 hours to 80 hours;
- reduced the number of errors in CSR drafts by 50%, in categories such as data, messaging, citations, terminology and typography;
- shortened the elapsed time to produce first drafts of CSRs from two to three weeks to three to four days.
The two time figures measure different things: 180 to 80 hours is effort on a reviewed first draft, while weeks to days is calendar time. Merck has not published how many studies were measured or how errors were counted.
A related result sits in market access rather than trials. Williams told Reuters that Merck has used Google’s technology to cut down by half the time and cost of compiling the dossiers many countries require before they reimburse a new medicine. “This isn’t a pilot,” he said, adding that the company is submitting dossiers in markets with it and scaling it globally (Reuters via BNN Bloomberg). That figure is a single executive statement, without a stated baseline.
What didn’t go to plan
No public source reports a setback or reversal, but the evidence is thinner than the headline suggests. Every number comes from Merck itself, in an announcement co-branded with the consultancy that helped build the platform. Merck has not reported publicly whether it met its end-2025 target of scaling across the late-phase pipeline; in April 2026 Williams described the CSR work as having “some success”, a more measured phrase than the original launch.
The gain also depends on people the platform does not replace. Merck stresses that medical writers oversee every draft, and the 80-hour figure is for a fully human reviewed draft. Regulators expect that: the European Medicines Agency’s reflection paper on AI in the medicinal product lifecycle says generative language models are “prone to include plausible but erroneous or incomplete output”, and asks for quality review of model-generated text before it goes to regulators (European Medicines Agency). The paper also puts responsibility on the sponsor to ensure that models and data pipelines are fit for purpose. Merck has not disclosed what the platform cost.
Lessons
- Automate the structured work before the prose. Merck’s platform pairs table mapping, data extraction and validation with generation, which is where a CSR’s errors in data and citations come from.
- Measure effort and quality, not just speed. Merck reported hours per reviewed draft and error counts alongside calendar time, which makes the claim testable.
- Budget for the operating model, not just the model. It took a team of more than 80 and retraining in prompt engineering and data pipelines to put the platform into live submissions.
- Keep the expert accountable for the output. Medical writers review every draft, matching the EMA’s expectation of human review before submission.
- Reuse the pattern on adjacent documents. Merck now reports a similar approach halving the time and cost of reimbursement dossiers.
Adapt it in your organization
Is this for you?
- You produce long, templated, regulated documents built from structured source data, such as study reports, dossiers or safety reports.
- Drafting sits on a critical path where days saved have a clear value.
- Your source data and tables are available in machine-readable form.
- You have experienced writers who can own review and sign-off.
- You can count errors in drafts today, so you have a baseline to compare against.
A 90-day version
Pilot: run AI-assisted first drafts in parallel with the normal process on three to five completed studies, comparing effort, elapsed time and reviewer-logged errors against the human drafts. Owner: head of medical writing. Success measure: reviewed first-draft hours down at least 30% with no increase in errors found at quality control.
Metric to track
Expert hours to a fully reviewed first draft, with errors found per draft at quality control. Merck reports 180 hours falling to 80 and errors halving, a ceiling to aim at rather than a baseline. For a comparable documentation result in clinical care, see the Kaiser Permanente case.
Failure modes to pre-empt
- Reporting calendar-time savings to the board when expert review effort is the real constraint.
- Letting plausible but wrong text reach a regulator because review was shortened along with drafting.
- Underfunding the data pipeline, so the model drafts from inconsistent tables.
- Scaling before you can show the error count, leaving the gain unverifiable, as Merck’s public record still is.
Questions to take to your leadership team
- Which regulated documents sit on our critical path, and what is a week of drafting worth?
- Who signs off AI-drafted text, and how would we prove to a regulator that it was reviewed?
- Do we measure error rates in drafts today, or would we have no baseline?
Executive FAQ
How much time did AI save Merck on clinical study reports?
Merck says the effort on a fully human reviewed first draft fell from an average of 180 hours to 80 hours, with elapsed time down from two to three weeks to three to four days.
Is Merck still using AI for clinical study reports?
Yes. In April 2026 its chief information and digital officer said the company had used AI on CSR sections for about two years and plans to expand it.
About the curator & AI disclosure
Curated by Luciano Ayres
Luciano Ayres is Director of Artificial Intelligence at ArcTouch, with more than 20 years of experience in software and technology leadership and specialized studies in deep learning. Connect with him on LinkedIn.
AI-generated briefing. This case was produced through an automated AI research and writing workflow curated by Luciano Ayres. Sources are linked for independent verification; summaries and analysis may contain errors and do not replace professional due diligence.