The AI vendor numbers in medical publishing are real and they are large. Up to 75% MLR cycle-time reduction. First-draft clinical study reports from 180 hours to 80. Screening that took six weeks reduced to days.
What almost nobody publishes is the arithmetic that turns those into a number for your team. So here is one, fully assumption-stated. Disagree with the assumptions rather than the conclusion — that is what they are there for.
The model
A five-person publications team inside medical affairs. Roughly 200 productive hours per person per month, so about 1,000 team-hours. A realistic split, based on how this work is usually described:
- Drafting and revision — 35%. Manuscripts, abstracts, posters, summaries.
- Review co-ordination and MLR chasing — 20%. Routing, following up, reconciling comments.
- Literature and evidence work — 15%. Searching, screening, extracting.
- Tracking and status admin — 15%. The spreadsheet, the status meeting, answering "where is X".
- Congress and submission logistics — 10%. Deadlines, formats, portals, disclosures.
- Everything else — 5%. Meetings that are genuinely about strategy.
Now apply the published figures honestly
The mistake is multiplying each bucket by the vendor headline. That gives an absurd answer, because the headlines describe the compressible portion of a task, not the task.
Drafting, 350 hours. First-draft generation is where the big numbers live, but first drafting is perhaps 40% of drafting-and-revision time; the rest is revision after review. Compress the first-draft portion by half and you save around 70 hours, not 175.
Review co-ordination, 200 hours. Pre-review checks and content reuse genuinely reduce cycles. But if your reviewers are the bottleneck, faster routing produces a faster queue, not a shorter one. Assume a third: about 65 hours.
Literature work, 150 hours. The most compressible bucket, because screening is genuinely mechanical. Half is defensible: about 75 hours.
Tracking, 150 hours. Highly compressible and rarely attempted. Two thirds is realistic once state is centralised: about 100 hours.
Logistics, 100 hours. Deadline monitoring and format checking automate well; the human negotiation does not. A third: about 33 hours.
The total, and the honest reading
Adding those: roughly 340 of 1,000 team-hours — about a third. Not the 75% the headlines imply, and considerably more than nothing.
Three things worth noticing about where it comes from.
First, the largest single saving is tracking — the bucket nobody markets, involving no unpublished science, and requiring no security review. The least glamorous line is the biggest one.
Second, the drafting saving is the smallest relative to its size. It is the bucket everyone starts with and the one where the constraint is tightest.
Third, a third of team capacity is not a third of headcount. Recovered hours go back into work that was being skipped — the publication that slipped, the summary written properly rather than hurriedly. Teams that treat this as a headcount calculation usually get neither the saving nor the quality.
Do this with your own numbers
The model above is a worked example, not a benchmark. The version that matters is yours, and it takes an afternoon:
- Ask each person to split last month into those six buckets. Approximations are fine; you need the shape, not accounting.
- Multiply the compressible share of each bucket by a factor you can defend to a sceptic — half at most, and less where humans are the constraint.
- Rank by hours recovered per unit of difficulty, not by hours recovered.
- Start at the top of that list, which in most teams is not the interesting one.
And the prior test before any of it: if a process takes under two hours a week in total, no automation recovers its build cost. See AI for medical publishing for where those hours sit across the workflow, and AI tools for medical publishing for what addresses each.