Somewhere in every medical affairs department there is a spreadsheet. The publication plan. The MLR tracker. The congress calendar. The author-disclosure log. Often all four, in one file, across seven tabs.
It has a colour code that only one person fully understands. It has survived a reorganisation, a platform migration and, frequently, the departure of the person who built it. It is the single most load-bearing object in the department, and it appears on no system diagram anywhere.
It is not a sign of backwardness
The reflex in AI content is to treat this as evidence that a team is behind. That reading is both unkind and wrong.
Somebody built that file because nobody supplied a system. They mapped a real process, encoded genuine institutional knowledge, and kept it current through changes that broke other things. The spreadsheet was right. It has simply outgrown its container. Any conversation about replacing it that does not begin from that premise will fail, because the person who built it is usually still in the room.
Where the risk actually is
Roughly 88% of spreadsheets contain errors. In most industries that is a rounding difference on a forecast. In medical publications it is something else:
- An author disclosure that was never collected, discovered at submission.
- The wrong version sent for MLR review, so the approved asset is not the one that shipped.
- A congress abstract deadline noticed the week after it closed.
- A publication counted as complete on one tab and in progress on another, so nobody chases it.
None of these are exotic failure modes. They are the ordinary consequence of state being held in a file that several people edit, that has no validation, and that cannot tell you what changed or who changed it.
Why this is the right place to start with AI
Counter-intuitively, the most valuable first automation in medical publishing is not the drafting. It is this.
- It holds process metadata, not science. Titles, dates, owners, statuses, versions. No unpublished manuscript leaves your environment — which matters, because Wiley, Wolters Kluwer and Elsevier all warn against putting unpublished material into general-purpose LLMs.
- Nothing needs a security review. That removes the single biggest delay on a first project in this industry.
- The payback is immediately visible. People notice within a week that they stopped asking each other where things are.
- It builds the audit habits you will need later. Logging who changed what, when, and who approved it is exactly the discipline required before AI touches anything closer to content.
What "automating the tracker" actually means
Not a dashboard. Dashboards show you the state of a file that is already wrong. The work is upstream of that.
- One source of state. The status of a publication lives in one place, and every view derives from it rather than duplicating it.
- Automatic reconciliation. The tracker reads from where work actually happens — mailboxes, review systems, submission portals — instead of waiting for someone to update a cell.
- Exception surfacing. Not "here is everything", but "these four things are overdue, this one is missing a disclosure, this deadline is in nine days and nothing has moved".
- An audit trail by default. Who changed what, when. Free once state is centralised, impossible in a shared file.
The honest limits
Two things this does not do, and any supplier who implies otherwise is overselling.
It does not fix an undefined process. If two people genuinely disagree about when a publication counts as "in review", automation will encode the disagreement rather than resolve it. That conversation has to happen first, and it usually takes an afternoon.
It also does not, by itself, speed up MLR review. Veeva reports up to 75% cycle-time reduction with AI-supported review, but that figure assumes the operating model changed too — pre-review checks, content reuse, tiered workflows. A better tracker tells you where the queue is; it does not shorten it. What it does is stop the queue being invisible, which is usually the first step to shortening it.
Before you automate anything, count
Two numbers, and you can get them in an afternoon without spending anything: how many times a week does someone touch the tracker, and how long does one reconciliation take. Multiply.
Under roughly two hours a week in total, leave it alone — no automation recovers its build cost. Between two and ten, a modest tool or a small automation is usually right. Above ten, especially where several people are maintaining overlapping views of the same truth, it is worth building properly.
We have set out where the hours sit across the whole publishing workflow in AI for medical publishing, and compared the tools by job in AI tools for medical publishing.