The work is a chain, not a pool
Cleaning the dataset has to finish before the tables can be produced; the tables have to exist before the results can be interpreted; the interpretation has to be settled before the report can be written. You cannot shorten this phase by adding people the way you can shorten recruitment by adding sites, because each step consumes the output of the previous one. The phase ends when the last step ends, so every hour saved early in the chain moves the finish line forward by an hour.
Checking is where the time actually goes
A trial result is a claim that will be scrutinised for years, so the outputs are checked against the raw data and against the analysis plan that was fixed before the trial began. This is not bureaucratic padding. If a table is wrong, the error is discovered by someone outside the team, and the correction is far more expensive than the original check. The checking is also repetitive: the same consistency rules are applied to hundreds of tables and listings, which is precisely the kind of work that is slow for people and fast for software.
A queue, not just a task
Statistical and medical-writing specialists usually work on several trials at once. Even a well-run analysis waits its turn. This matters for judging AI's effect: automating part of the work shortens the task, but it also frees specialist time, which shortens the queue behind it.