Everyone asks about hallucinations. The feeling is that if you take care of hallucinations, the product works. But we all know a regulated document is far more complex than that. It has to convey a coherent story, grounded on real data. And it's a big document. Drafting a coherent, big document that's also accurate is a very difficult task.
So accuracy is where the conversation starts. It shouldn't be where it ends. I've seen tools get the facts right and still get dropped, because getting the facts right was never the whole job.
Here's what has to work on top of accuracy.
The draft has to be built around the right story. A regulatory document isn't a stack of sections to fill in — it's making a case. A clinical study report tells the story of a trial. A submission tells the story of why the benefit outweighs the risk. If the AI hands back a draft where every section is there and every fact is cited, but it's organized around the wrong story — leading with what doesn't matter, burying what does — the writer can't just tidy it up. They have to take it apart and rebuild it around the right argument. That's not editing, it's a rebuild, and rebuilding a big document costs more than the blank page would have. You don't find out you've been handed a detour until you're already deep in it.
Whoever uses this is the one who signs off, so they have to be able to check it — fast — and believe it. Their name goes on the document. They answer for it in front of a regulator. So they can't just take the machine's word for anything; they have to verify it. And if verifying one claim means digging through source files for ten minutes, and there are hundreds of claims in the document, the writer does the arithmetic and goes back to doing it by hand. The tool has to show its work — every line pointing back to where it came from, so a check is a click and not an investigation. This isn't about how you feel about the AI. It's practical: can you confirm what it wrote faster than you could have written it yourself? If not, it hasn't saved you anything. It's just moved the work from writing to checking.
No first draft is right, so the writer has to be able to get in and really change it. The writer knows things the model doesn't — what the sponsor is trying to say, where the data is soft, which way the story needs to lean. So they need to jump in and not just swap a word, but move sections, change the framing, push the whole thing in a different direction — and the tool has to come with them. If changing one part quietly breaks three others, or the tool keeps pulling the draft back toward its own idea of how it should read, the writer ends up fighting the software instead of writing. That's usually when people give up on a tool — not because it got something wrong, but because it won't let them steer.
Then the quieter things that decide whether anyone actually keeps using it. It has to work inside the tools writers already use, not make them learn a new system before they can start — a lot of tools lose people right there, on setup, before a word is written. The sign-off has to be defensible, because a person is still attesting to the document. And it has to reflect the guidance that's current, not what a model picked up two years ago.
I build one of these, so factor in the bias. Though it runs the wrong way for me — I'd love it if there were a single thing to fix. There isn't. The tools that work get most of this right at the same time. The ones that don't tend to do one thing well, demo that one thing, and lose the writers on everything else.