It floors me that with all the AI/ML expertise being deployed on relatively narrow problems, the best productivity suites can do to predict which documents are relevant to your interests is "sort by access_time desc"

How about

"Team members who viewed this Gantt chart also viewed the project plan"
"You comment on 100% of documents with Bitcoin in title. Maybe this one also relevant to your interests."
"A paragraph you wrote copy/pasted to here; it is visible to you; maybe we should flag it."

"You searched, opened this document, and closed it 3 seconds later; probably not what you were looking for so maybe I won't put it in Recent Documents."
"You start every day by checking email and updating your daily action plan. I wonder what doc you will open today."

"In the collaboration software that we coincidentally own and that runs your entire business, the person who we know to be your manager told you to remember to update the Q4 planning document. There is a document named 'Q4 planning document.'"

"It's performance review time, a ritual know about because we are capitalists, too. If we couldn't have guessed the calendar event 'write perf review' a good clue. You will probably be looking for impactful work you have done. We have some suggestions based on various signals."

"Based on our calculations you just joined the 'incident review randomly-generated-name' meeting 30 seconds ago then opened docs. Every time you join an incident review, you open the appropriate doc. Lets take a wild guess at behavior over next 30 seconds"