A senior EA at a mid-size firm told me she stopped reorganising her shared drive halfway through a Saturday afternoon project. Her principal had asked her to find a contract from 2022 that referenced a specific clause, and she had been planning to fix the folder structure first so this kind of thing would be easier next time. Then she dropped the folder into NotebookLM, asked it which file mentioned the clause, and had her answer in eleven seconds. She closed her laptop. The folder structure is still a mess. It no longer matters.
The bigger shift in how organised EAs work isn’t better folders. It’s the realisation that most of the time you spent organising files was time spent compensating for search that didn’t work. Now search works. The compensation is no longer needed.
Why the old organising playbook is mostly obsolete
The classic EA advice (consistent naming, dated versions, a clear folder taxonomy) was correct for a world where finding a file meant remembering where you put it. That world ended quietly in the last eighteen months. Tools that read the contents of files rather than just their names now make the question “where did I save it” the wrong question. The right question is “what was in it.”
This isn’t a small change. It’s the same shift that happened when web search replaced bookmarks. The EAs who internalise the shift early get a quiet, compounding edge. The ones who keep refactoring their drives every quarter are doing the digital equivalent of re-alphabetising a filing cabinet that nobody uses anymore.
NotebookLM is the tool to start with
If you only learn one new AI tool this year, make it NotebookLM. It does one thing well: take a set of documents (PDFs, Google Docs, slide decks, transcripts, web pages) and let you ask questions in plain English with answers that cite the source line. For an EA, this is the closest thing to a searchable institutional memory you can build without IT involvement.
The workflow is unromantic. Create a notebook for each long-running thread you support: a board pack, a client account, a strategy refresh, a hiring round. Drop in everything related (briefing docs, meeting transcripts, prior emails saved as PDFs, slide decks). When your principal asks something at 4:55 p.m. on a Friday, you ask the notebook, and you reply with the answer plus the citation. The citation is the part that matters. It tells your principal you didn’t make this up, which is exactly what AI critics say AI is doing.
A specific failure mode to watch for. NotebookLM will not pull from documents you forgot to add. If you set up a client notebook in March and your principal asks you in October about a contract negotiated in July, the answer will be confidently wrong unless you’ve been keeping the notebook current. Add new documents on a recurring weekly slot. Fifteen minutes a Friday is enough.
The retrieval prompt that beats keyword search
For one-off questions across a large drive, ChatGPT and Claude both accept folder uploads (with size limits) and can run reasoning across them. The prompt that produces the best answer is more specific than most EAs initially write:
“I’m looking for information across the documents I’ve uploaded. The question is: [QUESTION IN PLAIN ENGLISH, AS IF YOU WERE ASKING A COLLEAGUE]. Tell me which specific documents contain the answer, quote the relevant lines verbatim, and tell me which documents you looked at but did not find anything relevant in. If you cannot find a clear answer, say so explicitly rather than guess. If documents disagree with each other, flag the disagreement. The context I am asking in is: [WHO IS ASKING, WHAT THEY WILL DO WITH THE ANSWER].”
The instruction to list the documents that did not contain the answer is the part most people skip. It’s the instruction that catches the model when it’s narrowing too aggressively. If you ask “what was the deadline for the merger filing” and it tells you “March 14” without telling you it only looked at one folder, you have a problem. If it tells you “March 14 according to file X, files Y and Z were checked but did not contain a deadline reference,” you have a properly defended answer.
The folder cleanup prompt that actually saves time
Some folder reorganisation is still worth doing, particularly when you inherit a predecessor’s drive or take on a new principal. The prompt for this is not “rename my files” (which AI cannot do, it has no file system access). It is “design the convention I will then apply, and write me a script if I want one”:
“Below is a list of [N] filenames from a shared drive that supports [PRINCIPAL ROLE]. The current naming is inconsistent. Please: (1) identify the underlying categories the documents fall into, based purely on the names. (2) Propose a naming convention that captures category, date in YYYY-MM-DD format, document type, and version. (3) For each existing filename, give me the proposed new name in a two-column table, original then proposed. (4) Flag any names where the existing information is too ambiguous to rename safely, do not guess. (5) If I want to apply these renames in bulk, give me a Google Apps Script or Python snippet I can adapt. Filenames: [PASTE].”
You’re not asking the AI to do the renaming. You’re asking it to do the thinking, which is the part that took an afternoon. The actual rename is a five-minute job once you have the table.
A counterintuitive insight from EAs who have run this prompt: the categories the model proposes are usually better than the ones a human would have written, because the model is reading the actual filenames rather than the categories you imagined when you set the folder up two years ago. The drive has evolved. The model sees the evolution. You usually didn’t.
What organisation actually means now
The old definition of an organised EA was someone who could put their hands on any document in under thirty seconds. The new definition is someone whose principal never has to wait for an answer that lives somewhere in the firm’s collective memory. Those are not the same skill. The first one is about your drive. The second is about your toolkit.
Adopt one search-capable tool. Build the habit of asking it questions before you start scrolling. Keep one or two notebooks current for the threads that matter most this quarter. The rest of the cleanup work, the colour-coded folders and the perfect naming conventions, is now optional in a way it wasn’t two years ago. Spend the time you used to spend on it on the work that actually compounds: the briefings, the relationships, the things your principal cannot have an AI do for them.