How to Organise Shared Files So AI Tools Can Use Them

Give each kind of work one home folder, name files so the date, client and document type are readable, keep one current version and archive the rest, make scanned documents searchable, and tighten sharing first, because Copilot and Gemini can see every file the person using them can open. A small team can do this over three afternoons.

AI assistants that search your drive don't understand your folders the way your team does. They look for files that match the question and that the user is allowed to open, then answer from whatever they find, including last year's price list if it's still sitting there. So the work splits into three afternoons: sharing first, then structure and naming, then retiring old versions and making contents readable, with a ten-question test at the end to prove it worked.

Follow me on Instagram@sagnikteaches

How AI tools find your files, and why old files cause wrong answers

When you ask Microsoft 365 Copilot, Gemini in Google Workspace or a chat assistant connected to your drive a question, it searches the files you have access to, picks out the passages that look most relevant, and writes an answer from them. Two consequences follow.

Connect on LinkedInSagnik Bhattacharya

Permissions decide what it can see. Microsoft says Copilot only surfaces organisational data that the individual user has at least view permission for. Google says Gemini has the same access to Workspace data as you do, with one useful exception: if a Drive file's owner has blocked downloading, copying and printing, Gemini can't access it even when it's shared with you. So a file shared too widely is a file every AI user in the business can pull into an answer.

Subscribe on YouTube@codingliquids

Everything relevant competes. If three versions of your price guide exist, the tool may quote any of them. It can't tell that "Prices FINAL v2" was replaced by "Prices 2026 new" unless something in the files says so. Most wrong answers from AI search in small businesses trace back to this: not a faulty tool, but an out-of-date file that nobody retired. What ChatGPT's connected apps (formerly called connectors) can see in your Drive and inbox covers how this works for ChatGPT specifically.

Afternoon 1: fix sharing before anything else

Do this first, because tidying folders doesn't help if the wrong people can open them. Work through five checks:

1. Find files shared with "anyone with the link". Open the sharing settings on your main folders and on the files you've sent to clients and suppliers, and look at the general access setting. Most link-shared files were sent once to a client or supplier and never closed. Whoever administers your Google Workspace or Microsoft 365 account can tell you whether your plan includes a sharing report that lists them faster.
2. Remove former staff and freelancers from shared folders and shared drives.
3. Lock down the sensitive folders. HR, payroll, contracts and anything with health or financial details about clients should be open only to the people who need them.
4. Check client folders. Does everyone need every client's folder, or only the people working on that client?
5. Move files out of personal drives into shared locations. Work saved on someone's own drive or laptop is invisible to colleagues and to the AI tools they use, and it leaves with the employee.

Check 1 usually produces a long list, and it sorts quickly into three piles. At the wedding business described below, the 140 link-shared files went like this: about 100 were old mood boards and timelines sent to couples whose weddings were long over, so their links were simply switched off; around 35 were current planning documents, re-shared with the named couple and suppliers instead of "anyone with the link"; and five were brochures and the public price sheet, which stayed link-shared because they were meant to be public. The pile that mattered was the second one, which included two signed contracts.

Check 5 has a trap of its own. Move files into the shared drive; don't copy them. One planner copied her live wedding folders across and kept working in the originals out of habit, so for three weeks every run sheet existed twice, and the AI quoted whichever it found first. A move leaves nothing behind to compete.

Here's how skipping this tends to surface. A junior planner asks the assistant, "What margin do we usually make on a full-planning wedding?", expecting a rough rule of thumb. It answers with an exact percentage and cites "Owner - pricing workings 2025.xlsx", a spreadsheet the owner had shared with everyone in the company two years earlier so a bookkeeper could see one tab. Nobody had opened it since. The AI did nothing wrong: the file was shared, so it was fair game. Check 3 above would have caught it.

If you're on Microsoft 365, cleaning up SharePoint permissions before turning on Copilot goes step by step through the SharePoint side, which has more places for over-sharing to hide.

Afternoon 2: one home for each kind of work

A folder structure that works for AI is the same one that works for a new starter: shallow, predictable and with one obvious place for everything. Three to four levels deep is enough for most small businesses. Here's an illustrative structure for a wedding planning business:

Company shared drive
  00 Reference  (current versions only)
     Price guide
     Contract template
     Supplier directory
     Policies and FAQs
  01 Weddings
     2026
        2026-06-13 Smith-Jones - Riverside Barn
           01 Contract and payments
           02 Planning (timeline, run sheet)
           03 Suppliers
           04 Guests and seating
           05 Inspiration and photos
     2027
  02 Marketing
  03 Finance        (restricted)
  04 Team           (restricted)
  99 Archive        (superseded and closed work)

The numbers keep folders in the same order for everyone. "00 Reference" is the most important folder for AI: it's where the tool should find the current answer to "what do we charge?" or "what's our cancellation policy?". Keep only current documents there. Each wedding gets the same five subfolders, so a question like "what's the final head count for the June wedding?" leads to the same place every time.

A business organised around ongoing clients rather than one-off events needs the levels the other way round. An illustrative four-person bookkeeping practice put the client first and the year second, because most questions start with the client:

01 Clients
   [Client name]
      00 Engagement and permissions   (engagement letter, access list)
      2026
         01 Bank and card statements
         02 Working papers
         03 Filed returns and reports
         04 Correspondence            (decisions written up, not raw threads)
      2025

The same principle holds in both: identical subfolders for every client or job, so "where are last year's working papers for this client?" always has one answer.

File names that people and AI can read

A good file name tells you the date, who it's about and what it is, without opening it. That helps search tools pick the right file, and it helps you check the answer. A simple rule: date_client_document-type_version.

Year-month-day dates sort correctly and can't be misread. Don't rename thousands of old files by hand. Apply the rule to new files and to anything in "00 Reference" and the current year's work; archived material can keep its old names.

An AI assistant can draft the new names if you give it the old ones with the facts it needs, but watch what it does with gaps:

Rename these files using date_client_document-type_version.
Use only the dates and client names I've given; if either is
missing, write MISSING rather than guessing.

run sheet FINAL final.docx   (Smith-Jones wedding, 2026-06-13, 4th draft)
Scan0034.pdf                 (signed contract, no other details)
seating plan (2).xlsx        (Smith-Jones wedding, 2026-06-13)

ILLUSTRATIVE REPLY
2026-06-13_Smith-Jones_run-sheet_v4.docx
2026-06-13_Smith-Jones_signed-contract.pdf
2026-06-13_Smith-Jones_seating-plan_v2.xlsx

The first and third are right. The second ignored the instruction: the scan came with no client or date, so the assistant borrowed both from the neighbouring lines. The client happened to be right; the date wasn't, because the contract was signed in February, not on the wedding day. A signed contract filed under the wrong date or the wrong wedding is a serious error, so open every scan before accepting a name for it. The "(2)" becoming "v2" is also a guess worth checking against the file's history.

One current version of everything

This is the rule that prevents the most wrong answers. For every document that gets updated (price guides, contract templates, supplier lists, policies), there should be exactly one current version outside the archive.

• When a document is replaced, move the old one to "99 Archive" the same day, or delete it if it has no record-keeping value.
• Put the effective date inside the document too, in the first line: "Price guide, effective 1 January 2026". AI answers often quote the text, and a date in the text makes a stale answer obvious. Asked "What do we charge for day-of coordination?", an assistant replying "Day-of coordination is $1,450, according to the price guide (effective 1 January 2025)" tells you in the answer itself that it found the wrong file. Without that line, the same wrong figure reads as fact.
• Delete true duplicates: the "Copy of…" files and the same attachment saved five times from five emails.

The one legitimate exception is a document with two versions in force at once. Each autumn, the wedding business starts taking bookings for the following year at new prices while this year's couples stay on the old ones. Both guides live in "00 Reference", and each says in its first line which weddings it covers: "Price guide for weddings taking place in 2027, effective for bookings from 1 October 2026". Asked "What would a 2027 wedding cost?", the assistant can now tell the two apart, because the text says which is which.

If you only do one thing from this tutorial, do this for the files in "00 Reference". Then give that folder a single owner, the one person allowed to add or replace files there, and a ten-minute check at the end of each month to confirm nothing old has crept back in. Shared folders stay tidy when someone is answerable for them; they drift when everyone is.

Afternoon 3: make the contents readable

AI tools read text. Some files look full of information to a person and are nearly empty to a machine:

• Scanned PDFs without a text layer. A scanned contract is a picture of text. Run it through the text recognition (OCR) built into most PDF and scanning apps, or open it in Google Docs, which converts images of text. Going paperless before you add AI covers scanning properly.
• Photos of documents taken on a phone. Same fix, or re-scan.
• Spreadsheets built for printing, with merged cells, blank spacer rows and notes typed in random cells. Keep one clean table per tab. The wedding business's supplier directory was a typical case: a merged title row, suppliers grouped under bold category headings with blank rows between, and "(closed now)" typed in red next to one florist. Rebuilt as one table with the columns Category, Supplier, Contact, Phone, Email, Status and Last used, the same information became something an AI tool could filter, and "Status: Closed" replaced the red note it couldn't see.
• Slides where the content is an image. A pasted screenshot of a timeline carries no searchable text.
• Emails saved as PDFs with long quoted threads. Where the decision matters, write it into the relevant document instead.

You don't need to convert everything. Start with "00 Reference" and the current year's client folders, which is where most questions will land. For large volumes of paperwork, AI document processing turns PDFs and scans into usable data at scale.

Three afternoons at a six-person wedding planning business

The business in this illustration plans about 40 weddings a year, with roughly 9,000 files across a shared drive, two planners' personal drives and a laptop. Before the clean-up, a planner asked their AI assistant what the business charges for day-of coordination and got last year's figure, from a price guide saved inside an old wedding folder.

• Afternoon 1 (sharing): 140 files were shared with anyone holding the link, including two signed contracts. A former freelancer still had access to the whole weddings folder. Both fixed in about three hours.
• Afternoon 2 (structure): the owner set up the numbered folders, moved current reference documents into "00 Reference", and gave the planners a week to move their live weddings into the standard five-folder layout.
• Afternoon 3 (versions and readability): seven old price guides and four contract templates went to the archive, and the signed contracts for current-year weddings were run through text recognition.

Asked the same pricing question afterwards, the assistant quoted the current guide and named the file it came from. The owner's verdict was that the clean-up helped the team as much as the AI: new planners stopped asking where things were. A wedding planner's AI workflow from enquiry to final timeline shows what that team built on top of it.

Test it with ten questions before you rely on it

Write ten questions whose answers you know, covering reference material, client work and something that should be off limits. Ask them through the AI tool your team will use, logged in as an ordinary team member rather than the owner:

1. What do we charge for our main service?
2. What's our cancellation policy?
3. Who is our usual supplier for [something you buy often]?
4. What's the date and venue for [a current client]?
5. What's outstanding on [a current client's] payment schedule?
6. What changed in our contract template this year?
7. Where is the signed contract for [a client]?
8. Summarise the planning notes for [a current project].
9. What is [a colleague]'s salary? (Should fail: restricted.)
10. What did we charge in [two years ago]? (Should come from the archive and say so.)

Pass mark: eight correct, the salary question refused, and every answer naming a current file.

The wedding business's first run, logged in as a planner, scored 7 out of 10 (illustrative):

A retest a week later scored 9, with the salary question still refused. The remaining miss, a payment schedule question, turned out to be a planner who hadn't moved one live wedding out of her personal drive yet.

If a question returns an old version, find and archive that file. If the salary question returns anything, go back to afternoon 1. Once the drive passes, you're ready for the next step: building a company knowledge base AI can answer from, which goes beyond tidy files to pages written specifically for AI to answer from.

For the vendors' own descriptions of how access works, see Microsoft's Copilot data, privacy and security page and Google's page on what controls Gemini's access to Workspace data.

Further reads

• AI Security Checklist Before Connecting Tools to Email and Files — The security checks to run before connecting any tool to your files.
• Gemini Notebook (Formerly NotebookLM) for Small Business Teams — Point AI at a hand-picked set of files instead of the whole drive.
• Copilot in Word: Draft Proposals From Your Own Files — What tidy files make possible: proposals drafted from your own documents.
• Business Data Backup Checklist Before You Connect AI Tools — Back up the drive before you start moving thousands of files.
• How to Prepare Your Business Data for AI, Step by Step — The equivalent job for records and spreadsheets.
• Staff Offboarding Checklist for AI Tools and Shared Accounts — Keeps sharing clean when people leave.
• How to Review an AI Tool After 90 Days: Keep, Fix or Cancel — An evidence pack, a ten-point scoring sheet, a worked split verdict for four Copilot seats and a checklist for cancelling without loose ends.
• Do You Need a Lot of Data to Use AI in Your Business? — How much data each common AI task needs, from three example emails to two years of sales history, and why scattered data is the bigger problem.
• Can ChatGPT Read PDFs, Spreadsheets and Photos? What Breaks — What ChatGPT reads well in PDFs, spreadsheets and photos, the features that make it misread or skip things, and a five-minute test before you trust a figure.
• How to Create SOPs From Screen Recordings With AI — Record a task once, talk through the why, and let AI draft the procedure. Covers Gemini video uploads, Loom, Scribe and Tango, redaction and the cold-run test.
• Airtable vs Google Sheets for Business Data AI Can Use — Which home suits business data you want AI to read: Sheets for flat lists, Airtable for linked records, with prices, limits and a worked school example.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Learn on data, privacy and security for Microsoft 365 Copilot; Google Workspace help pages on what controls Gemini's access to Workspace data. Checked September 2026.