How to turn a Google Drive into an AI knowledge base

A Drive becomes a knowledge base when an agent can find the current version of a fact without guessing. That takes structure more than a new tool. This is the order we work in, with a printable checklist at the end.

Published 28 September 2026 by Attacca

Most shared drives grew one folder at a time, named by whoever needed it that week. People cope because they remember where things are. An AI agent has no such memory. Given a folder with three versions of the same proposal, it will answer from whichever one it reads first, and it will sound equally sure either way. The steps below remove the guessing.

1. Take an inventory before you move anything

Export a list of every file with its name, type, owner, size, last-modified date and folder path. A spreadsheet is fine. The count alone is useful: at one creative agency we worked with, a single client group held 6,213 files, with one brand name spelled two ways across all of them and the same logo duplicated dozens of times. Nobody knew that until the list existed. Write the totals down so you can show what the clean-up removed.

2. Dedupe, then triage in priority order

Group files with the same name and size and keep one. Drop system files, temporary exports and screenshots from one-off checks. Export native Google documents to a plain text format, because an agent reads text far more reliably than a layout. Then fill the shelves in the order people actually need things. For a brand team that order was brand books, manuals, logos, fonts, the latest results report, the last two content grids, then approved pieces. The same agency put its rule plainly: better 600 MB useful than 800 MB of copies.

3. Name things one way

Pick one pattern and apply it everywhere: client, project, document type, then a date in year-month-day order, such as acme_spring-launch_brief_2026-09-14. Agree one spelling per client before renaming anything. Words like final, new and v2-real carry no information; the date does that job. When the name says what a file is, both a person and an agent can tell which version is current without opening it.

4. Put every file on one of five shelves

We use five top-level shelves. Projects hold work with an end date. Clients hold one folder per client, for the things that outlive any project. Areas are ongoing responsibilities such as finance, hiring or sales. Resources are reference material the team reuses: templates, brand guidelines, process maps. Archive holds finished or inactive work, kept but out of the way. Every file lives on exactly one shelf. If a file seems to belong on two, it usually belongs to a project and should link to the client.

The pile is inventoried, deduped and renamed before anything lands on a shelf. Canonical facts and the task ledger sit beside the shelves as structured records. The agent reads all of it with read-only access; anything it proposes to write waits for a person.

5. Write canonical facts once

Each client gets one short file of facts that should never be guessed: legal name, the people you deal with, what you sell them, the numbers that come up in every report, and the house rules. Each fact carries its source and the date it was confirmed. When someone corrects a fact, the correction goes here and nowhere else. At a Florida marketing agency, a client’s active store count had been assumed at 37; once verified at 33 and locked in the client file, the wrong number stopped resurfacing in drafts.

6. Keep a task ledger next to your task tool

You do not need to replace the tool your team already uses for tasks. The ledger holds what that tool usually lacks: where a commitment came from (which meeting, which email), who owns it, what counts as done, and a receipt when it is closed. It is what lets an agent answer a question like “what did we promise this client last month”, and it is how a commitment survives when the person who heard it leaves.

7. Decide what is a table and what is text

Swipe to compare

Where each kind of information belongs
Kind of informationStore it asWhy
Tasks, assets, client fields, anything you count or filterRows in a database table or a well-kept spreadsheetYou need to ask how many, which ones and since when
Briefs, process maps, house rules, meeting notesText documents, in plain text or Markdown where possibleThe meaning lives in sentences, and agents read text well
Logos, videos, layouts, signed contractsFiles that stay in Drive, referenced by pathThey are evidence and assets, not something to rewrite

8. Grant read-only access

Connect the agent through Google’s OAuth consent with a read-only Drive scope, from an account you control and can revoke in one step. Read-only means the agent can find and quote files but cannot move, rename or delete them. Anything it wants to change, such as a new task or a corrected fact, should arrive as a draft for a person to approve.

9. Point an agent at it, then test it

Put a short instruction file at the root that explains what each shelf means, where canonical facts live and what the agent must never do. Label instructions, references and examples separately; one agency learned this after a reference image placed in the middle of a slide was read as a locked creative decision. Then ask five questions you already know the answers to, and check each answer against the file it came from. If two of the five are wrong, the fix is almost always structure, not a better model.

The checklist

Print this page, or work through it with your team.

Inventory

  • Export a list of every file: name, type, owner, size, last modified, folder path.
  • Write down the total file count and total size, so you can see what the clean-up removed.
  • Mark the folders nobody on the team can explain.

Dedupe and triage

  • Group files with the same name and size; keep one, and record where the others were.
  • Drop system files, temporary exports and screenshots taken for a one-off check.
  • Export native Google documents to a text format an agent can read.
  • Fill the shelves in priority order, starting with the documents people ask for most.

Structure

  • Agree one spelling for every client and project name.
  • Rename files to one pattern that ends in an ISO date.
  • Create the five shelves: Projects, Clients, Areas, Resources, Archive.
  • Move every kept file to exactly one shelf.

Facts and work

  • Write one canonical facts file per client, with a source and a date for each fact.
  • Start a task ledger: owner, due date, where the commitment came from, what counts as done.
  • Decide which information is a table (tasks, assets, client fields) and which is text.

Access and first test

  • Connect the agent with a read-only Drive scope, from an account you can revoke.
  • Put an instruction file at the root explaining the shelves and the rules.
  • Ask five questions you already know the answers to, and check each answer against its file.

When you do not need this

If your Drive holds a few hundred files, one person owns all of them and you can find any document in seconds, you do not need a knowledge base yet, and you certainly do not need to pay anyone for one. A clear naming pattern and an archive folder will carry you until a second person, a second client or an agent has to rely on what you know.