A business owner I work with asked her AI agent a simple question: "Can you get me our current logo?"
It returned the wrong file.
Her first reaction was to blame the agent. Then we opened the folder. Months earlier, someone had uploaded the new logo and never renamed anything, so an older file was still sitting there called Current Logo. The agent read the name, believed it, and handed her a three-year-old logo. It did exactly what her Drive told it to do.
Here is the part that matters. Every person on her team knew which logo was real. Humans compensate for messy systems. They remember that Proposal FINAL FINAL v4 is the newest proposal, that the signed contract lives in an email attachment, that New Folder 7 somehow contains the onboarding documents. AI reads the structure you actually built, and it inherits every shortcut you took.
That makes this the right moment to clean up your Drive, because your Drive stopped being storage a while ago. It is your business's memory, and increasingly it is the memory your AI tools will work from. These ten prompts, run in ChatGPT Work with your Drive connected, audit it the way a new employee would experience it: no context, no folklore, just what the files actually say.
Before You Start: Three Rules
Run these in ChatGPT Work. ChatGPT Work is the agent mode inside the new ChatGPT desktop app, the July 9, 2026 update that merged the Codex app in. Connect the Google Drive account that actually holds your business files (Settings, Connectors, Google Drive), open a Work session, and paste the prompts in order. Work mode carries findings from one prompt to the next, which is exactly what the final roadmap prompt needs. Two precision notes: the app is just called ChatGPT, Work is the mode inside it, and Work usage is metered, so a deep audit spends credits. If you want to conserve them, every prompt here also runs in a regular chat with the Drive connector. Our ChatGPT Work tutorial covers the setup in ten minutes.
Expect it to search, and expect limits. ChatGPT Work reads your Drive through search. It can read what your documents say, but it usually cannot see sharing settings, file owners, or view history. These prompts are written for that reality: each one asks the model to say plainly what it can and cannot see, and to separate what it confirmed from what it guessed. When it tells you "I can't see permissions from here," that is the prompt working, and the prompt tells you what to check manually.
Nothing changes without you. Every prompt here is written read-only: audit, cite, propose, stop. ChatGPT Work can take actions through your connectors when you approve them, which is exactly why the habit matters. If it offers to fix something as it goes, decline and keep it in audit mode until prompt 10 gives you the full plan. Audit, review, approve, then execute. In that order, always. When you eventually hand this work to an agent on a schedule, that habit is the thing protecting your business records.
One more thing: the first run will expose a mess. Duplicate proposals, a contract in the wrong client's folder, a price list from two years ago that looks current. Do not treat that as failure. The mess was already there, quietly making every project slower. Now you can see it.
Run the prompts in order. The first two show you the truth, the middle six find the risk, and the last two turn it into a plan.
The Prompts
1. Test how findable your business actually is
This is the logo story, turned into a test you can run in five minutes. Most owners skip straight to organizing. Run this first, because it shows you exactly what a stranger, or an agent, cannot find.
I have connected my Google Drive to this conversation. Before we organize anything, I want to test how findable my business actually is. Here are real questions about my business. Answer each one using only what you can find in my Drive: 1. What is our current logo file? 2. What is our current pricing? 3. Where is the signed agreement for our newest client? 4. What do we send a new client on day one? 5. [ADD 5-10 MORE: the questions your team actually asks each other] For every answer: name the exact file you used, and rate your confidence high, medium, or low. If two files disagree, show me both and say which one you would bet on and why. At the end, list every question where you had to guess, and what about my Drive forced the guess. Do not change anything. This is a test, not a cleanup.
Every "low confidence" and every guess on that list is a place your business currently runs on somebody's memory.
2. Map what the AI can actually see
Before fixing anything, get an honest inventory, including an honest account of the blind spots.
Working in my connected Google Drive, build me a map of what you can actually see, and be honest about what you cannot. First: tell me plainly which information is available to you here (file contents, names, locations) and which is not (sharing settings, file owners, view history, anything else you cannot access). Do not fake precision you do not have. Then, searching as broadly as you can, describe the major areas of my Drive: - what business function each area seems to serve (sales, delivery, finance, operations, marketing) - what lives in each area, in plain English - areas that overlap or appear to duplicate each other - the areas that would hurt the business most if they were lost or leaked Cite the folders and files behind every observation, and separate what you confirmed from what you are inferring. No recommendations yet. This is inventory only.
3. Untangle the duplicates and the Final FINALs
Duplicate files rarely waste storage. They waste decisions, because different people edit different copies, and eventually somebody sends a client the old pricing.
Search my connected Google Drive for documents that exist in more than one version: exact copies, near copies, and the "Final, Final 2, FINAL FINAL v3" families. Start with the documents that matter most: proposals, contracts, price lists, brand assets, onboarding docs. [NAME ANY DOCUMENT YOU ALREADY SUSPECT HAS COPIES.] For each set you find: 1. List every version and where it lives 2. Recommend which one should be the single source of truth, using evidence from the content itself: which is most complete, which do other documents reference, which matches what the business actually uses 3. Give me your confidence, and the business risk if the wrong copy stays in circulation Where you cannot tell which version is current, say so and flag it for me instead of picking one. Do not merge, rename, or delete anything. This is a review list.
4. Fix the names that force guessing
A file named Current Logo might have been current three years ago. Humans read names with context. AI reads names as claims.
Review the file and folder names in my connected Google Drive and flag every name that would force someone who has never seen my business to guess. I mean names like: - "Current," "New," "Latest," or "Final" with no date or version - "Document," "Untitled," "Copy of Copy of," "New Folder" - client or project files with no client, date, or version in the name Propose one simple naming standard for my business, short enough that people will actually follow it. Then give me a rename preview as a table: current name, proposed name, reason. Mark any rename you are not fully confident about, and tell me what could go wrong if that file were renamed without a human checking it first. Rename nothing. This is a preview for my approval.
5. Check every client folder for gaps
Missing documents stay invisible until the worst moment: a dispute, a renewal, a client asking for the deliverable you cannot find.
Using my connected Google Drive, check my client records for gaps before a client finds them for me. My active clients and projects are: [LIST THEM, or tell me what you can infer from my folders]. For each one, search for the documents a complete engagement should have: - a signed agreement or contract - the proposal or scope - onboarding materials - key deliverables - invoices or payment records Give me a per-client checklist of found and not found, citing the exact file for every "found." For every "not found," say whether it might simply live under a different name or location, and list what you searched before concluding it is missing. Separate confirmed gaps from suspected gaps. Change nothing. This is a report.
6. Find the files living in the wrong place
Most "lost" files were never lost. They were filed somewhere that made sense to one person, once.
Search my connected Google Drive for documents whose content says they belong somewhere else: a signed contract sitting in a general uploads folder, a deliverable filed under the wrong client, a financial record in a marketing folder. For each one, give me: where it lives now, where its content suggests it belongs, the evidence, and your confidence. Anything ambiguous, flag for me to decide instead of deciding yourself. Present moves as a proposal only. Do not move anything.
7. Find the sensitive material sitting in the open
ChatGPT Work usually cannot see your sharing settings, so this prompt does what it can do well (find sensitive content in careless places) and then hands you a checklist for what only you can verify.
Search the contents of my connected Google Drive for sensitive material sitting where it should not be: contracts, pricing, payroll or financial records, personal information, and especially passwords or login details saved in plain documents. For each finding: the file, where it lives, why it is sensitive, and how bad it would be if the wrong person opened it. Rank the list by risk. Do not quote sensitive values back to me in full. Then be honest about your limit: you likely cannot see my sharing settings from here. Based on the sensitive files you found, give me a short manual checklist of exactly what to verify myself in Drive's own "Shared" and "Shared with me" views, including which specific files to check first. Change nothing.
8. Surface the abandoned corners
Abandoned folders cut both ways. Some are junk. Some hold years of client history from an employee who left, and nobody ever went looking.
Search my connected Google Drive for areas that look abandoned: folders whose contents reference projects, clients, years, or employees that are clearly in the past, half-finished reorganizations (two folder trees for the same purpose), and important-looking documents stranded outside the main working areas. You may not be able to see activity dates, so judge by content: what the documents talk about, who they mention, which years they reference. Sort what you find into three piles: 1. Archive candidates 2. Recovery candidates: valuable material that should be pulled forward, like old client work, templates, or forgotten intellectual property 3. Needs a human to decide Cite the evidence behind each call. Recommend review, not deletion.
9. Design the structure a stranger could navigate
Only now, after eight prompts of evidence, does redesign make sense. Structure first is how businesses end up reorganizing the same Drive every eighteen months.
Using everything you have learned about my Drive in this conversation, propose a simpler folder structure built around how my business actually operates, not around anyone's personal filing habits. Give me: 1. A proposed top-level structure organized by business function 2. The naming standard from earlier, finalized 3. A map of where today's content would go in the new structure 4. The test: could a brand-new team member, or an AI agent seeing this Drive for the first time, answer every question from our first prompt without asking anyone? Present it as a proposal for my review. Create and move nothing.
10. Turn it all into an approval plan
Cleanup should happen in small batches you approve one at a time, never as one heroic weekend reorganization.
Turn every recommendation from this conversation into a staged cleanup plan I can approve batch by batch. Order the batches from lowest risk and highest value down to highest risk. For each batch: 1. What changes 2. The specific files and folders affected 3. The business benefit 4. The risk if it goes wrong 5. What I need to check before saying yes Nothing gets renamed, moved, archived, or deleted until I approve that batch, and the actual changes will be made by me or my team in Drive itself. Your job is the plan, the evidence, and the order. Give me the full roadmap first.
What You Will Find
Nobody runs this audit and finds nothing. You will find two price lists that disagree, a contract filed under the wrong client, and at least one folder from someone who left in 2023. Expect prompt 1 to sting and prompt 5 to pay for the afternoon by itself, because a missing signed agreement is a problem you want to discover now, on your own terms.
Then rerun prompt 1 after the cleanup. Watching every guess turn into a confident, cited answer is the whole point, and it is the same test any AI agent will take on day one.
Common Questions
Is it safe to connect my Google Drive to ChatGPT Work? Use a paid plan and turn off model training in settings (Settings, Data Controls, turn off "Improve the model for everyone"), or use a Business or Enterprise plan where that is off by default. Every prompt in this audit is written read-only. ChatGPT Work can take connector actions when you approve them, so treat any offer to fix, rename, or move files as an approval moment, and hold changes until you have the full plan from prompt 10.
Do I need a paid plan to use ChatGPT Work? The new ChatGPT app puts Chat, Work, and Codex modes on every plan, including Free, but Work usage is metered against a shared credit bucket. A full first audit fits comfortably on a paid plan; on Free you may run out of credits mid-audit. The economical route: run the audit in a regular chat with the Drive connector, and save Work credits for the day you hand this checklist to it as one job.
Why can't ChatGPT Work see who owns a file or who it's shared with? The Drive connector reads file contents and names through search. Sharing settings, owners, and view history mostly are not exposed to it. That is why prompt 7 pairs the AI's content scan with a manual checklist for Drive's own sharing views, and why every prompt asks the model to say what it could not see instead of guessing.
How often should I run this? The full ten, once. After the cleanup, prompt 1 (the findability test) quarterly, and prompt 3 (duplicates and versions) whenever a document that matters gets revised. If the quarterly test starts failing again, the naming standard from prompt 4 is what slipped.
Bottom Line
These ten prompts are you auditing the Drive by hand, once. An AI agent runs the same checks on a schedule, catches the second Current Logo before anyone uses it, and hands you the exceptions list every Monday without being asked. If you want to know how ready your business is for that, start with the free Agent-Ready Score.