Create one project per client, switch it to project-only memory, and put that client's brief, prices and key correspondence into the project's files and instructions. Project-only memory stops chats in the project drawing on saved memories or chats from anywhere else, and chats elsewhere can't see into it. Shared projects always work this way.
Separation here means separation of what ChatGPT draws on, not of who can look. Anyone signed in to the account can open every project, and nothing stops you uploading one client's file into another client's project. The memory choice is also made per project: one set up in a hurry on default memory can still pull in details from other chats, and OpenAI says a later switch to project-only can take a few hours to take effect.
What project-only memory walls off, and what it doesn't
OpenAI's help pages describe project-only memory in three rules: your saved memories aren't used in the project's chats; chats in the project can refer to each other but not to conversations outside it; and chats outside the project can't refer to what happened inside it. That is exactly the wall you want between a hotel's refurbishment and a gallery's exhibition. It leaves four gaps open:
- Account access. Everyone with the login sees every project in it. A shared login defeats the whole arrangement.
- Connected apps. If you use an app such as Google Drive inside a project chat, its reach is still your whole account, not the client's folder. The limits of those apps are set out in what ChatGPT's apps can see in your Drive and inbox.
- Your own copying. Paste a paragraph from one client's chat into another and the wall means nothing.
- Library files. Files saved separately to your Library, which holds 20 GB on Plus and Business, live outside any project and survive when a project is deleted.
Project instructions also matter more than they look. They apply only inside that project and override your account-wide custom instructions, so a client's tone and sign-off can differ from your usual ones without any risk of the two blending.
Setting up a client project, step by step
Allow about 20 minutes per client for the first setup. Most of it is choosing which files to include.
- Name it so nobody can mix it up. Client type plus job, such as "Hotel group – guest-room art 2026". Two clients with similar names is how files end up in the wrong project.
- Choose project-only memory. Pick it when you create the project, or open the project's settings, choose Project-only memory under Memory, and save. Give it a few hours before relying on it.
- Write the instructions. Who the client is, who approves what, tone, which file holds the prices, and what ChatGPT must never do. A filled-in example follows below.
- Add the files that answer most questions. OpenAI's limits are 5 files per project on Free, 25 on Go and Plus, and 40 on Pro, Business, Enterprise and Edu, with up to 10 uploaded at once. Combine small documents where you can.
- Move the old chats in. Drag an existing chat onto the project, or choose Move to project from the chat's menu. A moved chat picks up the project's instructions and files. Chats started with a custom GPT can't be moved, so copy anything you need from those by hand.
- Test the wall. In the new project, ask something only another client's chats would know, such as "What glazing did we use for the gallery?" A properly walled project should say it doesn't have that information.
If you've been using a custom GPT per client, this is the moment to move. OpenAI is retiring custom GPTs, which stop running on 11 December 2026, and a project's instructions and files do the same job for client context. The trade-offs between ChatGPT projects, Claude projects and Gemini Gems (which become skills from November 2026) are compared in custom GPT vs Claude project vs Gemini Gem.
A picture framer's three commercial clients, split into projects
Consider an illustrative framing business on ChatGPT Plus that does retail work plus three commercial accounts: a hotel group refitting 120 guest rooms, a gallery that sends 8 to 15 pieces a month for exhibition framing, and a dental practice redecorating its waiting room. Before projects, every chat sat in one history. The failure that prompted the change: a delivery email for the hotel described "conservation-grade museum glass", which is the gallery's standard and would have added about $40 a frame to a 120-room job if the hotel had taken it at its word.
The setup, client by client:
| Project | Files included | Count (Plus limit 25) | Notes |
|---|---|---|---|
| Hotel group – guest-room art | Design brief; 60 artwork specs merged into one PDF; room schedule spreadsheet; price agreement v3; delivery plan; three key email threads saved as PDF | 8 | The 60 separate specs would have broken the limit on their own |
| Gallery – exhibition framing | Conservation standards sheet; gallery price list; 12 recent job sheets | 14 | Job sheets rotate: oldest out when a new one goes in |
| Dental practice – waiting room | Brief; quote; photos of the wall; approval email | 4 | Deleted when the job is invoiced |
The sums behind it: three projects at about 20 minutes each took an hour to set up. Before, the owner spent roughly 4 minutes at the start of each client chat pasting background, across about 30 client chats a month, so around two hours a month went on re-explaining. That time mostly disappears, and the museum-glass mistake can't recur, because the hotel project never sees the gallery's chats. Retail customers stayed out of projects entirely; one-off questions about them go in ordinary chats.
The instructions for the hotel project
Client: hotel group, 120-room refurbishment (rooms 101 to 460).
Our job: framing and installing guest-room and corridor artwork.
Contacts: [first name], interiors lead, approves designs.
[first name], operations, arranges access. Address emails to the
interiors lead and copy operations.
Tone: short and practical, no sales language, British English.
Facts: prices come only from "Price agreement v3". Moulding is the
oak 25 mm profile unless a spec says otherwise. Glazing is
anti-reflective acrylic in all guest rooms.
Never: quote a price from any other file, mention other clients,
or promise a delivery date that isn't in the newest delivery plan.
If a question needs a file you don't have, name the file and stop.A first test prompt and an illustrative reply:
Draft the email confirming next week's delivery for floor 3.Hi [first name], we'll deliver 34 framed pieces for rooms 301 to 317 on Tuesday 13 October between 8 and 10am, with installation the same day. Could [first name] confirm lift access for that morning? Glazing is anti-reflective acrylic throughout, as agreed.
What was wrong: the newest plan said Wednesday 14 October. An older delivery plan was still in the project, and ChatGPT used it. Two fixes made it stick: delete superseded files rather than letting versions pile up, and put the date in every file name ("Delivery plan 2026-10-06"). A standing line in the instructions, "use the most recently dated delivery plan and say which file you used", now catches the rest.
The gallery project, and a different kind of error
The gallery project failed in another way, which is worth knowing because it looks like competence. The prompt:
From this month's job sheets, list every piece that needs UV-filtering
glazing and every piece over 100 cm on the long edge.Seven pieces need UV-filtering glazing: GAL-231 to GAL-235, GAL-238 and GAL-240. Three are over 100 cm on the long edge: GAL-233 (120 cm), GAL-236 (104 cm) and GAL-241 (110 cm).
GAL-241 didn't exist. Its job sheet was a phone photo of a handwritten form, and ChatGPT had read a smudged "GAL-247, 70 cm" as something else and filled the gap with a plausible answer. It showed up when the framer went to cut the moulding and couldn't find the piece. The fix was one more line in the project instructions, "if any job sheet is hard to read, list it as unreadable instead of guessing", plus a rule that job sheets go in as typed forms or clear scans. Since then, the answer ends with "unreadable: 1 sheet" whenever a photo is poor, which takes seconds to deal with.
Keep shared house material in its own project
Some material belongs to every client: your standard terms, moulding catalogue, email sign-offs, installation checklist. The temptation is to copy it from one client's chat into the next, which is exactly the habit that breaks the wall. A cleaner pattern is an internal project, "House material", holding those documents and nothing client-specific. When a client project needs the standard terms, upload the same file into that project too. It costs a slot against the file limit, but no client's detail ever travels through the house project, and when the terms change you know which projects need the new version.
Which ChatGPT plan gives a client project what
Project limits and defaults change with the plan. As OpenAI's help pages put them in September 2026:
| Plan | Files per project | Project sharing | Training on your content |
|---|---|---|---|
| Free | 5 | Yes, by invitation | On unless you switch it off in data controls |
| Go and Plus | 25 | Yes, by invitation | On unless you switch it off |
| Pro | 40 | Yes, by invitation | On unless you switch it off |
| Business | 40 | Yes, plus groups and a workspace link | Off by default |
| Enterprise and Edu | 40 | Yes, plus workspace controls | Off by default |
For a sole trader with a handful of clients, Plus with the training switch turned off is usually enough: 25 files covers most clients once small documents are merged. The case for Business over Plus is the workspace side: sharing by group or workspace link, admin control over who's in the workspace, the 40-file limit, and business data excluded from training by default. The seat maths is in the next section.
Sharing a client project with staff on ChatGPT Business
OpenAI's help pages say project sharing now works on every plan, Free included. What Business adds is the workspace around it: you share with colleagues by name, email address or workspace group, choosing chat access (they can use the project) or edit access (they can change its instructions and files), and a workspace link lets anyone in the workspace join with chat access until you upgrade them. A workspace project can have up to 100 collaborators. Whatever the plan, a shared project runs on project-only memory, and OpenAI doesn't let you switch it back to default memory.
An illustrative garden centre with a small landscape design service shows how that fits. The designer owns one project per design client, with edit access; two assistants get chat access to the projects they support, so they can draft planting lists and client updates but can't swap the price file. Business needs at least two seats, so three people cost $75 a month at $25 a seat billed monthly, or $60 a month on the annual rate of $20 a seat. The same setup in Claude is covered in setting up Claude Projects as a shared team assistant, if your team prefers Claude.
Two rules keep sharing tidy. Give edit access to one person per project, so there's a single owner of its instructions and files. And review the sharing list whenever someone changes role, because a colleague who moved to another department keeps access to every project they were added to until you remove them.
How a wrong file slips past the wall
Project-only memory can't protect you from your own uploads. An illustrative music teacher kept one project per exam student. In a busy week before reports were due, she uploaded one student's progress notes into another student's project. The end-of-term summary for the second student then mentioned a Grade 5 pass that belonged to the first. It surfaced when a parent replied asking when their child had taken Grade 5.
The damage was small, one email and an apology, but the fix is worth copying. File names now start with the student's initials. Every report prompt ends with "list the files you used", and the teacher checks that list against the project name before sending. It takes ten seconds per report and would have caught this one. For work you'd rather not put in a project at all, anonymising client data before it goes into AI is the stronger habit.
Closing a client project cleanly
Client work ends, and so should its project. OpenAI's help pages say deleting a project permanently removes its chats, instructions and any files stored only in that project, while files saved separately in your Library stay put. Deleted items are removed from OpenAI's systems within 30 days unless the company must keep them for legal or security reasons. That makes deletion a real decision, so run a short close-down first. Filled in for the framer's dental job:
| Close-down step | Dental practice job |
|---|---|
| Copy out anything you must keep | Final quote and approval email saved to the client folder on the shop's drive |
| Check the Library for this client's files | Two wall photos found in the Library and deleted |
| Remove collaborators | Not shared, nothing to remove |
| Check your contract's retention terms | No requirement beyond keeping the invoice |
| Delete the project | Done the day the invoice was paid |
| Note the date | Recorded in the job sheet: "AI project deleted" |
The Library check is the step most people skip, and it's the one that leaves client material behind. For long-running clients, don't delete: prune instead, removing files that are superseded every month or two so the project answers from current documents. Reusing old client work for a new client is its own risk, covered in reusing past client work with AI without leaking client data.
Quick signs a project is working
After a month, three checks tell you whether the split is doing its job. Ask each project a question only another client could answer; it should say it doesn't know. Read the last five outputs and look for any detail from another client, such as a price, product or name. And count how often you still paste background at the start of a chat: if it's more than once a week for a client, that background belongs in the project's instructions or files.
ChatGPT Projects for client work: common questions
Can a client see or join my ChatGPT project?
Only if you invite them. Projects can be shared on every ChatGPT plan, so adding a client is technically possible, but they would then see every file and chat in it, including internal notes and pricing workings. Keep projects as your own working space and send clients finished documents. If a client needs to collaborate, use a shared document or folder you control instead.
Does project-only memory stop OpenAI training on my client files?
No, that's a separate setting. Project-only memory controls which chats and memories ChatGPT draws on inside the project. Training is governed by your plan and the model-training switch in data controls: business plans don't train on your content by default, while individual plans need the switch turned off. Check both before putting client material in any project.
Is one project per job better than one per client?
Use one per client when jobs share context such as contacts, prices and tone, and one per job when a job is large, has its own files, or will be deleted on a different date. A picture framer with a hotel group might keep one project for the whole refurbishment, but a separate one for a later, unrelated tender with that group.
What happens to a project when a staff member leaves?
On a business plan, check which shared projects they owned or edited, give edit access to someone who stays, and remove the leaver from the sharing list. Projects they kept private in their own seat should be reviewed before the seat is removed, because deleting a project permanently deletes its chats and project-only files.
Further reads
- ChatGPT Temporary Chat: When Your Staff Should Use It — For quick one-off questions that shouldn't sit in any project.
- ChatGPT Plus vs ChatGPT Business: Which Plan Does a Team Need? — Which plan you need for shared projects and 40-file limits.
- When Does a Shared ChatGPT Plan Stop Being Enough for Your Team? — Signs one shared account can no longer keep clients apart.
- Is It Safe to Put Customer Data Into ChatGPT? — What client data belongs in ChatGPT at all.
- How to Mention AI Use in Client Contracts and Proposals — How to tell clients you use AI on their work.
- How to Set Up Company AI Accounts Instead of Personal Logins — Move client work off personal accounts before it grows.
- How to Give AI Your Business Context So Answers Fit — Build a one-page business profile and task instructions, store them where each AI tool reads them, and test that answers fit your business.
- Does ChatGPT Remember Conversations? Memory Settings for Business — The three ways ChatGPT remembers, where the memory switches are in September 2026, and how to stop one client's details turning up in another's work.
- How to Write Event Proposals With AI in Half the Time — Where the hours in an event proposal really go, the kit to build once, four prompts to run in order, and a timed worked example.
- ChatGPT Prompts for Bookkeepers: Client Queries, Chasers, Notes — Twelve copy-ready prompts for client replies, record chasers and file notes, each shown with an illustrative output and the fix it needs.
- How Recruiters Use AI to Write Candidate Summaries Clients Read — A summary format hiring managers actually read, the prompt that fills it from your screening notes, a sample output with fixes, and the five-minute claim check.
- How HR Consultants Use AI for Policies, Letters and Cases — Three workflows for HR consultants, with a client fact sheet, letter before-and-after, a grievance chronology prompt and the lines AI must never cross.
- How Estate Agents Use AI to Prepare Valuation Appointments — A two-day prep routine for market appraisals: AI builds the property brief, comparables commentary and vendor pack, and the valuer sets the price.
- Writing Pharmacy SOPs With AI: A Step-by-Step Workflow — Record how a task is really done, feed AI your regulator's wording, then gap-check, bench-test and sign off. A pharmacy SOP in about two hours.
- AI Employee Onboarding: Plan a New Starter's First Two Weeks — Use AI to draft a realistic day-by-day plan for a new starter's first fortnight, a reading path through your documents, and an assistant they can ask.
- How Much Does a Chat-With-Your-Documents System Cost? — Build a document-chat budget that includes file preparation, access checks and the work needed to keep answers current.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI help pages (Projects in ChatGPT; Memory in ChatGPT; File uploads FAQ; Using Library to manage files; Managing data, sharing, and privacy in ChatGPT Business; Custom GPT retirement and migration FAQ), checked September 2026.