Before inviting staff to ChatGPT Business, an owner should check workspace discovery and roles, paid seat capacity, plugins and connected apps, sharing, memory, data controls and credit spending. Record each setting and test it with a non-sensitive sample account. The important distinction is between a control you can actually set in Business and a policy your team must enforce.
Do this while the workspace has only its initial owner accounts. A member can invite another member when a paid seat is available, workspace discovery is on by default and connected apps start enabled. Business also cannot disable shared links for everyone, according to OpenAI's current help. Those details change what a sensible pre-invite checklist looks like.
Record the starting state before changing controls
Open the intended Business workspace from the profile switcher, then open Workspace settings. Write down the workspace name, account used as owner, billing cycle, number and type of paid seats, and the renewal date. Two workspaces with similar names can exist, so confirm the correct one before making changes. The owner should keep a short decision log: setting or policy, current state, chosen state, reason, person responsible and date tested. A screenshot alone is weak evidence because it does not say why a choice was made or whether staff were told about it.
For an illustrative removals firm, a starting record might say: two Standard seats paid; one occupied by the owner, one available; monthly renewal on the 18th; discovery on; automatic acceptance off; no apps connected; credits auto top-up off. That is a useful baseline because a later third invitation can be traced to a seat purchase rather than confused with ordinary member reassignment. Do not put customer names, addresses or inventory photographs in the test record.
Assign a second trusted owner if the business truly has one. Owner is a powerful role: OpenAI says owners control billing, identity management and workspace configuration. Admins can help with routine membership, while analytics viewers have a narrower information role. Give working staff Member unless their job requires administration. A shared owner login is a poor backup because it erases accountability; a separately named owner account is better. If a second owner is unavailable, document how the business will regain access when the first owner is absent.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowClose the unintended joining path
In Workspace settings, inspect Identity & access and the workspace discovery choice. OpenAI says Business discovery starts on, and owners may allow people with a verified company domain to request access. Automatic acceptance is a separate choice. For a small team, leave automatic acceptance off until the owner has an approved list of people and seats. Turning discovery off may be appropriate if invitations should be the only entry path. Then send a test join request from an approved account and check that the owner, rather than the applicant, controls approval.
There is a limit to this lock-down. OpenAI currently says members can invite other members using the default seat type, and that you cannot prevent members from doing so where capacity is available. Therefore write a staff rule: invitations require the owner's approval before they are sent. Check pending invitations weekly and investigate any name you do not recognise. The technical gate is the availability of a matching paid seat, not a complete prevention of invitations.
A landscaping company with a generic office email domain could otherwise mistake a join request for a normal hire. Its sample approval record would say: requester, employment status, manager approval, seat type and requested access date. If the person has only a short project, set a review date. When the project ends, remove the member and separately reduce the billable seat if there will be no replacement. The renewal seat audit explains that second step in detail.
Choose roles and seat defaults deliberately
In Members, inspect each person's role and seat type. Standard and Premium are paid capacity, while Member, Admin and Owner describe permissions. These are separate questions. A trusted office manager might be an Admin on a Standard seat; a heavy user might have a Premium seat but still be a Member. Do not award Owner just to give someone more usage. OpenAI lists Analytics Viewer as another built-in role, suitable only when someone needs the available workspace information without routine administration.
OpenAI's current guide puts the default seat type under Workspace settings > Billing > Seat preferences. It says the default is Standard unless the workspace has only Premium seats, and an invitation's selected seat type overrides it. Check the displayed default, make it Standard for ordinary staff if appropriate, and review the type on each invitation before sending. A Premium selection can change the next invoice materially even though the recipient sees an ordinary invitation email.
Use an illustrative five-person tutoring centre to test the rule. The owner and four tutors all need ordinary drafting and document review. A proposed mix of four Standard seats and one Premium seat costs $225 a month on monthly billing, using OpenAI's current USD list prices: four times $25 plus $125. Five Standard seats cost $125. The $100 difference is justified only if one person's measured use needs Premium limits. The owner can start with Standard, record a specific usage problem, then change a seat type after reviewing the charge. Prices are list figures before any regional checkout variation or tax.
Review plugins and connected apps before connection
This is the default most owners miss: OpenAI's admin guide says apps (formerly connectors) are enabled by default in ChatGPT Business, whereas new apps in Enterprise and Edu start disabled. Open Admin Console, select the workspace and open Plugins. Each plugin has an installation policy (Available for installation, Installed for workspace or Disable), and each included app can show Actions for reads and writes plus a New actions rule; Disable new actions blocks only actions added later, not ones already switched on. Google Drive's combined Docs, Sheets and Slides actions are also on by default for Business. Business administrators decide whether an app is enabled for the whole workspace; role-by-role app access is an Enterprise and Edu control. OpenAI also says changing an app permission does not disconnect the app or revoke access already granted at the provider, so a disconnection at the source service may be needed too.
Next, set the workspace-wide approval default under Workspace settings > General > Settings > Plugin permissions. The options OpenAI lists for that selector are Always ask, Allow read actions and Allow low-risk actions; Allow all actions is not offered workspace-wide, though it can appear for an individual app. Before staff are briefed, Always ask is the cautious choice. Where an app offers Manage domains, add the company email domain so a member cannot connect a personal mailbox or drive to the Business workspace; OpenAI says an empty list allows accounts from every domain.
Start with no external account connected, then approve one narrow use case at a time. For each app, record the staff role, source account, source permissions, data it may read, actions it may take and who reviews an action before it is final. Read access and action access are different risks. A bookshop might want an assistant to summarise an inventory sheet, but not to edit product prices or send supplier orders. Test with a copy of the sheet containing invented stock-keeping units and verify what the app can actually see.
OpenAI's app-permission guide says some actions may need additional review, including sending a message, creating or editing a file or record, changing access or sharing settings and making a purchase. Because Business switches apps on or off for the whole workspace, do not promise staff that a connected app can be permitted for only one named member. A sensible fallback is to leave the app disabled, use a narrower source account, or keep the job manual.
A sample before-and-after illustrates the difference. Before: the general office mailbox has access to all customer enquiries and is connected to an assistant for reply drafting. After: a dedicated test mailbox receives a small set of copied, non-sensitive enquiries; staff draft but do not send; an owner approves expansion only after checking the app's scopes and the quality of replies. If the connection is later withdrawn, disconnect it inside ChatGPT and at the mail provider, then verify a fresh request fails. The mail access permissions tutorial covers that boundary more closely.
Set sharing rules around what Business can enforce
Business shared conversation links are restricted to authenticated members of the originating workspace, according to OpenAI. That is useful, but a shared chat may include uploaded files or images. A recipient can also continue a shared conversation as a new private chat. Ask staff to inspect the entire thread before sharing, including any earlier material that a link may expose. Restrict the use of real client details until the business has approved a data classification rule.
OpenAI says Business does not have an Enterprise-wide switch to disable shared links. Staff can manage links under Settings > Data controls > Shared links and delete one they created. Treat a policy as a policy, not as a locked setting: only share within a named project or with an approved colleague; check attachments first; revoke a link if the audience changes. A shared link is not a substitute for a document access review. It also does not make all the creator's private chats visible to an owner.
Illustrative incident: a tutor shares a lesson-planning chat that began with a pupil's full report, even though the final message is a harmless worksheet. The recipient sees more of the thread than the sender expected. The corrective step is to start a fresh conversation using a redacted extract, share that narrower conversation and remove the old link. The owner should practise this with fake data before staff are invited. A real-task trial plan can include exactly this sharing rehearsal.
Check memory, personal context and data controls
Memory can make repeated work faster, but it can also carry an old instruction into a new task. OpenAI's Business memory FAQ says workspace owners can control whether memory and personalisation are available across the whole workspace, and each user can also turn Memory off in their own settings. Decide before staff join, because the same FAQ says that if an owner later turns Memory off for the workspace, members' existing saved memories are deleted. Memories belong to each individual account and do not transfer to colleagues. If memory stays on, test it with a harmless preference such as a house style for customer replies. Ask the assistant what it remembers, correct a deliberately outdated preference and verify the next answer reflects the correction.
Business workspace prompts and responses are excluded from OpenAI model training by default, according to OpenAI's current Business privacy guide. That is a vendor commitment, not a permission to paste any customer record. Record which data classes are approved for use, whether anonymisation is required, and who can approve an exception. Check how staff select the Business workspace from the profile menu so that a person with both a personal and a Business workspace does not put work content in the wrong one.
Test this with an illustrative draft. The raw note says: “Reply to [first name] about their order at 24 Example Road; include their phone number and the reason for the missed delivery.” The safer input for a style test says: “Draft a courteous missed-delivery reply using placeholders for name, address and contact details; do not invent a new delivery date.” Compare the output with the actual order system before sending. The difference is specific: the model needs the policy and tone, not every identifier.
Retention and export need their own check. OpenAI's Business privacy page says members cannot export their own Business data through Data controls; it describes an owner export option when available. The general FAQ still says self-service export is unavailable. Treat those pages as a reason to verify the actual workspace and support route, not to promise a guaranteed complete export. Before inviting staff, decide where final business documents will live outside chat, who owns them, and what must be copied into the normal records system. This matters when a member later leaves.
Cap credit spending without mistaking it for seat spending
Open Workspace settings > Billing and inspect credits, automatic reload and the monthly credit limits. Standard and Premium seat subscriptions are recurring commitments; each seat uses its included allowance first, and workspace credits then cover eligible extra usage where flexible pricing and spend controls allow it. OpenAI's credits guide says no seat or user has a credit limit by default, so set a monthly limit per seat type, with per-user overrides where needed, before staff join. Owners and admins can set those limits; only owners can buy credits or turn on automatic reload. If reload is switched on, fill in the Monthly recharge limit: OpenAI says leaving it blank allows unlimited automatic purchases each month. Purchased credits are valid for 12 months. Only owners can manage paid seats and billing information, according to OpenAI's Business billing guide, so record who reviews the monthly invoice.
For example, a removals firm budgets $75 a month for three Standard seats on monthly billing, plus a separately approved credit allowance of $20. The subscription is still $75 if credits are never used. If the owner approves a fourth paid seat for a new starter, the recurring seat total becomes $100, with a prorated charge for the current cycle. If credits are bought, they add a separate line of cost. A single “$95 AI budget” label would conceal those two mechanisms and could lead the owner to miss a recurring seat purchase.
Check the next invoice date and seat count after your test invitation. A pending invitation does not itself reserve a seat; a successful join needs suitable capacity. If there is already an unused paid seat, assigning it does not buy another. The two-seat minimum remains usual, although OpenAI now documents a limited one-seat checkout for some new workspaces. The one-person cost comparison explains that exception and the decision it changes.
Run a complete rehearsal with fake data
Prepare a rehearsal account belonging to a real authorised person, then follow the same path a new staff member will follow. Invite it as Member on the chosen seat type, verify the email and workspace name, select the Business workspace, and ask it to perform one permitted task. Check that it can reach approved features and cannot complete a disallowed external action. Then try a shared link, a memory-dependent request and a connected-app request. Record what happened rather than what the settings screen appeared to promise.
A compact filled-in test record might read: “Test account: office manager; role: Member; seat: Standard; discovery join: owner approval required; inventory app: disabled; chat link: visible to signed-in member only; customer-data test: placeholders used; credit top-up: off; result: pass, except sharing policy needs staff briefing.” That exception is important. It distinguishes a configuration defect from a behaviour staff must learn. The briefing should tell staff which workspace to use, what data to omit, when to request an app, and whom to ask before inviting a colleague.
Finally, remove the rehearsal member if it was temporary. Verify access ends, check whether its paid seat is still in the subscription, and decide whether to reassign or schedule a reduction at the next billing cycle. Do not merge a personal workspace into Business merely to test onboarding: OpenAI says that merge has no undo path back to personal. A regional checkout comparison is useful if the invoice currency or tax treatment also needs review before real staff join.
Give the first joiner an instruction sheet
Settings are easier to keep when staff know the reason for them. Send a one-page instruction sheet with the invitation, then ask the first joiner to acknowledge it. It should name the correct workspace, approved task types, forbidden data, the app-request process, sharing rules, who may approve a new member and where final documents are stored. Avoid a vague rule such as “use AI responsibly”; give an example that can be checked. For a bookshop, the approved task might be drafting a supplier chase from a redacted purchase-order summary, while putting a customer's payment details into a chat is outside scope.
Use a short escalation route: staff member stops, saves the task in the ordinary work system, and asks the named owner when a prompt needs customer records, a new app or an external action. If the owner approves a new use, update the decision log and rerun the relevant rehearsal. That creates a controlled change, rather than an informal exception that quietly becomes the new default. Review the list after the first week of real use. Count the requests for exceptions and inspect one finished output for each task type; both reveal where the settings and the written policy disagree.
Further reads
- ChatGPT Business Alternatives for EU Firms: Data Terms Compared — Compare the contractual data boundaries.
- Is ChatGPT Available for Commercial Use? What Owners Should Check — Commercial use is allowed, but plan terms decide training, confidentiality and indemnity. A checklist for owners, with the vendor terms linked.
- ChatGPT Productivity Tool for Plumbers: Free, Plus or Business? — Which ChatGPT plan a plumber should pay for: break-even minutes, seat costs for one to five people, three worked jobs and the data rules.
- ChatGPT Business for Law Firms: Seats, Privacy Settings and Cost — What ChatGPT Business costs a small law firm at three sizes, the privacy and sharing settings to fix before the first prompt, and what it cannot provide.
- ChatGPT Business vs Claude Team: Seat Costs for a Small Consultancy — Seat sums for two, four and eight consultants, where the real cost differs, and a two-hour test of data terms you can run yourself before buying.
- ChatGPT and Copilot for HR: Which Suits a Small People Team? — A four-question rule, seat costs for a three-person people team, and the employee-data handling that matters more than which assistant you choose.
- ChatGPT Business or Enterprise: When a Small Firm Should Upgrade — A practical upgrade test for small firms comparing seats, identity controls, retention, support and the true cost of an Enterprise quote.
- ChatGPT Work Usage Limits and Credits: What a Business Plan Allows — A five-person usage calculation shows where Business seat allowances stop, how workspace credits begin and what owners can control.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI help, Managing members, seat types, and roles in ChatGPT Business; Managing data, sharing, and privacy in ChatGPT Business; Managing app permissions in ChatGPT; Memory in ChatGPT; Managing billing and seats in ChatGPT Business; OpenAI Business pricing, checked 5 November 2026.