ChatGPT Business for companies: plans, agents, data and rollout (2026 guide)

ChatGPT Business lets every employee hand whole tasks to ChatGPT: summarizing three sets of meeting notes, analyzing a spreadsheet, preparing a presentation. It is OpenAI’s plan for teams, called ChatGPT Team until August 2025: a shared workspace administered by the company.

An example from OpenAI’s release notes of September 29, 2026: a team task gathers a project’s progress and blockers every week, then posts the summary to an approved Slack channel.

Every agent run consumes usage and acts with the rights of the connected account. Cost and risk are therefore settled in the admin console and the choice of plan, more than in the price per seat. Facts checked against OpenAI on October 14, 2026.

Key takeaways

• ChatGPT Team and ChatGPT Business are the same plan, renamed on August 29, 2025. To get started without the technical parts, go straight to the practical part.
• A Standard seat is listed at $20 per month billed annually, from two users. A Premium seat gives five times the usage for $100 per month billed annually.
• Beyond the included usage, the workspace consumes prepaid credits. No cap is set by default, but the admin can set one on day one, and with no credits left the feature stops rather than billing more.
• Custom GPTs will be retired on December 11, 2026 and replaced by plugins; their custom actions do not migrate.
• OpenAI does not train on Business data, but responses are not computed in Europe on Business: only Enterprise offers that.
• A booby-trapped document can hijack the assistant: OpenAI writes that an instruction hidden in a file can change ChatGPT’s behavior.

What ChatGPT Business is in October 2026

Since July 9, 2026, the new ChatGPT desktop app for macOS and Windows brings three modes together.

• Chat: conversation. GPT-6 has been rolling out there since October 7, 2026 for eligible Business workspaces.
• Work: an agent that carries a long task end to end. It researches, analyzes, works in your files and tools, and delivers a document, spreadsheet, presentation or Site, a small web app hosted by OpenAI. You follow its progress and approve important actions.
• Codex: OpenAI’s tool that writes and fixes code for developers, and also handles files and folders. Its app was merged into the ChatGPT app on July 9.

Work reaches your tools through connectors, which OpenAI now calls apps: authorized links to a piece of software. The pricing page, checked on October 14, 2026, cites Microsoft 365, Google Drive, Slack, GitHub, Linear and Figma.

The same page also lists shared projects, scheduled tasks, the built-in browser, company knowledge, record mode for meetings and add-ins for Excel, Word, PowerPoint and Google Sheets.

On the admin side, the plan includes single sign-on (SSO), which lets employees sign in with their company account, multi-factor authentication, an admin console, email domain verification and spend tracking. The OpenAI API, which developers use to build the models into their software, is not included.

Plans and pricing: Standard seats, Premium seats and credits

Seats

Taken on October 14, 2026 from OpenAI’s help center and the French pricing page:

	Standard	Premium	Enterprise
Billed annually	$20 per month (€21 on the French page)	$100 per month	On quote
Billed monthly	$25 per month (€26)	$125 per month	On quote
Included usage	Baseline, capped in five-hour windows	Five times Standard, no five-hour window	Credit pool set by contract
Payment	Credit card	Credit card	Invoice, volume discounts

These are published list prices: the French page does not say whether tax is included, and it does not yet show Premium in euros.

The five-hour window is a rolling period during which a Standard seat’s usage is capped: once the limit is reached, the user waits for it to free up. Premium, introduced on August 24, 2026, has no such window and resets weekly.

A Premium seat costs five times a Standard seat and gives five times the usage, so the price of included usage is the same. Premium is for people the five-hour window interrupts, typically developers working with Codex all day.

A Business workspace requires at least two seats of either type, and the two types can be mixed. Invoice billing, purchase orders and zero data retention require, according to OpenAI, a contracted offering.

Credits

Beyond each seat’s included usage, the workspace draws on a shared credit pool, prepaid by the workspace owner. According to the help article on credits, they are valid for twelve months.

OpenAI does not publish a unit price per credit: it is shown at purchase. Its rate card, updated October 14, 2026, does publish consumption once included usage is used up:

Usage	Credits
Message to GPT-6	10
Deep research	50 per task
One end-to-end workspace agent run	Typically 5 to 25
Task in Excel or Google Sheets	5 to 20 per message
Task in PowerPoint	10 to 50 per message

Messages to the default fast model remain unlimited. An order of magnitude is therefore counted in runs: a hundred agent runs a month come to 500 to 2,500 credits. For Codex, OpenAI estimates the average cost at $100 to $200 per developer per month, with wide variation.

Setting a cap on day one

Four settings, all in Workspace settings > Billing, keep the bill under control:

1. a monthly credit limit per seat type, Standard and Premium, and per user if needed;
2. usage alerts, in Billing > Usage Alerts;
3. automatic reload turned off, or capped with a monthly limit: left blank, that limit is unlimited;
4. a modest first credit purchase, since credits are prepaid.

By default, no cap is set. But according to OpenAI’s page on flexible pricing, when the pool is empty, the feature is blocked and the user asks the owner to add more: spending cannot exceed the credits bought.

Worked example: 10, 50 and 200 seats

The calculations use the help center’s dollar prices, billed annually, with one Premium seat per developer or heavy user.

Organization	Mix	Monthly calculation	Per month	Per year
10 seats	8 Standard, 2 Premium	8 × 20 + 2 × 100	$360	$4,320
50 seats	45 Standard, 5 Premium	45 × 20 + 5 × 100	$1,400	$16,800
200 seats	180 Standard, 20 Premium	180 × 20 + 20 × 100	$5,600	$67,200

Billed monthly, the 200 seats cost 180 × 25 + 20 × 125 = $7,000 per month, 25% more than annual billing.

Agents come on top. If the 200-seat organization runs twenty team agents once a week: 20 × 4 × 5 to 25 = 400 to 2,000 credits per month, charged first to included usage, then to the pool.

Since OpenAI publishes no typical consumption per organization, a one-month pilot with a cap in place gives the real figure before any annual budget.

Where to start

For an employee or team lead without a technical background, five steps are enough.

1. Open the ChatGPT desktop app and choose Work. Describe a real task with files attached. For example: “From the three attached meeting notes, write a summary of the decisions and the list of actions with their owner and due date.”
2. Follow the work. Answer the agent’s questions and approve or decline the actions it proposes, especially sending a message or editing a file.
3. Check the deliverable against the source documents before sharing it.
4. Connect only the tools the task needs, and only those your organization has approved.
5. Turn recurring work into routines. A weekly task becomes a scheduled task, then a team task or a shared agent.

If you use ChatGPT Business for your own work or your team’s, you can stop here. The rest is for IT leadership, CISOs (chief information security officers), DPOs (data protection officers) and rollout leads.

Moving from personal accounts to the Business workspace

Many employees already use ChatGPT on a personal account. OpenAI’s page on onboarding employees with existing accounts describes what happens.

• The invitation goes to the work address, and the employee must accept it with that same address.
• On Business, the employee chooses to keep their personal workspace separate or merge it into the company workspace; admins cannot require a merge or a deletion.
• Merging cannot be undone: supported content is copied, the personal workspace is deactivated, chats cannot be picked one by one and Sites do not transfer. Merged content then falls under the company’s retention rules.
• OpenAI recommends a separate account, on a personal email address, for private use.

One point is worth telling teams: on the individual Free, Plus and Pro plans, OpenAI writes “we may use your content to train our models”, unless the “Improve the model for everyone” setting has been turned off.

Company documents pasted into a personal account in the past may therefore have been used for training.

Workspace agents, team tasks and plugins

Workspace agents

A workspace agent is an assistant configured once for a repeatable task, then shared in the workspace. According to the ChatGPT Business release notes, it began rolling out on April 22, 2026, has been generally available since May 21, and since July 6 has been billed by the tokens consumed.

Tokens are the units of text the model reads and produces.

An agent is built from a template or from scratch, connects to Google Drive, Google Calendar, Slack or SharePoint, is previewed before publishing and can run on a schedule. Its builder sets, app by app, which actions are allowed; the admin tracks each agent’s activity in the console.

Team tasks and @ChatGPT

Since September 29, 2026, team tasks let colleagues share recurring work, as in the opening example. The same day, @ChatGPT arrived in Slack and Microsoft Teams: the organization installs it, the admin sets its instructions and connections, and employees mention it in a channel.

The documentation is explicit: for a team task, the connected account determines what the task can read and do. If the shared connection belongs to a director, the task sees what that director sees. Colombani.ai recommends a service account with limited rights for these shared connections.

Plugins and skills

A plugin is a reusable package that combines, for a given job, skills (sets of instructions ChatGPT applies when a request calls for them), reference files and connectors. On July 9, 2026, the plugin directory replaced the app directory.

Since October 1, admins manage plugins in the console, choose for each whether it is offered to employees or installed by default, and can import the organization’s own plugins from GitHub.

Custom GPTs are going away: timeline and migration

Timeline

According to OpenAI’s FAQ on GPT retirement, updated in early October 2026:

• December 11, 2026: custom GPTs retire, on all plans;
• February 11, 2027: retirement only for Enterprise workspaces with an approved deferral;
• October 26, 2026, a planned date for affected Enterprise workspaces: creation of new GPTs ends.

For a Business workspace, the applicable date is the one in the in-product notice. Past conversations with a GPT remain accessible after retirement.

What carries over and what does not

Migration starts from My GPTs, via Migrate to plugin. It uses the GPT’s latest published version: drafts do not transfer.

Carries over	Does not carry over
Instructions, which become a skill in the plugin	Custom actions
Knowledge files, copied as reference files	The model selected for the GPT
Connected apps, added to the plugin	Sharing settings: the plugin starts private

Rebuilding a custom action may require an MCP server (Model Context Protocol, the standard protocol that connects an agent to an external tool) and technical work. The original becomes read-only after migration. The FAQ says nothing about conversation starters or the GPT’s capability settings: check them during testing.

Method

1. Inventory the GPTs in use, with their creators and users; the admin analytics console lists them.
2. Publish useful drafts, without necessarily sharing them publicly.
3. Migrate, then test the plugin on a few familiar prompts and one hard case.
4. Rebuild custom actions with IT, after reviewing their permissions.
5. Share the plugin and ask a real user to install and try it.

Your data: training, retention, certifications, residency

Training and retention

For Business, OpenAI states that it does not use prompts, responses or files to improve its models unless the organization has signed a separate agreement. Response feedback does not permit it, and admins can turn feedback off.

On Business, admins set the retention period, according to OpenAI’s enterprise privacy page. A deleted conversation is removed within 30 days, unless the law requires otherwise or it is reasonably necessary to protect the services or third parties. Zero data retention is only available under contract.

What the admin can see

Workspace admins can view, export and delete employees’ conversations. Employees should be told before rollout.

In France, a tool that allows this kind of oversight also involves employee representatives: article L2312-38 of the French Labour Code requires, in companies with at least fifty employees, the CSE, the works council, to be informed and consulted before the decision to implement any means of monitoring employees’ activity.

Certifications

SOC 2 Type 2 is an audit, run by an independent firm, of how effective security controls are over a period. According to OpenAI, Business has passed it.

ISO 27001, 27017, 27018 and 27701, international standards, cover information security management, cloud security, personal data in the cloud and privacy management. The pricing page ticks them, at plan level, for Enterprise only.

Data residency and inference residency

Data residency sets where content is stored. Inference residency sets where the model computes responses, inference being that computation. The EEA is the European Economic Area: the European Union plus Iceland, Liechtenstein and Norway.

	Business	Enterprise and Edu
Storage in a chosen region	Rolling out, not yet available to all; region chosen at checkout	Ten regions, including Europe (EEA and Switzerland), for new customers
Responses computed in region	No	Europe (EEA and Switzerland), United States, United Arab Emirates
Copy in the United States	Every prompt and response, for a limited time, for abuse monitoring	Enterprise-specific terms
Excluded features	Shopping in ChatGPT	Codex Web, third-party GPTs, Sites, beta features

Sources: OpenAI’s page on Business storage and its page on residency for Enterprise and Edu. The pricing page marks residency as absent from Business: rolling out, storage at rest only.

Even with inference residency, OpenAI says some processing stays outside the region: extracting text from a PDF, request routing, authentication.

The copy kept in the United States is held by a US company, subject to the CLOUD Act, the US law on authorities’ access to data held by service providers. Our Mistral guide details what that law says.

The contract

OpenAI offers a DPA (data processing agreement) for Business, Enterprise and the API: it is the contract that governs OpenAI as a processor under the GDPR, a provider that handles data on the company’s behalf. It is requested through a form; Sites have their own agreement.

For Enterprise, the pricing page mentions “custom legal terms” for eligible customers. OpenAI’s sub-processor list is public and names Microsoft, CoreWeave, Oracle, Google Cloud and AWS, among others; the DPA gives 30 days to object to a change.

Rolling out ChatGPT also makes the organization a deployer under the EU AI Act. Our article on what has applied since August 2, 2026 covers those obligations.

Security: connectors, injection, Sites, memory

Connectors

A connector never grants more rights than the account that connects it: OpenAI states that access in ChatGPT does not override permissions in the connected system. It does inherit all of that account’s rights.

Admins choose which apps are enabled and, for some, which write actions are allowed: posting in Slack, creating a reminder, uploading a file. Enabling read before write remains the simplest rule.

Injection through documents

Prompt injection means hiding, in content the assistant reads, an instruction that hijacks its behavior. OpenAI’s page on Lockdown Mode says so: a malicious instruction hidden in an uploaded file can still change ChatGPT’s behavior and produce a wrong answer.

Lockdown Mode limits the outbound requests that could exfiltrate data: browsing limited to cached content, agent mode and deep research disabled. It does not apply to Codex and does not stop every injection, OpenAI notes.

On Business, each user turns it on in Settings > Security. The documentation only provides for admins to impose it in managed workspaces, through a dedicated role using role-based access control, which is reserved for Enterprise.

Public Sites

On Business, Sites are enabled by default. Public publishing, through an address open to anyone, was not available in the EEA, Switzerland or the United Kingdom at its launch on July 9, 2026. Since September 3, a Site can be shared with named people outside the workspace.

A Site that reads connected apps requires admin approval, plugin by plugin; that setting is off by default. Admins can disable created Sites in Workspace settings > Sites.

Memory and history

Since June 25, 2026, Business memory draws on past conversations. Each user can view a summary of what it retains, correct it or turn it off.

Computer History, on macOS, records activity in selected apps and sites to give ChatGPT context. It is off by default, must be allowed by the admin, and has been available in the EEA since August 20, 2026.

Admin console decisions

Defaults and supporting references come from the OpenAI documentation cited above.

Decision	Default	Recommendation	Basis
Monthly credit cap	None	Cap per seat type and alerts before the first purchase	OpenAI advises reviewing seats, credits, reload, alerts and limits together
Automatic reload	Unlimited if the monthly limit is blank	Off, or capped	Credits page
SSO and domain verification	To configure	Turn on before inviting employees	Employee onboarding page
Connectors and write actions	Set per app	Read-only first, write case by case	OpenAI recommends enabling only trusted apps and avoiding writes with uncertain visibility
Plugins	Offered or installed by default	Approved list, a named owner	Plugin console
Sites	On, without connectors	Decide who publishes before opening up	OpenAI advises deciding who creates and publishes before enabling broadly
Computer History	Off	Keep it off unless a documented need exists	Release notes
ChatGPT for Word	On since October 1, 2026	Decide, since it draws on the shared allowance	Release notes
Rollout	None	Start with a small group	OpenAI recommends testing with a small group

Business or Enterprise: three questions to decide

1. Does your data require the model to run in Europe, or ISO certifications at plan level? Only Enterprise provides them.
2. Do you need to manage accounts at scale? SCIM automatically syncs joiners and leavers with the company directory, and role-based access control grants rights by group of employees. Both are Enterprise-only.
3. Do you need a custom contract: invoice billing, custom retention, compliance logs, an IP allowlist, or EKM (enterprise key management), where the company manages its own encryption keys? These are Enterprise features.

Three noes: Business is enough. A single yes justifies requesting an Enterprise quote.

Enterprise pricing is not public. According to the pricing page, the quote depends on the number of seats, how usage is billed (in credits or in tokens), volume discounts and the level of support. Data residency is included at no extra cost.

ChatGPT compared with Claude, Copilot and Mistral

We train on all four tools. The criterion that most often separates them is where your documents and identities already live.

Tool	Team entry price, billed annually	Price basis	What sets it apart
ChatGPT Business	€21 per seat per month	Published list price, tax not stated	Chat, agent and coding in one app, shared agents
Claude Team	€18 per seat per month	Excluding tax	Long tasks on folders of files, shared skills
Microsoft Copilot	€26.00 per user per month	Excluding VAT	Inherits Microsoft 365 permissions
Mistral Vibe Team	$24.99 per user per month	Excluding tax	French vendor, on-premises deployment possible on Enterprise

Prices taken on October 14, 2026 from the vendors’ French pages; Mistral lists its prices in dollars. These tools are not mutually exclusive: an organization can deploy several, depending on teams and data sensitivity. To go deeper on each:

• our Claude Cowork guide, on long tasks and their security;
• our Copilot Studio guide, on Microsoft’s agents, their credits and their governance;
• our Mistral for business guide, on Vibe and sovereign deployment.

Training your teams

The console sets what the tool can do. Usage sets what employees entrust to it. Three situations need preventing: a connector attached to an over-privileged account, an agent acting without review, a deliverable shared unchecked.

Training therefore covers three habits: knowing which data goes into which workspace, scoping a task handed to Work or an agent, and checking what it produces before approving it.

Colombani.ai is certified Qualiopi, the French quality certification for training providers, which is a condition for public and pooled funding.

• For executives and business teams, the ChatGPT training comes as a half-day briefing and a day in which each participant hands their own tasks to ChatGPT Work and agents.
• For teams working on files, folders or code, the Codex training has each participant configure an agent on their real tasks.

French companies can request funding from their OPCO, the skills operator that funds private-sector employee training. For a public body, training is paid from the employer’s training budget.

To prepare the rollout, our kit brings together the console settings table with defaults and our recommendations, the GPT migration checklist to complete before December 11, a template of usage rules for employees and a thirty-day pilot plan.

Colombani.ai is Qualiopi-certified for training activities. Its founder, Ulysse Trin, holds the Claude Certified Architect Professional, issued by Anthropic, and nine practitioner certifications issued by OpenAI, including ChatGPT Deployment Practitioner.