Skip to content
← AI articles

Practical guide · AI at work

AI assistants at work: start with a plan that scales

The Anthropic, OpenAI and Mistral logos on three façades separated by two diagonal cuts.
AI-generated illustration.

In short

Many businesses want something simple: give their teams ChatGPT, Claude or Mistral, and stop personal accounts being used for work. This guide describes the rollout step by step, with one clear position: start with a managed plan that grows with the business, rather than hosting a model yourself.

  • A managed plan grows one account at a time. The vendor provides the interface, the models and the updates. You start with one team, add accounts later, and change plan when your size requires it.
  • Provide a tool before you ban anything. A ban with no alternative pushes people towards unapproved tools.
  • The rollout takes six steps: decide on uses and data, choose the plan, configure access, protect files, decide what must be blocked, then inform, train and test.
  • No training on your files, provided you set it up. Depending on the vendor, this is the default setting or a box to untick; step 4 gives the exact path and six tests to run.
  • Prices change; the criteria stay. This guide does not dwell on pricing: it gives the questions to ask, which will still apply the next time price lists change.
  • Self-hosting is justified by control, not by cost. It applies to a few specific cases, and the question comes up later.

Each step separates what management decides, what the IT provider configures and what you need to check before buying. A limited first rollout can be quick; its timetable depends on contracts, security requirements and any required consultations.

Why a managed plan rather than self-hosting

Why personal accounts need rules

  • The business controls nothing : not the settings, not the access, not the retention period, and no data processing agreement binds the vendor to it.
  • The account leaves with the employee, history included.
  • Conversations may be used for training, depending on the vendor and the settings. At Mistral, this is the default for individual users (Mistral). It is not universal: OpenAI and Anthropic offer individuals a setting.

Provide a tool before you ban anything: a ban with no alternative pushes people towards unapproved tools.

What a managed plan provides

  • A start with no infrastructure. No server to buy or run: you open a company workspace, link your domain and create the accounts.
  • Growth in stages. Ten accounts today, a hundred tomorrow: you add seats, then move from the team plan to the Enterprise plan as administration needs grow.
  • A complete product. The interface, document search, connectors and support are included. With self-hosting, you have to build and maintain them.

Why not self-hosting, to begin with

Hosting a model yourself gives you full control over your data. For a first assistant rollout, it is not the right starting point.

  • Fixed costs weigh heavily at small scale. Servers, operations and staff must be paid for whatever the number of users.
  • Hardware ages quickly. It is depreciated over several years, while vendors’ models and prices change with every generation.
  • The models are not the same. In the Artificial Analysis index launched at the end of October 2026, the best closed models remain well ahead of open-weight models (ActuIA). Measure the gap on your own tasks.

Self-hosting is justified by control, not by cost: data that must not leave the business, or a client that requires SecNumCloud-qualified hosting (the French state security qualification). In that case, a qualified hosting provider is often a middle way. For everyone else, the question comes back later, with several hundred active users and measured usage.

Step 1: define permitted uses and data

For management to decide. Before choosing a tool, state what may be put into it.

Category Examples In the provided tool
Public or internal Documentation, procedures, meeting notes Permitted
Confidential Client contracts, margins, code According to the uses set by the business
Personal Customer or employee data For uses defined after analysis, with the DPO if there is one
Secret or access data Passwords, access keys, documents classified as secret Excluded

At the start, allow conversation and file uploads. Keep web search, connectors, external sharing and features that can take actions switched off until they have been approved.

Step 2: choose the plan

Check before buying. Vendors generally offer two tiers: a self-service team plan and a negotiated Enterprise plan. The table keeps what matters for getting started and then growing, as of 11 November 2026.

Plan Size Things to know
ChatGPT Business Usually 2 to 200 seats Single sign-on included; European data residency not offered for Business in the documentation reviewed
ChatGPT Enterprise On quotation European data residency for eligible customers; network blocking of personal accounts
Claude Team 2 to 150 members Single sign-on and domain linking included (Claude)
Claude Enterprise Minimum of 20 seats Network blocking of personal accounts; European data residency to be requested by contract; usage billed on top of the seat
Mistral Vibe Team Not specified Domain verification; no SAML SSO; training to be switched off by the administrator
Mistral Vibe Enterprise On quotation SAML SSO; training off by default

The criteria that matter for growth

  • The size covered. A minimum and a maximum number of seats: beyond the team plan’s ceiling, you have to move to the Enterprise plan.
  • Identity. Linking the email domain and single sign-on, so that access can be created and removed from the company directory.
  • Administration settings. Switch a feature off for everyone or for a group, set the retention period, export the logs.
  • Quotas. What the seat includes, and what happens when a user goes over the limit.
  • Exit. How to get your data back if you change vendor.

Training. By default, OpenAI and Anthropic do not train their models on data from their business plans; at Mistral, the administrator must switch it off on the Team plan, and it is the default on Enterprise. Check it in the contract, distinguishing three things: training, retention and access by subprocessors.

What about price?

Price lists change often; check them on each vendor’s website on the day you buy. As of 11 November 2026, team plans are listed at around $20 to $25 per seat per month, and Enterprise plans are usually negotiated on quotation. Two identical listed prices are not always equivalent: some plans bill usage on top of the seat. In the budget, also include setup by the IT provider, support and any filtering, then set a ceiling.

Ten questions to ask the vendor

  1. Where does inference take place, and can it leave the EU, for example during load peaks?
  2. Where are prompts, files, memory and logs stored, and for how long?
  3. Which subprocessors are involved, and in which countries?
  4. Is zero retention possible, and with what exceptions?
  5. Can connectors and web access be restricted by user group?
  6. Which actions can the assistant carry out without approval, and can this be prohibited?
  7. Can the logs be exported to our security tools?
  8. What are the quotas, and how are overages billed?
  9. For a SecNumCloud service: which qualification number, which scope, which expiry date?
  10. How do we get our data and settings back if we leave?

Step 3: configure access and features

For the IT provider to configure.

  1. Verify the email domain, then, if you wish, turn on automatic joining for company addresses. At Mistral, this is a separate option from verification (Mistral Docs).
  2. Turn on single sign-on when the plan includes it, then build the tool into the joiner and leaver procedures. Single sign-on alone is not enough to remove access: disable the account in the directory, revoke the sessions, and test that a disabled account no longer has access.
  3. Switch off training if it is not the default setting.
  4. Switch off what is not permitted : web search, connectors, external sharing, action features. How fine-grained these settings are varies by vendor and plan.
  5. Set conversation retention when the plan allows it, usually on Enterprise only.
  6. Check Microsoft 365 or Google Workspace : which third-party applications can employees authorise there? A personal AI account can connect to it.

Adoption statistics, security logs and conversation content are three different things. Exporting logs for compliance is reserved for Enterprise plans at OpenAI and Anthropic; at Mistral, it needs to be confirmed and tested.

Step 4: protect the files given to the assistant

Configure, then test. This is the question everyone asks: “If I upload a contract to the assistant, is it used to train the model?” The answer comes down to three distinct settings, to be handled separately: training, retention and access.

Where a file actually goes

The same document can end up in four places, each with different rules.

  • In a conversation, as an attachment: only the user can see it, and the file lasts as long as the conversation.
  • In a project or shared space, as a knowledge base: every project member can query it, and it stays there as long as the project exists.
  • Behind a connector (Google Drive, SharePoint, email): the file is not uploaded; the assistant reads it at the source with the user’s permissions.
  • In a share link : a conversation or generated document made accessible to others, sometimes outside the business.

Four places for a file, three settings to protect them

Company workspace at the vendor Employee, company account ConversationAttachmentVisible to its author Shared projectReference filesVisible to members ConnectorRead at the sourceEmployee’s permissions Share linkPublished conversationVisible to outsiders 1 · TrainingTurned off in the console 2 · RetentionPeriod set, projects wiped 3 · AccessMembers and rights checked Company workspace at the vendor Employee, company account ConversationAttachmentVisible to its author Shared projectReference filesVisible to members ConnectorRead at the sourceEmployee’s permissions Share linkPublished conversationVisible to outsiders 1 · TrainingTurned off in the console 2 · RetentionPeriod set, projects wiped 3 · AccessMembers and rights checked
  • Setting to apply
  • File location
  • Company workspace
The three settings apply to all four locations, not only to conversations.

Setting 1: training

The default setting depends on the vendor, and it should be checked in the admin console before inviting the first user.

Plan Default setting What to do
ChatGPT Business and Enterprise No training on workspace content (OpenAI Academy) Check that each employee is actually working in the company workspace, and not in a personal account left signed in in the same browser
Claude Team and Enterprise No training on inputs and outputs, unless the user sends thumbs feedback: the conversation concerned may then be used (Anthropic) Switch off feedback: organisation settings, “Data and Privacy”, “Rate chats”
Mistral Vibe Team Training on until the administrator switches it off; attached documents count as input data (Mistral) On admin.mistral.ai: “Manage”, “Vibe”, “Privacy”, then switch off “Allow your interactions to be used to train our models”
Mistral Vibe Enterprise No training Check that the setting has not been switched back on by an administrator

Three pitfalls come up often.

  • The setting only covers the plan where it is made. At Mistral, the Vibe setting and the API setting are separate: switching one off does not switch off the other.
  • The feedback button is a back door. A thumbs up or down can send the conversation concerned to the vendor. At Anthropic, it is then kept for up to five years.
  • A setting needs checking more than once. Repeat the check after every plan change, and keep a dated screenshot in the decision sheet.

Setting 2: retention

“No training” does not mean “no retention”. By default, conversations and their files stay with the vendor until someone deletes them.

  • ChatGPT. Conversations are kept until the user deletes them. On Enterprise, the workspace owner can set a retention period, with a minimum of 90 days (OpenAI Academy).
  • Claude. Data is kept indefinitely by default. On Enterprise, an owner can set a period in “Data and Privacy”, with a minimum of 30 days. A project’s period overrides that of the conversations it contains, and changing the project’s files restarts the clock (Anthropic).
  • Mistral. The period and how to set it need to be confirmed in the admin console and in the contract.

In practice: the assistant is not a storage location. The reference file stays in the company’s own tool, and the copy uploaded to the assistant is meant to be deleted. On a team plan with no retention setting, deletion relies on users: write it into the usage policy, and name an owner for each shared project, responsible for emptying it once it is no longer in use.

Setting 3: access

  • Shared projects. Only upload files that every member is allowed to read. A “Management” project and a “Sales” project are better than a single project where everything is mixed.
  • Connectors. In Claude, an owner enables them for the organisation under “Connectors”, then each member connects their own account. The assistant inherits existing permissions: it only reads what the user can already open (Claude). In ChatGPT, administrators choose which applications are available to the workspace (OpenAI).
  • Overly broad permissions. A connector creates no permissions, but it makes existing ones visible. An “HR” folder opened to the whole company by mistake becomes searchable in one sentence. Clean up shared folder permissions before enabling a connector.
  • Actions. Claude’s Google connector can also send a message, share or delete a file. By default, the assistant asks for confirmation before each action; owners decide whether members can remove this confirmation. Leave it in place.
  • Share links. Switch off public sharing of conversations when the plan allows it, or limit it to members of the organisation.

Six tests to run with a dummy file

Create a fictitious document, marked with a recognisable sentence, and check before opening the tool to teams:

  1. Training. With an administrator account, then with a standard user account, check that the setting is switched off and that a user cannot switch it back on.
  2. Feedback. Check that the thumbs have disappeared from the interface, or that sending them is disabled.
  3. Separation. User A uploads the file to a private project; user B searches for the sentence and must find nothing.
  4. Connector. A user with no permission on a folder asks the assistant to summarise a file in it: the request must fail.
  5. Deletion. Delete the conversation and the file, then check that they no longer appear in the history or in search.
  6. Leaver. Disable a test account and check that it can no longer reach conversations or projects.

Step 5: decide what must be blocked

Decide, then configure.

The foundation is organisational: a provided tool and a usage policy. Technical protections are added according to the data, the devices and remote working, not according to headcount: a ten-person practice may handle more sensitive information than a business of two hundred.

Situation Measure to consider Limitation
Limited uses, data of low sensitivity, no IT team Administered plan, strong authentication, simple rules, leaver procedure Personal accounts remain accessible
Managed workstations, staff often working remotely Managed browser or endpoint protection, filtering, restricted extensions A managed browser does not cover other applications
Need to enforce the company workspace on the chosen service Plan that supports blocking outside accounts, and a gateway applied to the relevant traffic Test browsers, applications and workarounds
Work use on mobile Device or work app management Does not block personal accounts on its own
Sensitive data, traceability requirement Contractual commitments, access control, filtering, cloud service control Check the exact capabilities of each product

Block outside accounts on the chosen service. The company gateway adds an identifier to every request, and the vendor refuses any account outside the organisation. OpenAI documents this for ChatGPT Enterprise (OpenAI), Anthropic for Claude Enterprise (Claude). For Mistral, nothing is confirmed in the public documentation as of 11 November 2026. The mechanism requires a gateway that inspects encrypted traffic, installed on site or provided as a cloud service. It only covers traffic that passes through it: a work laptop at home escapes it if its traffic goes out directly, and so does a personal phone on 4G.

Step 6: inform, train and test before a wider rollout

Employment law. Before introducing logging that makes it possible to monitor employees’ activity, define its purpose, the information collected, how long it is kept and who is authorised to consult it. Under French law, employees must be informed before any collection (article L. 1222-4 of the Labour Code), and in businesses with at least 50 employees the works council (CSE) must be consulted before the decision (article L. 2312-38). Monitoring must remain justified and proportionate.

Tests before a wider rollout : sign in with a personal account from a company workstation, in the office and remotely; try the desktop and mobile apps; disable an account and check that access is lost; check that disabled features really are disabled.

A usage policy in five rules

  1. Use the authorised work space. Company information is only processed in approved tools. A personal account, even a paid one, is no substitute.
  2. Limit the information shared. Only what the task requires; never passwords, access keys or secret documents. If in doubt, ask the designated contact person.
  3. Do not open new access without authorisation : connecting an email account or business software, installing an extension, enabling an action feature, sharing a conversation externally.
  4. Check before using the result. Check facts, figures, sources and recipients; have anything with legal, financial or human consequences approved.
  5. Report errors quickly. Data sent by mistake or suspicious behaviour should be reported to [contact]. A report made in good faith is not, in itself, grounds for disciplinary action.

Decision sheet to complete

Decision Expected answer
Authorised tool and workspace Name, plan, access link
Equipped staff Number of accounts, roles
Budget Monthly total, commitment, extras, ceiling
Permitted data Concrete examples and exclusions
Permitted features Conversation, files, web, connectors, sharing
Technical coverage Office, home, mobiles, applications
Owner Administration, leavers, support, incidents
Tests carried out Access, blocking, revocation

Growing without starting over

A well-framed rollout at ten accounts extends to two hundred without changing method. What changes is the plan and the level of control.

  • Add accounts in waves. One team first, then the next ones, repeating the usage policy, the training and the step 6 tests with each wave.
  • Change plan at the right time. Team plans have a ceiling: 150 members for Claude Team, usually 200 seats for ChatGPT Business. Enterprise plans add what is missing at that size: at OpenAI and Anthropic, network blocking of personal accounts and log export for compliance, and usually the retention setting.
  • Connect the tool to the directory. The more accounts there are, the more single sign-on and the joiner and leaver procedure matter: a forgotten account is an open door.
  • Open features one at a time. Web search, then read-only connectors, each after approval. A read-only connector widens the data within reach without turning the tool into an agent.
  • Track usage and spending. Look at adoption statistics, identify users who reach the quotas, and review the ceiling set in step 2.
  • Keep an exit route. The data rules, the usage policy and the decision sheet do not depend on the vendor: they can be reused as they are if you switch.

Four points are enough for an assistant. They do not replace advice from the DPO or a lawyer.

  • Train your teams. The European AI Act requires businesses that deploy an AI tool to take measures to ensure a sufficient level of AI literacy among their staff (Cooley).
  • Apply the GDPR. The CNIL, the French data protection authority, recommends starting from a concrete need, limiting the personal data entered, checking reuse by the provider, adopting a usage policy, providing training and involving the DPO from the start (CNIL, July 2024).
  • Protect what the GDPR does not cover. Plans, margins, code and configurations may qualify as trade secrets, provided in particular that they are subject to reasonable protection measures (article L. 151-1 of the French Commercial Code). The data classification in step 1 is part of this.
  • Comply with employment law as soon as logging makes it possible to monitor employees’ activity: see step 6.

What comes next?

Two topics go beyond this guide. They come after a well-established assistant, not before.

  • Agents. An assistant answers; an agent acts. It reads files, runs code, writes to your applications and can send messages. You then need to know what it can read, what it can do and where it can send data. Until these features are explicitly authorised, leave them switched off.
  • Self-hosting. It becomes an option again when headcount, measured usage or a control requirement justifies it. The rules set out here still apply.

Conclusion

The right first step is neither choosing the best model nor setting up your own servers. It is providing an administered tool, saying what may be put into it, and being able to add or remove an account in a few minutes. Models and prices will keep changing; this framework will remain.

Sources

Checked on 11 and 12 November 2026.

Plans and settings

Public pricing

Legal framework and models