An EU firm should compare Mistral Vibe for EU hosting and deployment control, Claude Team for a separate commercial workspace, Gemini in Google Workspace for existing Google data, and Microsoft Copilot for existing Microsoft 365 data. None wins on location alone. Check each contract, storage and processing location, retention rule and app permission against the data staff will actually enter.
The product name matters here: Mistral's current business assistant is Vibe, formerly Le Chat. Its Enterprise deployment choices differ from the ordinary Vibe Team chat product. Similarly, a vendor saying it does not train on business data does not mean all data stays in one region or disappears immediately after a user deletes a conversation. Make those separate rows in the purchase record.
Write the data requirement before opening a pricing page
Start with three example tasks, not a generic promise of “secure AI”. For each, note what is entered, who owns it, whether it includes personal data, where the source already lives, who needs the output and how long it must remain available. A procurement choice is defensible only if the same scope is put to each vendor. A short list might include drafting a public event notice, summarising an internal supplier contract and reviewing a spreadsheet of named customer complaints. Those are materially different data classes.
Use a simple classification. Public material can be used in an ordinary evaluation. Internal operational material needs a business account, access review and retention plan. Named customer records need the correct contractual basis, documented purpose, minimised inputs and legal review. If the team cannot say which class a document is in, do not upload it during the comparison. The pre-invite admin checklist gives practical controls for an OpenAI workspace; the same questions can be put to other vendors.
For an illustrative independent bookshop, the public class includes the opening-hours page and publisher catalogue descriptions. Internal material includes wholesale discount terms. Personal data includes a reading-club member list with email addresses. A trial prompt about opening hours tests writing quality without creating a privacy question. A contract summary tests how the tool handles long documents, but only after the firm has accepted the applicable business terms. The member list should be replaced by a synthetic sample until the legal and security review is complete.
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 nowCompare the terms that change the buying decision
| Offer | Contract and training position | Location and retention question | Likely fit |
|---|---|---|---|
| ChatGPT Business | Business Terms; workspace data excluded from training by default | Storage-region rollout is limited; processing and safety copies have separate treatment | Shared ChatGPT workspace when those limits are acceptable |
| Mistral Vibe Team or Enterprise | Team admins can switch training off for the organisation; Enterprise is opted out by default; review the data processing addendum | Hosted in the EU by default; Enterprise deployment choices; zero data retention does not apply to Vibe chat | Firms prepared to contract for a defined deployment |
| Claude Team or Enterprise | Commercial terms; review data processing addendum and connected tools | Traffic can be routed across several regions by default and data is stored in the US; confirm any contractual exception | Separate team workspace with project collaboration |
| Gemini in Workspace | Workspace agreement and processing terms; no training of Workspace customer data without permission or instruction | Admin-set Gemini history retention; check data-region settings by service, as some features sit outside them | Work already centred on Gmail, Drive and Docs |
| Microsoft Copilot | Microsoft product terms and data protection addendum; no foundation-model training on commercial prompts | EU Data Boundary commitments, except web search queries and Anthropic models; configured retention policies | Work already centred on Microsoft 365 files and permissions |
This table is a screening tool, not a legal opinion. A data processing addendum, or DPA, describes how a supplier handles personal data on a customer's behalf. It does not itself prove that a particular prompt never crosses a border. A storage promise concerns data at rest; inference is the processing that produces an answer; an abuse-monitoring copy may follow a separate path. Ask for each answer in writing for the exact plan, feature and date. The rules can differ between the chat application, connected apps, API and enterprise deployment.
OpenAI: storage choice has a narrower scope than many expect
OpenAI says Business data is not used for model training by default. Its current help page also says a Business storage-region option is rolling out gradually and is not available to everyone. A selected region controls primary customer content stored at rest, including prompts and replies. It does not include inference residency. Abuse-monitoring logs stay in the US whichever region is chosen, and if a non-US region is selected, OpenAI says a copy of every prompt and response is also stored in the US for a limited time for safety and enforcement. That is a material qualification for a firm with a strict processing-location requirement.
The public Business comparison still marks data residency as unavailable while the newer Help page describes a limited checkout rollout. Treat that as a live product discrepancy: inspect your actual checkout and ask OpenAI which wording governs the proposed account. Capture the available region choices, the help-page version and a written answer to the safety-copy question. Do not tell an auditor that “Business is EU-only” merely because a storage region appeared during signup. If the requirement includes in-region inference, seek a plan or vendor commitment that expressly covers it.
An illustrative tutoring centre wants to summarise lessons without sending named pupil notes outside an approved processing region. Its owner sees a storage choice at checkout and initially marks OpenAI as passed. The Help page shows that processing location is not fixed by that choice. The corrected result is “conditional”: test with synthetic lessons, request a written data-flow answer, and keep named records out of the trial. That one-word change in the procurement sheet prevents a misleading green tick.
Mistral: separate Vibe chat from enterprise deployment
Mistral now markets Vibe as its work assistant, with Team and Enterprise choices. Its product page says Enterprise can be deployed on premises, in a private cloud or on Mistral Cloud with full data residency. Those are choices to scope and contract, not features to assume in an ordinary Team subscription. For the standard service, Mistral's data-location help says data is hosted in the EU by default unless a customer deliberately uses its US API endpoint, and that subprocessors outside the EU are bound by contractual transfer safeguards. Ask which subprocessors handle the features and connectors you plan to use.
Training is the setting to fix on day one. Mistral's help says ordinary Vibe users are not opted out of model training by default, Team administrators can disable training for the whole organisation, and Enterprise customers are opted out by default. On a Team plan, an administrator should open the Admin panel, choose Vibe and switch off the toggle labelled “Allow your interactions to be used to train our models” before anyone uploads a document; Mistral treats uploaded documents as input data. Record the date the toggle was changed in the procurement file.
Mistral's zero-data-retention FAQ is explicit that zero data retention is unavailable for Vibe Work or Chat at any subscription tier because conversation history and application state have to be stored. Its API offers a separate zero-retention option for specified stateless endpoints, only on pay-as-you-go and only after Mistral approves a request, but that is not the same product as an employee using Vibe. For Vibe itself, Mistral says conversations are removed when a user deletes them or the account, typically after a short grace period. A buyer who needs no retained chat history should not treat an API promise as a Vibe Team promise. Deleting a conversation and retaining no prompt at all are also different requirements.
Consider an illustrative landscaping company whose procurement team wants staff to review supplier terms but keep contract files under a private-cloud arrangement. Its short list should compare an Enterprise Vibe deployment quote with the other vendors' actual enterprise offerings, not compare Vibe Team's public price with a bespoke hosting promise. A proof request can read: “Show the chat, file, index, backup and monitoring locations for this exact deployment, plus the deletion procedure when a project ends.” If the vendor cannot answer a line, leave it open in the decision record.
Claude: commercial privacy is not automatic regional isolation
Claude Team is Anthropic's team offering; Enterprise adds further contracting and administration options. Anthropic's commercial data guidance distinguishes work accounts from consumer accounts. Its server-location FAQ says traffic may be routed to selected countries across several regions by default unless otherwise agreed or instructed, and it notes that data is stored in the US. That is more useful than a broad statement that a provider has offices nearby: routing depends on the service and contract. Ask whether a specific regional arrangement is available for the chosen tier, and obtain the applicable subprocessor list.
Anthropic's organisation-retention FAQ says its work products keep chats in the product until users delete them, keep inputs and outputs from a chat flagged for a usage-policy breach for up to two years, and keep feedback submissions, such as a thumbs-up or bug report, for five years. Incognito chats are deleted within 30 days unless flagged. The exact path still depends on the feature, legal duties and any enterprise settings, so copy the current text into the procurement record rather than promising instant erasure. A firm needing a fixed deletion window should request the schedule for conversations, files, connected-app results, safety logs and backups separately. The fact that a staff member cannot see a deleted chat is not evidence that every copy has gone.
For an illustrative removals firm, Claude might be evaluated on turning 18 anonymised job notes into a clearer handover. The sample contains invented booking IDs, no names or addresses. The owner checks whether Claude's project sharing supports the handover, then asks Anthropic how project files are retained when a member leaves. The output can be excellent and still fail the firm's contractual requirement; quality and data terms are different columns. For broader task comparison, see the assistant choice by admin controls.
Gemini: use the Workspace agreement, not consumer settings
Google's Workspace privacy guidance says customer data in its generative AI services is handled under the Workspace agreement and is not used to train or fine-tune supporting models without customer permission or instruction. That statement applies to Workspace's commercial service, not automatically to every personal Gemini account a staff member might use. Check the administrator's enabled services and third-party extensions, then the retention controls: Google says admins can let users delete Gemini conversations and set automatic deletion after a period of inactivity, such as 3 months, 18 months or 3 years, or keep history indefinitely. Data-region settings need a feature-by-feature check; Google says files uploaded to Gemini Notebook are copied outside Drive and the organisation's data-region settings do not apply to them. A strong Google account configuration can be undermined if staff paste a customer document into a separate personal account.
Workspace pricing currently includes Gemini features across business tiers, with broader Gmail, Docs and Meet integration on Standard than Starter. That makes it attractive to a firm already paying for Workspace, but “included” is not the same as “approved for every file”. Examine Drive sharing first. If a user can already read a broadly shared folder, an assistant grounded in that user's access may surface material the firm meant to keep narrow. Fix the source permissions before judging the AI response.
A bookshop stores wholesale agreements in Drive. The test account can see a supplier folder marked “anyone in the organisation”. Gemini answers a query about discounts using that folder, even though the tester is not on the buying team. The error is an overbroad Drive permission, not an AI model inventing access. The remedy is to restrict the folder, rerun the same prompt and note the difference. Put that observed result beside the contract review; both are needed for a sound decision.
Copilot: existing file permissions become the AI boundary
Microsoft says commercial Copilot and Copilot Chat are covered by its product terms and DPA, with Microsoft acting as a processor. Prompts, responses and Microsoft Graph data are not used to train foundation models. Its enterprise data protection page says Copilot respects existing permissions, inherits sensitivity labels, applies the organisation's retention policies and supports audit. For an EU firm the key line is the EU Data Boundary: Microsoft lists it among its commitments, but says it does not apply to web search queries and that Claude models from Anthropic used inside Copilot are currently excluded from it. That can simplify procurement for a firm already using the suite, provided the file estate is properly governed.
It also creates a practical test: the assistant can surface content a user is authorised to read even if that access was granted accidentally years ago. Review broad SharePoint and OneDrive permissions before turning on grounded work. The ten-person licence comparison helps separate the incremental AI licence cost from the base suite cost, but the privacy decision should be made on effective access and retention. Ask your Microsoft administrator whether web grounding and Anthropic models are enabled for staff whose work must stay inside the boundary, and check each agent's own privacy statement and terms, as Microsoft advises.
Illustrative failure: a manager asks Copilot to draft a pay review and it cites an old personnel spreadsheet from a widely shared team site. The model has used the manager's existing permissions; the surprise is the site's sharing configuration. Remove the broad access, verify the manager can no longer open that file, and repeat the query. Keep screenshots of both access tests, with all names replaced in any vendor demonstration. This is a better control test than asking the vendor whether its model is “private”.
Make a scored choice with one worked comparison
Use mandatory gates before scoring convenience. An illustrative independent bookshop has six staff, uses Workspace already, and wants three tasks: public event copy, a monthly stock summary, and replies drafted from customer complaints. It sets four gates: commercial DPA applicable to the selected plan, no training on its business prompts by default, acceptable storage and processing routes, and a documented deletion path. A vendor that fails a mandatory gate is not rescued by a good writing score. The customer-complaint task stays synthetic until the gate review finishes.
Suppose each surviving tool is then scored out of five on task quality, fit with existing files, permission control and operating effort. For illustration, Gemini scores 4, 5, 4 and 4, totalling 17; Copilot scores 4, 2, 4 and 2, totalling 12 because the firm does not use Microsoft 365; Claude scores 5, 2, 3 and 3, totalling 13; Vibe scores 4, 2, 3 and 2, totalling 11 pending an Enterprise quote, and on a Team plan it passes the training gate only once an administrator has switched training off. These are invented decision scores, not vendor ratings. The arithmetic makes the reason visible: the existing file estate matters more than a one-point writing advantage.
The firm should then run the same redacted tasks with the top two choices. Record input preparation time, output correction time and any refusal or permission failure. If a draft saves six minutes but takes eight minutes to verify, it is not a useful win for that task. If a vendor's data answer remains unclear, write “unverified” and keep the task out of live use. Do not turn uncertainty into a soft score of three. The Business trial checklist provides a way to compare the work itself once the data gates are passed.
Ask the vendor questions that produce usable evidence
Send the same written questions to each shortlisted supplier: which signed or online terms govern this exact tier; which entity processes data; which locations hold prompts, replies, files, indexes, backups and safety logs; where inference runs; whether any content trains models; what happens to feedback; how deletion and export work; and which subprocessors handle connected-app content. Request links to the current contractual text and the date of the answer. Ask the vendor to distinguish default behaviour from optional configuration and sales-assisted commitments.
A filled-in entry might say: “Requirement: prompts stored and processed only in the contracted region. ChatGPT Business: storage choice may be offered at checkout, but inference is excluded and safety copies have separate handling. Result: not met on current public documentation; ask for a different plan or contract.” Another might say: “Requirement: no model training on customer content. Google Workspace: public Workspace privacy statement says no training without permission or instruction. Result: provisionally met, subject to our account and enabled-service review.” The first conclusion is narrower than “OpenAI is unsafe”; it is specific to one requirement and plan.
Keep the legal review proportionate. A business processing ordinary supplier copy may need a simpler assessment than one using sensitive client records, but both should document the decision. For personal data or regulated work, have a qualified data-protection adviser check the contract and transfer terms before live upload. After purchase, revisit the record when a new app is connected or a plan is changed. A procurement comparison expires when the actual data flow changes.
Further reads
- ChatGPT Business vs Claude Team: Seat Costs for a Small Consultancy — Compare the two direct team rivals.
- ChatGPT Business for Law Firms: Seats, Privacy Settings and Cost — See a more sensitive data case.
- ChatGPT Business for One Person: Plus, Pro or the Two-Seat Minimum? — Check the smallest workspace commitment.
- ChatGPT Business Seats: What You Pay Per Seat and How to Trim at Renewal — Model the seat commitment as well as data terms.
- 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 Business pricing and Where your ChatGPT Business content is stored; Mistral Vibe product page and pricing, data storage, training opt-out and zero data retention FAQs; Anthropic Claude pricing and organisation retention and server-location FAQs; Google Workspace pricing and Generative AI in Google Workspace Privacy Hub; Microsoft Learn Enterprise data protection in Microsoft Copilot and Copilot Chat, checked 5 November 2026.