No training on your company data: the business plans of ChatGPT, Claude and Copilot compared

As of 27 July 2026. All prices, training commitments and contract details in this article come from the providers’ pricing pages, help articles and legal documents, each with a retrieval date. This article is not legal advice; for specific questions, consult a data protection lawyer or the FDPIC, the Swiss federal data protection authority. Sources at the end.

Anyone introducing AI in a company sooner or later asks the same question: can customer data go in there? The answer depends less on the tool than on the contract behind it. The same models are available with or without training use, with or without a data processing agreement, and with or without European data residency - depending on how you buy them. This article sorts the tiers: what the business plans from OpenAI, Anthropic, Microsoft and Google cost, what they actually commit to - and when an API key or a Swiss cloud region is the better route.

The training line runs between personal and business accounts

At OpenAI and Anthropic, content from personal accounts may flow into model training - and the business tiers of all four major providers exclude that by default. Anyone working with a private login has a data protection problem regardless of how well-crafted the prompts are.

At OpenAI this concerns the Free, Go, Plus and Pro plans: “we may use your content to train our models”, OpenAI states in its own help centre; an opt-out is available via the privacy portal or the data controls (retrieved 27 July 2026). Two footnotes: temporary chats stay out of training, and conversations rated with thumbs-up or thumbs-down feedback can be used despite an opt-out.

Anthropic redrew its line in 2025: with the update to its consumer terms of 28 August 2025, users of the Free, Pro and Max plans had to decide by 8 October 2025 whether their chats may be used for training; for those who agreed, Anthropic announced a retention period extended to five years. The rule explicitly covers Claude Code when it runs on a personal account.

On the other side of the line the picture is uniform: ChatGPT Business and Enterprise, Claude Team and Enterprise, and Microsoft 365 Copilot do not use company data for training by default; Google commits to not using Workspace data for models outside Workspace without permission. An AI business plan is the contract tier of an AI provider at which company data does not flow into model training by default and a data processing agreement is part of the deal - the latter is what Article 9 of the Swiss Data Protection Act (revDSG) provides for commissioned processing, more on that below. These two commitments are what you buy with the tier upgrade - and they are what is missing when employees quietly work with personal accounts. The consequences are covered in ChatGPT in your company: 6 mistakes to avoid.

01 · Personal plan

Training possible

Content may go into training by default or after a pre-set choice. No data processing agreement.

02 · Business plan

No training, DPA included

ChatGPT Business from CHF 15, Claude Team from USD 20, Copilot CHF 24.30 per seat and month.

03 · API

No training, 30 days

Retention up to 30 days, zero data retention on request. Pay per use.

04 · Cloud route

Processing in Switzerland

Azure Switzerland North and Bedrock Zurich - at the price of older model generations.

What the business plans cost and commit to

Entry into a business tier starts at CHF 15 per user per month: that is ChatGPT Business in Switzerland with annual billing (CHF 19 monthly, minimum two seats; pricing page, retrieved 27 July 2026). What you buy is the same pair of commitments everywhere: no training on company data, plus a data processing agreement. ChatGPT Business was called ChatGPT Team until 29 August 2025 - if you search for the old name, you find the same product. Business and Enterprise are SOC 2 Type 2 audited; the Enterprise tier has no list price and runs through sales.

Claude Team costs USD 20 per seat per month with annual billing, USD 25 monthly; the premium seat is USD 100 or USD 125 (pricing page, retrieved 27 July 2026). Anthropic publishes US dollar prices only; there is no CHF price list. Team and Enterprise together form “Claude for Work” and are exempt from the consumer training rule; the data processing agreement is incorporated into the commercial terms by reference and needs no separate negotiation.

Microsoft 365 Copilot costs CHF 24.30 per user per month in Switzerland with annual billing, plus VAT (Swiss pricing page, retrieved 27 July 2026) - as an add-on to a qualifying Microsoft 365 plan, whose cost comes on top. Microsoft commits that prompts, responses and data accessed through Microsoft Graph are not used to train the underlying models. A cheaper side entrance: Copilot Chat is included at no extra cost for users with an Entra account and a qualifying Microsoft 365 subscription, with the same protection commitments (enterprise data protection), but without the Office integration of the paid licence.

Google solved the tier question differently: since 15 January 2025, Gemini has been included in the Business and Enterprise plans of Google Workspace, and the separate Gemini add-ons were discontinued. The list price for Business Standard in Switzerland is CHF 14 per user per month (pricing page, retrieved 27 July 2026; currently advertised at a discount). Google states that it does not use Workspace data to train models outside Workspace without permission.

Plan	Price per seat and month	Training on company data	DPA
ChatGPT Business	CHF 15 (annual) / CHF 19 (monthly), min. 2 seats	no, by default	yes
Claude Team	USD 20 (annual) / USD 25 (monthly)	no, by default	yes
Microsoft 365 Copilot	CHF 24.30 plus VAT (annual), plus M365 plan	no	yes
Gemini in Google Workspace	included in Business/Enterprise (Business Standard: list price CHF 14)	no, without permission	yes

All figures from the providers’ pricing pages and documentation, retrieved on 27 July 2026.

Data residency: where the plans hit their limits

If you require data to be stored in Europe or even in Switzerland, the subscription model runs out of road quickly - this is where Business and Enterprise tiers part ways. OpenAI offers data residency in the “Europe (EEA + Switzerland)” region only for ChatGPT Enterprise and Edu, at no extra cost but only for new customers; ChatGPT Business has no residency option (help article, retrieved 27 July 2026).

Anthropic offers no European data residency for the Claude apps first-party at all: the configuration knows only “US” and “global”. US-only inference on usage-based Enterprise plans carries a surcharge of factor 1.1 on token prices. European data residency for Claude models exists only via the API route through cloud platforms - more on that below.

Microsoft promises the EU Data Boundary for Copilot, which also applies to Swiss customers, and lists Switzerland as its own datacentre region with Advanced Data Residency including Copilot. The fine print sits in the documentation: data is stored in the chosen region, but the model calls themselves can go to other datacentres within the boundary - a Europe guarantee, not a Switzerland guarantee. The optional Anthropic models in Copilot are excluded from the EU Data Boundary and deactivated by default in Europe.

Details like these belong in the contract, not in trust - and they co-determine how easy a later provider switch will be. The contract side is covered in depth in Vendor lock-in with AI platforms.

The API route: no training, 30-day retention, zero on request

For technical teams, the API key is often the most direct route to reliable commitments: both major APIs do not use customer data for training by default - at OpenAI since 1 March 2023, at Anthropic per the commercial terms (“Anthropic may not train models on Customer Content from Services”).

On retention the two sit close together: OpenAI stores API inputs and outputs for up to 30 days for abuse detection, Anthropic deletes within 30 days. Both offer zero data retention - at OpenAI after approval for ZDR-eligible endpoints, at Anthropic by arrangement through the sales team. The exceptions sit in the small print: Anthropic’s frontier models Fable 5 and Mythos 5 mandate 30-day retention and are not available under ZDR; content flagged by safety systems can be retained by Anthropic for up to two years, even under ZDR. A data processing agreement is available on both platforms; for Swiss customers, OpenAI Ireland Ltd. is the contracting party.

OpenAI also runs a European endpoint: API data residency in the “Europe (EEA + Switzerland)” region via eu.api.openai.com, configurable per project - though only for new projects, after approval of modified abuse monitoring controls and with a modified retention amendment to the contract.

The flip side: an API key is not a workplace. It needs a front end - built in-house or bought from a third party whose data protection then matters too - plus someone to manage keys, costs and access. For a team that simply wants a chat workplace, the business plan is the shorter path.

Cloud route: Swiss processing, but older models

Processing in Switzerland is possible - but only with older model generations. With a regional deployment in Switzerland North, Azure OpenAI processes prompts and responses within the Swiss geography; currently available there are GPT-4.1, GPT-4.1-mini, GPT-4o plus embedding and Whisper models (Microsoft documentation, retrieved 27 July 2026). From a Swiss resource, the current GPT-5 family is available only as a global or EU data zone deployment - then processing leaves Switzerland. Customer data is not used for training here either without consent, and abuse monitoring can be modified on application for vetted customers.

At AWS the picture is similar: Amazon Bedrock has a Zurich region (eu-central-2), but only the older models Claude 3 Haiku and Claude 3.5 Sonnet run in-region there. Newer Claude models are available from Zurich only via geographic routing within Europe or globally. AWS and the model providers do not use Bedrock data for training.

Google drops out for this purpose: as of 27 July 2026, no generative Gemini model is available in the Vertex region Zurich (europe-west6), only embedding models - the nearest Gemini regions are in the EU.

The honest summary of the cloud route: if you need Swiss processing, you can have it - and you pay with one to two model generations of lag. Whether your use case tolerates that is a calibration question, not a question of principle.

What Swiss law requires - and why the DPF status matters

The Swiss Data Protection Act permits commissioned processing only on a contractual or statutory basis (Art. 9 revDSG) - in the relationship with an AI provider, that means in practice: a data processing agreement, and one only exists from the business tier upwards. OpenAI concludes it for ChatGPT Business, Enterprise and the API; the personal plans get none. In their business DPAs, the providers declare themselves processors.

For transfers to the US, certification status matters: the Swiss-U.S. Data Privacy Framework has been in force since 15 September 2024. I queried the participant list on 27 July 2026 - Microsoft and Google are certified with active status, OpenAI and Anthropic are not in the register. The latter two instead rest on standard contractual clauses contained in their respective DPAs; Anthropic’s includes a dedicated Swiss addendum for the revDSG. The FDPIC notes that the Swiss Data Protection Act is technology-neutral and therefore applies directly to AI-based data processing.

The full legal groundwork - which uses of AI trigger obligations under the revDSG, when a data protection impact assessment is due, and what the five core duties look like in practice - is in the companion article Using AI in Swiss SMEs - legally compliant and practical.

Frequently asked questions

Is ChatGPT Plus with training switched off enough for business use?

No. The training toggle solves only one of two problems: Plus remains a personal contract without a data processing agreement - the contractual basis Art. 9 revDSG provides for commissioned processing. Add the feedback exception: rated conversations can be used by OpenAI despite the opt-out. For tasks entirely without personal data, Plus is unproblematic.

What is the cheapest entry with solid commitments?

If Microsoft 365 is already in the house, you have it: Copilot Chat is included in qualifying plans at no extra cost, with enterprise data protection. Otherwise ChatGPT Business at CHF 15 per user per month (annual, minimum two seats) is the cheapest dedicated AI business plan - as of 27 July 2026. Google Workspace Business Standard lists at CHF 14 with Gemini included, but it is an office suite, not a pure AI subscription.

Does my data stay in Switzerland with a business plan?

No. The subscriptions offer Europe at best: OpenAI stores at rest in the EEA-plus-Switzerland region for Enterprise customers on request, Microsoft guarantees the EU Data Boundary - with processing possible anywhere within it. Processing guaranteed in Switzerland exists only via the cloud route, such as Azure Switzerland North with a regional deployment, and there only with older models.

Does the consumer training rule also apply to Claude Code?

Yes. Anthropic names Claude Code explicitly in the announcement of its updated consumer terms: if it runs on a Free, Pro or Max account, the account’s training choice applies. On Team and Enterprise plans and via the API, Claude Code is exempt from training use.

Practical takeaway

The tier question boils down to three checks. First: should company data reliably stay out of training? Every business tier delivers that, from CHF 15 per seat. Second: does personal data flow in? Then you need the data processing agreement that only the business tiers include. Third: does a customer or regulator require data residency in Europe or Switzerland? Then the path leads to the Enterprise tier, the API’s EU endpoint or a Swiss cloud region - with the model lag described above as the price. The reverse also holds: for a one-person company that never enters personal data, a personal plan with opt-out is defensible.

The next step is small: this week, count who in your company works with a personal account, and price the seats against the business tier. The difference is usually less than what a single data protection inquiry costs in effort. If you would rather not make the tier decision alone: in my team workshops, this calibration - what is enough, what is too much - is part of the programme.

---

Sources (checked 27 July 2026):

• What is ChatGPT Business? - OpenAI Help Center (updated ca. 24 July 2026)
• ChatGPT pricing - OpenAI (retrieved 27 July 2026)
• Enterprise privacy at OpenAI (updated 8 January 2026)
• How your data is used to improve model performance - OpenAI Help Center (updated ca. 13 July 2026)
• Data residency and inference residency for ChatGPT - OpenAI Help Center (updated ca. 14 July 2026)
• Introducing data residency in Europe - OpenAI (5 February 2025, updates through 16 January 2026)
• Data controls in the OpenAI platform - OpenAI Platform Docs (retrieved 27 July 2026)
• OpenAI Data Processing Addendum (effective 1 January 2026)
• Claude Pricing - Anthropic (retrieved 27 July 2026)
• Updates to Consumer Terms and Privacy Policy - Anthropic (28 August 2025)
• Is my data used for model training? - Anthropic Privacy Center (as of 16 March 2026)
• Anthropic Data Processing Addendum (effective 24 February 2025)
• API and data retention - Claude Platform Docs (retrieved 27 July 2026)
• Enable US-only inference for your organization - Claude Help Center (as of 15 June 2026)
• Microsoft 365 Copilot for business - Swiss pricing - Microsoft (retrieved 27 July 2026)
• Data, Privacy, and Security for Microsoft 365 Copilot - Microsoft Learn (9 July 2026)
• Enterprise data protection in Microsoft 365 Copilot - Microsoft Learn (29 May 2026)
• The future of AI-powered work for every business - Google Workspace Blog (15 January 2025)
• How Gemini in Google Workspace protects your data - Google Help (retrieved 27 July 2026)
• Region availability for Foundry Models sold by Azure - Microsoft Learn (24 July 2026)
• Data, privacy, and security for Foundry Models - Microsoft Learn (18 May 2026)
• Amazon Bedrock FAQs - AWS (retrieved 27 July 2026)
• Regional availability by models - Amazon Bedrock - AWS documentation (retrieved 27 July 2026)
• Deployments and endpoints (Locations) - Google Cloud Docs (24 July 2026)
• Federal Act on Data Protection, AS 2022 491 - Fedlex (25 September 2020, in force since 1 September 2023)
• Swiss-US Data Privacy Framework - FDPIC (14 August 2024)
• Data Privacy Framework - participant list (queried 27 July 2026)
• AI and data protection - FDPIC (as of 24 September 2025)