Not on the business plans, according to each vendor's own terms. Microsoft, OpenAI and Anthropic all say they do not train their models on customer business data by default. That promise covers training only. It does not cover who can reach your data inside your own firm, how long it is kept, or what staff paste into free accounts. This guide sets out what each vendor commits to, what Copilot costs in £, and the checks a UK firm should make before switching on an AI assistant.
What does Microsoft 365 Copilot promise about your data?
Microsoft says use of Microsoft Copilot and Copilot Chat by organisations is covered by its Data Protection Addendum and Product Terms, with Microsoft acting as a data processor. Its enterprise data protection page says prompts, responses and data accessed through Microsoft Graph are not used to train foundation models.
The same page says Copilot respects your identity model and permissions, inherits your sensitivity labels, applies your retention policies and supports audit of interactions. Web queries sent to Bing are a separate case with their own data handling practices, so check them separately if you allow web grounding.
The practical point is that Copilot works with what each user can already see. If a user has access to a sensitive folder they should not, Copilot can draw on it. Tidy permissions before you roll it out.
Do OpenAI and Anthropic train on business data?
OpenAI's enterprise privacy page, updated on 8 January 2026, says it does not train its models on business data by default for ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu and its API platform. It adds that if you explicitly opt in to share data, for example through feedback mechanisms, it may use that data. It also says customers of ChatGPT Enterprise, Healthcare and Edu control how long their data is retained, which is not stated for every plan.
Anthropic's Commercial Terms say the customer retains rights to its inputs and owns its outputs, and that Anthropic may not train models on Customer Content from the Services. These terms apply to commercial customers, not to consumer accounts.
Read the wording closely. "Not trained on" is a different commitment from "not stored", "not seen by staff for abuse checks" or "stays in the UK". Ask each vendor for the retention period, the sub-processors and the data location for the plan you intend to buy.
What does Copilot cost for a UK business?
Microsoft's UK Copilot business pricing page, checked on 2 November 2026, lists Microsoft 365 Business Standard with Copilot at £18.10 per user per month and Business Premium with Copilot at £24.60, both paid yearly and excluding VAT. The Copilot Business add-on for eligible existing Microsoft 365 business customers is listed from £13.80, down from £16.10.
These offers are time-limited and were checked on 2 November 2026, so confirm on the brand's own site before you buy. Microsoft says the discount runs from 1 July 2026 to 31 December 2026 and applies to the first year only. An annual commitment is required, and some advanced AI features are billed on usage.
Microsoft Copilot UK business prices per user per month
Published UK prices paid yearly, excluding VAT, for plans that include Copilot, with the add-on shown at its offer and earlier price.
View the data
| Category | £ per user per month excluding VAT |
|---|---|
| Add-on offer price | £13.80 |
| Add-on earlier price | £16.10 |
| Business Standard with Copilot | £18.10 |
| Business Premium with Copilot | £24.60 |
What is an unmanaged entry point for AI tools?
It is not a formal standard term, so treat this as our working definition. An unmanaged entry point is any route by which an AI tool can reach company data that IT has not approved, set up or monitored. Examples are a personal chatbot account, a browser extension, a script with an API key, or a connector that has been granted access to shared files.
To find them, list the AI tools in use, then check single sign-on records, application consent grants, browser and device inventories, and the admin consoles of the tools you have approved. For each connector, ask what it can read, whether it can write or send, and whose identity it uses.
Are there differences in data privacy between Agentforce and Copilot?
Yes, in where the data sits and who configures access. Salesforce's Agentforce privacy FAQ says Customer Data is not stored in Agentforce itself, because it is the agentic layer over underlying Salesforce services, where the data is stored. It says Salesforce has a zero data retention policy so that third-party language models do not retain Customer Data or use it for training, and that it does not currently use Customer Data to train generative AI models.
Copilot instead works across Microsoft Graph and the permissions each user holds. So the checks differ. For Agentforce, review what each agent is configured to do and which Salesforce data it can reach. For Copilot, review user permissions and sensitivity labels. Neither document proves a product is safer in your set-up, so test the real configuration.
Which enterprise AI tools keep prompts out of model training?
Every vendor covered here says its business offering does not train on your content, and the wording differs by product. Microsoft, OpenAI, Anthropic and Salesforce are covered above. For intranet software, Staffbase's AI FAQ says customer content is not used to train Staffbase or third-party AI models. It also says it uses a self-run model and OpenAI Services hosted on Microsoft Azure.
If "no training on our data" is a hard requirement, shortlist suppliers whose contract says so, and ask for the product-specific terms in writing. A web FAQ is a good lead, but it is not the contract.
Can you keep AI and collaboration data in the EU or Asia-Pacific?
Sometimes. Atlassian's data residency documentation lists an EU location made up of Frankfurt and Dublin, a Germany location of Frankfurt, and an Australia location of Sydney. It also lists data that cannot be pinned, including app analytics, cached content and data in transit for up to 30 days. Ask any vendor to map content, metadata, logs, backups and AI processing separately, because "regional hosting" rarely covers all of them.
What does the NCSC say about sensitive information and public AI tools?
The National Cyber Security Centre's blog on ChatGPT and large language models says the provider and its partners may be able to read queries, and advises you to understand the terms of use and privacy policy thoroughly before asking sensitive questions. It recommends not including sensitive information in queries to public LLMs, and not submitting queries that would cause problems if they were made public.
That advice is why an approved business plan matters. It moves staff from a public account, where you rely on the terms of a consumer service, to a contract that you have read and can enforce.
What does the ICO expect from UK firms using generative AI?
The ICO's executive summary of its generative AI consultation response covers lawful basis, purpose limitation, accuracy, individual rights and allocating controllership across the AI supply chain. It says it retained its position on purpose limitation, accuracy and controllership.
For a firm buying an assistant, that means deciding and recording who is controller and who is processor, what the tool is for, and how people can exercise their rights over data that passes through it.
How do you choose a safe AI assistant for regulated work?
Work through a short checklist before any rollout:
- Classify the data staff may use, and ban personal, special category and client data in consumer accounts.
- Get the training, retention and sub-processor terms in the signed contract, not only on a web page.
- Review permissions and sensitivity labels so the assistant cannot surface files a user should not see.
- Keep a person approving any consequential action the assistant takes.
- Keep audit logs and review them.
Run a small pilot with named owners, then widen it once the checks hold.
What changes next?
Vendors change terms and prices often. Microsoft's add-on discount ends on 31 December 2026, and OpenAI's page was last updated in January 2026. Put a diary date in to re-read the data terms of each tool before every renewal.
Sources
- Microsoft, Enterprise data protection in Microsoft Copilot, 2026
- Microsoft UK, Copilot business pricing, checked 2 November 2026
- OpenAI, Enterprise privacy at OpenAI, January 2026
- Anthropic, Commercial Terms of Service
- Salesforce, Agentforce privacy FAQ
- Staffbase, AI in the Staffbase platform FAQ
- Atlassian, Understand data residency
- NCSC, ChatGPT and large language models, what's the risk
- ICO, Response to the consultation series on generative AI, executive summary
