AI Data Governance: Which Data Goes Into Claude, ChatGPT and Copilot, on Which Terms, and Who Owns It
AI data governance is how a company decides what data may go into which AI tools, on which terms, and with what record. For most companies using Claude, ChatGPT and Copilot it comes down to six things: a rule for which data goes into which tool, business plans with the right training and retention terms, controlled access to company systems through connectors, a decision on where data is stored, logs, and a named owner.
It is not a new discipline. It is your existing data protection work, applied to a tool that employees use dozens of times a day and that can now read your systems directly.
What AI data governance means in practice
Data governance has always meant knowing what data you hold, who may use it, for what, and being able to show that. AI adds two things. People now send company data to an outside service many times a day, often without thinking of it as sending data. And with connectors, the AI reaches into the CRM, inbox and file store itself, so the question is no longer only what people paste in.
Most companies we meet have one of two problems. Either there is no rule, so careful people avoid AI for anything real and careless people use personal accounts. Or there is a long policy nobody has read. The fix for both is short, concrete rules and a setup where the safe route is also the easy one.
Which data may go into which tool
Start with one table. Our AI policy template uses three levels of tools and three levels of information, and the table that combines them is the part people remember:
| Open | Internal | Confidential | |
|---|---|---|---|
| Level 1: official tools, company accounts, connected to approved systems | Yes | Yes | Yes, through approved connectors |
| Level 2: tools approved for a team, not connected | Yes | Yes, company account | No |
| Level 3: everything else, private use | Yes | No | No |
Give concrete examples for each information level, because “confidential” means different things to different people. Open is anything already published. Internal is ordinary working material: meeting notes, plans, drafts. Confidential is client data, personal data about employees, contracts, unpublished financials. Then add the few things that never go into any AI tool at your company, if there are any, and say why.
Training and retention terms by plan
The biggest single factor is the account, not the product. On the business plans of all three main assistants, your content is not used to train the vendor’s models. On personal plans the terms are different, and the company has no control. That is why the first job is making sure people use the company account.
| Claude Team and Enterprise | ChatGPT Business and Enterprise | Microsoft 365 Copilot (work account) | |
|---|---|---|---|
| Used for training? | No, on commercial terms (Anthropic) | No, by default (OpenAI) | No (Microsoft) |
| How long chats are kept | Until deleted; Enterprise can set its own retention period | Until deleted, then removed within 30 days; Enterprise can set its own period | Retention policies set in Microsoft Purview |
| Where data is stored | United States; no EU-only storage for the Claude apps | Europe available on Enterprise, rolling out on Business (OpenAI) | Within the EU Data Boundary for EU and EFTA customers, with limited exceptions |
| Personal plans | Free, Pro and Max: each person chooses | Free, Plus and Pro: used for training unless turned off | Consumer Copilot: used unless turned off |
The detail, with sources and the caveats, is in our plan-by-plan guides: Claude, ChatGPT and Microsoft Copilot. Terms change; check them when you renew. If you are still choosing a plan, see Claude Team vs Enterprise and ChatGPT Business vs Enterprise.
Ask everyone to check which account they are signed into. Expect a surprise: someone on a personal Plus or Pro plan doing client work, often because it was faster than waiting for a seat. That is shadow AI, and the answer is giving them the company account, not a warning.
Access to company systems through connectors
Connectors change the data question. When Claude, ChatGPT or Copilot is connected to your CRM, finance system or file store through MCP, the AI reads the data directly. The information table above still applies, but it is no longer enforced by what people choose to paste. It is enforced by what each connection can reach.
Each connection needs four answers:
- Whose login? The person’s own, so the AI sees only what they may see, or a shared account that sees everything.
- Read or write? Read-only as the default.
- Which groups? Finance data for Finance, not for everyone.
- Where is it logged? So you can show what was read.
Watch for permissions that were too broad before AI. Copilot shows people what they are already allowed to open, which makes files shared too widely years ago easy to find. The same is true of any connector that uses a person’s own login. Fix oversharing before a wide rollout. Our guides to MCP governance and MCP security cover connectors in more depth, and what an MCP gateway is explains the layer that sets these rules once for every system.
Data residency, GDPR and UK GDPR
If you have employees or clients in the EU or the UK, GDPR or UK GDPR applies to the personal data that passes through AI tools. The obligations are the ones you already know; AI does not create a separate regime. The practical points:
| Requirement | What it means for AI tools |
|---|---|
| A processor agreement (Article 28) | A data processing agreement with each AI vendor, on the business plan. Personal accounts have none. |
| Records of processing (Article 30) | Add the AI tools and connectors to your records, with what personal data they handle. |
| Impact assessment (Article 35) | Where processing is likely to be high risk, a DPIA. The UK ICO expects one in most cases where AI processes personal data. |
| International transfers | Where data leaves the EU or UK, a transfer mechanism such as standard contractual clauses or the UK addendum. |
| Data minimisation | Connectors and groups set so the AI reaches only the personal data a task needs. |
| People’s rights | Being able to find and delete a person’s data includes chats and logs. |
Data residency is a business decision as much as a legal one. Transfer mechanisms make US storage lawful for most companies, but some sectors and clients require data to stay in Europe. Decide which you are before choosing a plan, because it differs by vendor, as the plan table shows. Remember the connectors and gateway too: where they run and where their logs are stored is part of the same question. The ICO’s guidance on AI and data protection is the UK reference; in the EU, the AI Act has required providers and deployers of AI systems to take measures for staff AI literacy since February 2025.
Logs: what you can show afterwards
Governance you cannot show is a promise. There are two kinds of log, and you want both.
Logs inside each assistant. What people asked and what the AI answered. Copilot logs prompts and responses for audit and eDiscovery through Microsoft Purview. ChatGPT Enterprise offers audit logs through its Compliance API, and Claude Enterprise has audit logs and a Compliance API that Team does not.
Logs of what the AI touched in your systems. Which person, through which assistant, read or changed what in which system. The assistants’ own logs rarely show this across systems, and each system’s own log shows only what it saw, often one shared account. A gateway that every connector passes through gives one log for all of them. Decide how long logs are kept and who may read them: they contain personal data too.
Who owns AI data governance
One person owns the whole: usually the COO, CIO or head of IT. They are supported by whoever handles data protection, the owners of the main systems, and security. The owner keeps the policy, the approved tools and connectors, and answers “may I put this into AI?” in one message. Without that short route, people decide for themselves.
A two-week start
- Write the three-by-three table: tools and information, with examples. Our AI policy page has a template.
- Move everyone onto company accounts on a business plan, and check who is still on a personal one.
- Confirm a data processing agreement with each AI vendor, and add the tools to your records of processing.
- Decide whether data must stay in Europe, and check each vendor against it.
- List every connector, whose login it uses and whether it can write. Make read-only the default.
- Turn on the logs your plans offer, and decide where the log of system access lives.
- Name the owner and the route for questions.
Where Deployed OS fits
Deployed OS handles the connector side of AI data governance. Your systems are connected once and used from Claude, ChatGPT and Copilot. Each person signs in with their own login where the system allows it, user groups decide who may read or write in which system, and every call is logged and can be exported. Guardrails, such as your rule on confidential information, come with every question. It is built by an EU company, and you choose where the data is stored.
If you want your data rules to apply in every assistant your team uses, rather than in a document, we set up Deployed OS on your own systems with your team. See every connector.
Frequently asked questions
What is AI data governance?
AI data governance is how a company decides what data may go into which AI tools, on which terms, and with what record. For companies using Claude, ChatGPT and Copilot it covers which information goes into which tool, the training and retention terms of each plan, what the AI can reach through connectors, where data is stored, logs and who owns the rules.
Do Claude, ChatGPT and Copilot train on company data?
Not on their business plans. Claude Team and Enterprise, ChatGPT Business and Enterprise and Microsoft 365 Copilot on a work account do not use your content to train the vendor’s models by default. Personal plans are different and the company has no control over them, so the first step is making sure people use the company account.
What data should not go into AI tools?
It depends on the tool, not only the data. A simple rule uses three levels of information, open, internal and confidential, and three levels of tools. Confidential information goes only into official company tools, ideally through approved connectors, while private accounts get open information only.
Is using AI tools compliant with GDPR and UK GDPR?
It can be, on business plans with the usual obligations met: a data processing agreement with each vendor, the tools added to your records of processing, an impact assessment where processing is high risk, a transfer mechanism where data leaves the EU or UK, and access limited to the personal data a task needs. Personal accounts have no processor agreement with your company.
Where is AI data stored?
It differs by vendor. Claude business data is stored in the United States, with no EU-only storage for the Claude apps. ChatGPT offers storage in Europe on Enterprise and is rolling it out on Business. Microsoft 365 Copilot is within the EU Data Boundary for EU and EFTA customers, with limited exceptions.
Who should own AI data governance?
One named person, usually the COO, CIO or head of IT, supported by whoever handles data protection, the owners of the main systems and security. The owner keeps the policy, the approved tools and connectors, and gives a quick answer when someone asks whether they may put something into AI.
Found this useful? Send it to someone who needs it.