Where Is Your Data Stored When You Use AI Tools?

On the AI vendor's servers, not your own computer. Chats, uploaded files and saved memories sit in the provider's cloud, and extra copies can exist in safety logs, backups and any app the tool connects to. Business plans from OpenAI, Microsoft and Google let you check or choose the storage region; free and personal plans don't.

Storage and processing are separate questions, and vendors treat them separately. OpenAI's help pages say a ChatGPT Business workspace keeps its content at rest in the region chosen at checkout, yet that choice doesn't decide where replies are generated, and abuse-monitoring logs stay in the United States whichever region you pick. So "our data is stored in X" is only ever part of the answer.

Follow me on Instagram@sagnikteaches

Six places a single prompt can end up

Take one ordinary request. A removals office manager pastes a customer's inventory list into an AI chat and asks for a quote email. That text can now exist in six places, and each has its own rules.

Connect on LinkedInSagnik Bhattacharya

1. Your chat history. The main copy, kept at rest in the vendor's cloud so you can scroll back to it. This is the copy a "data residency" option controls. On business plans it sits alongside the rest of your workspace content; on personal plans it sits wherever the vendor's standard setup puts it.
2. The processing location. The chips that generate the reply can be in a different region from the storage. Microsoft says Copilot's model calls go to the nearest data centres but can be sent to other regions when demand is high. Anthropic routes Claude processing to any available region unless an organisation pays for its US-only inference setting.
3. Safety and abuse-monitoring logs. Separate copies kept for a limited time to catch misuse. On OpenAI's API these logs are kept for up to 30 days by default. They are the reason zero data retention exists as a separate arrangement that only some customers can get.
4. Backups and replicas. Usually in the same region as the main copy when residency is switched on. OpenAI lists backups and replicas as in scope for ChatGPT Enterprise residency, which is the kind of detail worth looking for in any vendor's documentation.
5. Connected apps, web search and other model suppliers. When the assistant searches the web, reads a shared drive through a connector or hands work to a third-party model, that service handles the data under its own terms and in its own regions. OpenAI's residency article lists apps, MCP servers and web search as out of scope. Microsoft notes that Anthropic models offered inside Copilot currently sit outside its EU Data Boundary.
6. Your own devices. Downloaded exports, files the AI produced, desktop app caches and screenshots on staff phones. It is the least glamorous copy and often the least protected.

Whether the vendor trains its models on your data is a different question again. A vendor can store chats in your chosen region and, on a consumer plan, still use them for training unless the switch in privacy settings is off. If that is your worry, read how training works on business and free ChatGPT plans before you spend time on regions.

Subscribe on YouTube@codingliquids

Region controls by plan, as the vendors describe them

The table below summarises each vendor's own documentation as of September 2026. These pages change often, so treat it as a map of what to check rather than a permanent answer.

Two patterns stand out. Personal plans give you no say at all, and even the best business options control storage more tightly than processing. If a contract demands both, you are usually looking at an enterprise plan or an API route, not a $20 subscription.

A 14-person removals firm maps its AI data trail

The firm in this example is illustrative: three office staff, eleven people on the vans and roughly 60 moves a month. Its AI use grew piecemeal over a year:

• Microsoft 365 Business Standard for all 14 staff, with the Copilot Chat that comes with it.
• ChatGPT Business with three Standard seats at $25 each billed monthly, so $75 a month, used by the office for quotes and customer emails.
• A video-survey app that turns a customer's phone walk-through into an itemised inventory, with AI built in.
• A website chat widget that answers questions about packing, storage and moving dates.

The office manager spent about three hours filling in one row per tool: what goes in, where it is stored, where it is processed, who else gets a copy and how long it is kept. What she found:

• Copilot Chat turned out to be the easy one. The Data Location Card in the Microsoft 365 admin centre showed one committed location for the firm's email and files. Copilot Chat isn't listed on the card, but Microsoft's documentation says it is covered by the same data location commitments, so the honest answer was "wherever our email lives".
• ChatGPT Business showed no region anywhere she could find. OpenAI's help pages explain why: Business customers pick a region at checkout and the option is still rolling out, so this workspace follows OpenAI's default setup.
• The video-survey app said "hosted on a major cloud provider" and nothing more. Its subprocessor list named an AI model provider, and videos were kept for 24 months by default.
• The chat widget kept transcripts until someone deleted them, and the vendor confirmed each message goes to an AI model API whose abuse-monitoring logs keep content for up to 30 days.

One sum changed the priorities. Sixty moves a month is 720 home walk-through videos a year, and at 24 months' retention the survey app was holding about 1,440 videos of the insides of customers' homes at any one time. Cutting retention to 90 days after the job closes, while keeping the typed inventory in the firm's own job system, brings that down to roughly 180. That one setting mattered more than any region question.

The firm made four changes: shorter video retention, a rule that storage-unit codes and key-safe numbers never go into any AI tool, a written request to the chat-widget vendor for its storage region and subprocessors, and a line in the privacy notice naming the four tools. Total time was about four hours. New spend was zero.

When the storage region matters, and when it barely does

Location becomes a real requirement in a handful of situations. Everywhere else, what you put in and how long it's kept matter far more.

• A client contract names a place. An HVAC installer maintaining plant rooms for a hospital or a school may sign supplier terms that say where the client's data must be held. Once that clause exists, either use a plan with a region control and written confirmation, or keep that client's drawings and access details out of AI tools entirely.
• Your privacy notice promises a location. If your website says customer data stays in a particular region and your AI chat widget stores transcripts somewhere else, the notice is now wrong. Fix one or the other.
• Transfer rules in data-protection law. Rules such as the GDPR's cover moving personal data between regions, and they are usually satisfied by the safeguards in the vendor's contract rather than by the region itself. How the big assistants handle that is covered in whether ChatGPT, Claude, Gemini and Copilot are GDPR-compliant.
• Security details are involved. A locksmith's job notes might hold alarm codes, key numbers and "customer away until the 14th". For that data the question isn't which region it's stored in; the answer is that it shouldn't go into a chat tool at all.

Two quick contrasts. A cleaning company that holds keys for 40 clients can use AI freely for rota messages and quotes, as long as key tags, alarm codes and entry instructions stay in its job system. A landscaper drafting planting plans and seasonal newsletters has almost nothing location-sensitive to worry about. The rule of thumb: if a contract, a regulator or your own privacy notice names a place, you need a region control and it in writing; if none of them does, spend the effort on inputs and retention instead.

For the most sensitive material, there is also the option of not sending it to anyone's cloud. The trade-offs are set out in cloud AI versus on-device AI for business data.

Checking the region on your own account

Each major vendor exposes the setting in a different corner. These are the routes their documentation gives:

• ChatGPT Enterprise or Edu: open your profile menu, choose Workspace settings, then General, and look for the Data Residency Region field. Owners and admins can see it; it displays the region but isn't a control for changing it. For Business, the region is chosen at checkout, and OpenAI's page on where ChatGPT Business content is stored spells out what that covers.
• Microsoft 365: in the admin centre go to Settings, then Org settings, open the organisation profile tab and select Data location. The card shows a Current Geography and a Committed Geography. "No Commitment" means Microsoft stores the data where it best serves the service and can change that without notice, as its Data Location Card documentation explains. Microsoft describes the card as a tool for Global Admins, so ask whoever holds that role.
• Google Workspace: in the Admin console go to Menu, Data, Compliance, Data regions. The choices are No preference, United States or Europe, on Business Standard and higher editions.
• Claude: Team plans have no region setting. Usage-based Enterprise organisations can switch on US-only inference under organisation settings, then Data and privacy. On the API, each workspace's storage geography is set in the Console when the workspace is created and can't be changed later.

One shortcut that doesn't work is asking the assistant itself. A typical exchange looks like this (illustrative):

Prompt:  Where do you store the conversations I have with you?

Reply:   I don't store your personal data. Each conversation is
         processed in real time and isn't retained after our
         chat ends.

That reply is describing a generic chatbot from its training data, not your account. Your history is plainly stored, because you can scroll back through it tomorrow. The model has no view of your workspace's region, retention or backup settings, so only the vendor's documentation and your admin screens count as evidence.

An email that gets vendors to name their regions

For AI features inside other software, such as a scheduling app, a quoting tool or a phone-answering service, the answer rarely sits on a settings screen. Ask in writing, with questions narrow enough that a vague reply is obvious.

Subject: Data location questions about [product]

Hi [first name],

We use [product] for [job, e.g. customer quotes]. Before we put
more customer information into it, could you confirm in writing:

1. Where is our data stored at rest (region), including backups?
2. Where is it processed when the AI generates a response?
3. Which AI model providers or other subprocessors receive our
   data, and in which regions?
4. How long do you and each provider keep prompts, outputs
   and logs?
5. Can we choose or restrict the region, and at what cost?
6. Where is this documented (trust page, DPA, subprocessor list)?

Thanks,
[name]

Here is the kind of reply that comes back more often than it should (illustrative): "Your data is protected with bank-grade encryption and stored securely in the cloud. We partner with industry-leading AI providers and never sell customer data." Read it question by question and it answers none of them. Encryption says nothing about location. "The cloud" is a building somewhere specific. "Industry-leading providers" should have names, and "never sell" answers a question you didn't ask. A polite follow-up does the job: "Thanks. Could you name the storage region and the model providers, and send the link to your subprocessor list?" Vendors that take data protection seriously answer that within a day or two. The contract side of the same questions is covered in what to check in an AI vendor's data processing agreement.

Gaps that make a residency promise weaker than it sounds

Even a genuine region commitment has edges. These are the ones that catch small teams out:

• Web search and connectors. Ask an assistant to look something up online or read a document through a connected app, and that part of the job runs under the other service's rules. A cleaning company's office manager asking an enterprise chat tool to check a client's building access policy online has just sent that query outside the residency promise.
• Busy-period routing. Microsoft is candid that model calls can go to other regions when capacity is short. Processing guarantees are rarer and narrower than storage guarantees.
• Other companies' models inside a suite. Copilot can offer Anthropic models, which Microsoft currently excludes from its EU Data Boundary. Admins decide whether to allow them, so check that setting if the boundary matters to you.
• Caches. Google says your data region settings don't apply to data processed and cached by Gemini Notebook, even though Drive files you add to it are otherwise covered.
• Timing. ChatGPT Enterprise residency applies to workspaces set up with it, and OpenAI says new features aren't residency-eligible until it adds them to its list. A tool your team adopted last month may not be covered yet.
• Safety copies. On ChatGPT Business, choosing a region outside the United States still leaves a temporary copy of every prompt and response there for abuse monitoring.

API users have one more lever with a visible price. Anthropic charges 1.1 times the normal rate for US-only processing on its newer models, so a quote-drafting automation costing $40 a month in Claude API usage would cost about $44 with the setting on. Small in money terms, but it shows that processing location is a feature vendors price separately.

A one-page AI data map, filled in

Keep this map somewhere the whole team can see it, and review it whenever you add a tool or a vendor changes its terms. The first two rows are the removals firm's; the third is blank for your next tool.

AI DATA MAP                               Reviewed: [date]

Tool: ChatGPT Business (3 seats)
  Data going in:      quote details, customer names, move dates
  Never goes in:      storage-unit codes, key-safe numbers
  Stored at rest:     OpenAI default (no region chosen at checkout)
  Processed:          OpenAI, region not guaranteed
  Other copies:       abuse-monitoring logs; exports on office PCs
  Kept for:           until deleted by the user
  Who can see it:     each user's own chats; shared projects
  Checked against:    OpenAI help pages, [date]

Tool: Video-survey app
  Data going in:      walk-through videos of customers' homes
  Stored at rest:     vendor's cloud, region requested [date]
  Processed:          named AI model provider (subprocessor list)
  Kept for:           90 days after job closes (was 24 months)
  Who can see it:     office team, survey vendor support
  Checked against:    vendor email reply, [date]

Tool: ______________________
  Data going in:
  Never goes in:
  Stored at rest:
  Processed:
  Other copies:
  Kept for:
  Who can see it:
  Checked against:

A map like this takes an afternoon for most small teams, and it answers the question customers and insurers increasingly ask: where does my information go when you use AI? When you can answer that in a sentence per tool, you have done the part that matters.

More questions about where AI data lives

Does choosing a storage region cost extra?

It depends on the vendor. OpenAI includes data residency at no additional cost on ChatGPT Enterprise and Edu, and Business customers choose a region at checkout. Anthropic bills US-only inference at 1.1 times normal usage rates on newer models. Microsoft sells Advanced Data Residency as a separate licence. Ask for the price in writing before you plan around a region.

If I delete a chat, is it deleted everywhere?

Not immediately. Deleting removes the chat from your history at once, but back-end copies follow the vendor's schedule. Anthropic, for example, says deleted conversations leave its back-end storage within 30 days. Safety logs, backups and anything already passed to a connected app have their own timelines, so check each vendor's retention page.

Is it illegal for AI data to be stored in another country?

Usually not in itself. Data-protection law such as the GDPR allows transfers when safeguards are in place, typically contract clauses in the vendor's data processing agreement. Trouble starts when a client contract, a sector rule or your own privacy notice promises a specific location. If you are unsure, ask your data-protection adviser before you sign up.

Can I keep AI data entirely on my own computer?

For some jobs, yes. On-device features and small local models can process text without sending it to a vendor's cloud, although they are less capable and you then own the security and backups. Many small businesses use a mix: local tools for the most sensitive material and business-plan cloud tools for everything else.

Further reads

• Do You Need a DPIA Before Using AI Tools? — Decide whether your AI use needs a formal risk assessment first.
• Does Microsoft 365 Copilot Keep Your Business Data Private? — How Copilot handles business data beyond its storage location.
• Is Gemini Safe for Confidential Business Data in Workspace? — Gemini in Workspace, judged on confidential business data.
• SOC 2 and ISO 27001 Explained: Checking an AI Vendor's Security — Read a vendor's security reports without a compliance background.
• Who Owns AI Chat History When an Employee Leaves? — What happens to stored chats when someone leaves your team.
• What If Your AI Vendor Shuts Down? Checks Before You Commit — Stored data is only safe while the vendor is still trading.
• What to Check in an AI Tool's Privacy Policy and Terms — Seven checks for any AI tool's privacy policy and terms, the words to search for, and how ChatGPT, Claude, Gemini and Copilot compare.
• AI Tool Approval Process: How Staff Request a New AI Tool — The request form staff fill in, the checklist the reviewer works through, and the three replies, so new AI tools get approved in days, not months.
• n8n Self-Hosted vs Cloud: Real Costs for a Small Business — Both price lists side by side, the upkeep hours self-hosting really takes, and a roofing contractor's year costed both ways.
• How to Keep Customer Data Private When Your Team Uses AI — Eight steps to keep customer details out of the wrong AI tools, with a filled-in data table, redaction examples, policy wording and a quarterly audit.
• AI Security Checklist Before Connecting Tools to Email and Files — Twenty checks to run before an AI tool touches your inbox or shared drive, from reading the permission screen to rehearsing how to disconnect it.
• Business Data Backup Checklist Before You Connect AI Tools — A grouped checklist of what to copy before an AI tool gets access to your files, spreadsheets, CRM or accounts, with how to prove each backup restores.
• Are AI Meeting Note-Takers Safe for Client Calls? — When AI note-takers are safe on client calls: consent wording, the auto-join and auto-share settings that cause leaks, retention, and calls to never record.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help pages (Where your ChatGPT Business content is stored; Data residency and inference residency for ChatGPT) and API data controls guide; Claude API docs (Data residency); Claude help pages on US-only inference for Enterprise; Anthropic privacy pages on organisation data retention; Microsoft Learn (Data, Privacy, and Security for Microsoft Copilot; Data Location Card; Data Residency for Microsoft 365 Copilot); Google Workspace Help (Data covered by data regions; Choose a geographic location for your data; Gemini Notebook admin settings). Checked September 2026.