Direct answer: only in an approved business workspace, and only for the kinds of guest data you have decided it may hold. The same assistant behaves very differently depending on the account. On OpenAI’s individual plans (Free, Plus, Pro), “we may use your content to train our models” unless the user opts out (OpenAI: how your data is used). Google tells consumer Gemini users that human reviewers see some chats and asks them not to “enter confidential information that you wouldn’t want a reviewer to see” (Gemini Apps Privacy Hub). On business plans the default flips: OpenAI says “we do not use your business data for training our models” and offers a Data Processing Addendum (OpenAI enterprise privacy), and Microsoft says organisational use of Copilot is covered by its Data Protection Addendum “with Microsoft acting as a data processor” (Microsoft: enterprise data protection in Copilot). Under UK GDPR the hotel, villa company, tour operator or DMC remains the controller of its guests’ data either way, and is the one that has to show the tool was a sensible place to put it.
Kaize is a UK AI product studio for travel and hospitality. This article is based on OpenAI, Google, Microsoft and Anthropic product and privacy documentation, UK GDPR as published on legislation.gov.uk and Information Commissioner’s Office (ICO) guidance, all read on 17 October 2026. Vendor terms change often, so check the current versions before relying on them. It is not legal advice or a case study; the controls below are our recommendations.
Why this question is live now
In most travel businesses, staff did not wait for an AI policy. Reservations agents paste a guest’s complaint into a chatbot to draft a calmer reply; sales teams drop a group rooming list into one to reformat it; operations staff summarise a supplier email chain that includes passenger names, flight details and dietary notes. Much of this happens on personal accounts, because they are free, quick and already on people’s phones.
Meanwhile the vendor side has moved. OpenAI introduced UK data residency for its API Platform, ChatGPT Enterprise and ChatGPT Edu from 24 October 2025 (OpenAI: the next chapter for UK sovereign AI). Anthropic changed its consumer terms in 2025 so that Claude Free, Pro and Max users choose whether their chats train its models (Anthropic: updates to consumer terms). Google updated its Workspace generative AI privacy hub on 6 October 2026 (Generative AI in Google Workspace Privacy Hub). The tools are now good enough, and the business terms clear enough, that “just ban it” is no longer the only defensible answer — but the distinction between personal and business accounts has become the whole question.
Personal accounts versus business workspaces: what the vendors say
These are the vendors’ own statements, quoted from their current documentation. They describe defaults and commitments, not a guarantee that a given configuration is compliant.
OpenAI (ChatGPT)
- Individual plans: content may be used to train models; users can opt out by turning off “Improve the model for everyone”. Even after opting out, if a user gives thumbs-up or thumbs-down feedback, “the entire conversation associated with that feedback may be used to train our models” (OpenAI: how your data is used).
- Business plans (ChatGPT Business, Enterprise, Edu) and the API: not used for training by default; for Business, Enterprise and Edu, response feedback “does not permit training on the associated conversation” (OpenAI: how your data is used). OpenAI offers a DPA for ChatGPT Business, Enterprise and the API, and Enterprise admins control retention, with deleted conversations “removed from our systems within 30 days, unless we are legally required to retain them” (OpenAI enterprise privacy).
Google (Gemini)
- Consumer Gemini Apps: “Human reviewers (including trained reviewers from our service providers) review some of the data we collect”; with Keep Activity off, chats are still “retained with your account for 72 hours” (Gemini Apps Privacy Hub).
- Google Workspace: the Gemini app is governed by the organisation’s Workspace agreement, “including the Cloud Data Processing Addendum”, and chats and uploaded files “won’t be reviewed by human reviewers or otherwise used to train generative AI models outside of your domain without your permission”. Retention for Gemini app prompts and responses is “up to 36 months, as determined by admins” (Generative AI in Google Workspace Privacy Hub).
Microsoft (Copilot)
- Organisational Copilot and Copilot Chat: covered by the Microsoft DPA and Product Terms with Microsoft as processor; prompts, responses and Microsoft Graph data “aren’t used to train foundation models” (Microsoft: enterprise data protection).
- Two exceptions worth knowing. Web search queries sent to Bing are handled under different terms, where “Microsoft acts as an independent data controller”; and the EU Data Boundary does not apply to web queries, while “Anthropic models are currently excluded from the EU Data Boundary” (Microsoft: enterprise data protection). Microsoft also says Copilot calls to the model “can call into other regions where capacity is available during high utilization periods” (Microsoft Copilot privacy).
- Copilot “only surfaces organizational data to which individual users have at least view permissions” (Microsoft Copilot privacy). If a shared drive holding passport scans is open to all staff, Copilot will find it for all staff.
Anthropic (Claude)
- Consumer plans (Free, Pro, Max): users choose whether chats are used to train models; if they allow it, retention extends to five years, otherwise the existing 30-day retention applies. These terms “do not apply to services under our Commercial Terms”, including Claude for Work (Team and Enterprise) and the API (Anthropic: updates to consumer terms).
What UK GDPR asks of the travel business
Choosing a processor. Article 28 says a controller “shall use only processors providing sufficient guarantees” and requires a contract under which the processor “processes the personal data only on documented instructions from the controller” (UK GDPR Article 28). A business workspace with a signed DPA is built for that. A staff member’s personal account is not: there is no contract between your company and the vendor, and the vendor’s own consumer terms decide what happens to the data. The ICO’s controllers and processors guidance is the place to work through the roles, and the ICO notes it is under review following the Data (Use and Access) Act (ICO: controllers and processors).
Doing the assessment before, not after. The ICO says that when you buy an AI solution from a third party, “you are also required to specify your requirements at the procurement stage, rather than addressing trade-offs afterwards”, and that if you cannot assess whether a third-party solution would be compliant, “you should, as a matter of good practice, opt for a different solution” (ICO: accountability and governance of AI). Article 35 requires a data protection impact assessment where processing “in particular using new technologies” is likely to result in high risk (UK GDPR Article 35). Each use — drafting replies, summarising complaints, reformatting rooming lists — is a distinct processing operation that needs its own purpose and lawful basis (ICO: lawfulness in AI).
Special category data is everywhere in travel. Article 9 restricts processing of “data concerning health” among other categories (UK GDPR Article 9), and the ICO confirms the list includes health data and biometric data used for identification (ICO: what is special category data). In a hotel or tour operator, that means allergy and dietary notes that reveal a condition, mobility and assistance requirements, medical reasons for a cancellation, and sometimes religious information inferred from meal requests. These are exactly the details staff paste in when asking an assistant to “write a sensitive reply”.
Security and breaches. Article 32 requires security “appropriate to the risk”, including pseudonymisation where appropriate (UK GDPR Article 32). Where a personal data breach occurs, Article 33 requires notifying the ICO “without undue delay and, where feasible, not later than 72 hours” after becoming aware, unless it is unlikely to result in a risk to individuals (UK GDPR Article 33). Whether a paste into a personal chatbot account is a reportable breach depends on the facts; the point is that someone has to make that assessment, quickly.
International transfers. Sending personal data to a separate organisation outside the UK is a “restricted transfer” that must be covered by UK adequacy regulations, appropriate safeguards such as the International Data Transfer Agreement (IDTA) or the International Data Transfer Addendum, or an exception (ICO: international transfers). Business AI contracts normally address this; personal accounts leave it to whatever the consumer terms say.
Data residency is useful, but it is not the whole answer
UK data residency sounds like it settles the question. Read the scope. For ChatGPT, data residency keeps in-scope customer content “stored at rest” in the selected region; inference residency — where the model actually runs — was listed for Europe, the United States and the UAE, not the UK, on the page we read. Account data, billing, metadata and “logs without content” are out of scope, and data sent through external integrations such as apps, MCP servers or web search is “handled under that provider’s own storage, processing, privacy, and data-residency terms” (OpenAI: data residency and inference residency). Microsoft’s equivalent caveats are the Bing web-query and capacity-routing points above. Residency is a good control to choose; it does not replace the processor contract, the DPIA or deciding what data goes in.
A three-tier data rule staff can actually follow
Policies fail when staff have to interpret UK GDPR at the reservations desk. The ICO’s own advice for AI policies is to “define clearly what AI can and cannot be used for in your organisation”, make staff aware that what they share “can be accessed by others outside of their organisation”, and configure “settings that are the most privacy-enhancing” ([ICO: putting an AI policy in place]({ICO_POL})). For travel teams we recommend three tiers:
- Tier 1 — any approved tool, including web-only assistants: no personal data. Drafting a generic cancellation-policy explainer, rewriting a tour description, summarising a public regulation, preparing training material.
- Tier 2 — approved business workspace only, signed DPA, training off, retention set: ordinary booking and correspondence data needed for the task. Guest names, booking references, dates, itinerary details, the text of a complaint, a supplier email. Minimise first: replace names with booking references where the task allows.
- Tier 3 — not in a general-purpose assistant, even a business one, unless a specific DPIA has approved that use: card data, passport and visa scans, health, accessibility and allergy details, information about children, security credentials and API keys, and anything about staff disciplinary or grievance matters.
Card data deserves its own rule: it brings PCI DSS scope with it, which we cover in PCI DSS and AI travel checkout. Allergy and dietary information is covered in our piece on AI concierges and guest allergen questions.
Admin settings to check before you roll a tool out
- Contract: DPA signed for the exact product and plan staff will use; sub-processors and transfer mechanism noted in your records.
- Training: confirmed off by default for the plan, and response feedback settings understood.
- Retention: set to the shortest period that still supports the work; deletion behaviour documented.
- Residency: region selected at workspace creation where offered (for ChatGPT it applies to new workspaces), with its limits recorded.
- Connectors and web search: only the apps, connectors and web search you have assessed are enabled, because they can move data outside the vendor’s core terms.
- Permissions: SharePoint, Drive and shared-mailbox access cleaned up before enabling an assistant that can search them.
- Identity: single sign-on with company accounts, so you can remove access when someone leaves and stop personal-account use on company devices.
- Audit: conversation logs or compliance exports available to the people who need to investigate an incident.
If someone has already pasted guest data into a personal account
- Find out what was shared, which tool and account, and when. Do not make it a disciplinary first conversation, or the next one will not be reported.
- Delete the conversation and check the account’s training setting; note what the vendor’s terms say about retention and deleted chats.
- Assess whether it is a personal data breach and whether it is likely to result in a risk to the guests involved; if notification is required, the 72-hour clock under Article 33 runs from awareness.
- Record the decision either way, and fix the cause — usually that there was no approved tool for a task people genuinely needed to do.
A worked example (hypothetical)
A 40-person UK tour operator finds that half its sales and customer-service team use personal ChatGPT or Gemini accounts to draft emails. Rather than ban them, it picks one business workspace, signs the DPA, turns on single sign-on, sets retention to 90 days and disables connectors it has not assessed. It writes the three-tier rule onto one page, runs a 30-minute session per team, and adds a DPIA for the one higher-risk use it wants: summarising complaint files, which often mention medical reasons for cancellation. That use is approved only with names replaced by booking references. Six weeks later, the operations lead notices two staff still pasting passport details to check name spellings against tickets — a Tier 3 use — and replaces it with a field-matching check inside the booking system instead.
How to measure whether it worked
Track adoption of the approved workspace against headcount in the teams that handle guest data, the number of personal-account AI services still reached from company devices or networks, incidents and near-misses reported under the policy, and how long it takes to answer “where did this guest’s data go?” when someone asks. Fewer reported incidents is not automatically good news; in the first months it often means people have stopped reporting.
What this article is not
- Not legal advice, and not a recommendation of any one vendor; the quotes describe each vendor’s published defaults on the date we read them.
- Not a claim that business plans make any use compliant — the controller duties, DPIA and data minimisation still apply.
- Not a case study; the example above is hypothetical and Kaize reports no client results here.
Questions
Is turning off “Improve the model” on a personal ChatGPT account enough?
No. It stops new conversations being used for training, but feedback can still bring a whole conversation into training on individual plans (OpenAI: how your data is used), and there is still no processor contract between your business and the vendor (UK GDPR Article 28).
Does UK data residency mean the data never leaves the UK?
Not on the terms we read. ChatGPT residency covers in-scope content at rest; UK inference residency was not listed, and metadata and external integrations are out of scope (OpenAI: data residency and inference residency).
Are temporary or incognito chats safe for guest data?
They reduce training use, but they do not create a business contract. Google, for example, still keeps temporary Gemini chats for 72 hours (Gemini Apps Privacy Hub).
Is this the same question as letting an AI agent into our booking system?
No. This is about staff using general assistants. Giving an AI agent credentials to read or change bookings raises different controls, covered in how travel companies should control AI agents and prompt injection in travel AI agents.
Kaize point of view
Give staff a better approved tool than their personal one. Banning AI assistants in a travel business mostly moves the problem out of sight: people still need to answer an upset guest at 7pm, and the personal account is one tap away. The defensible position is a business workspace with a signed DPA, training off and retention set, a three-tier data rule short enough to remember, and a no-blame route for reporting mistakes. Then look at what staff keep pasting in. If a team copies the same booking data into an assistant every day, that is a workflow asking to be built properly — inside the booking system or CRM, with the right permissions — not a habit to police.
Related reading: UK GDPR and ATOL when travel AI changes a booking, automated decisions and profiling in a travel CRM, disclosing AI before a guest chat under the EU AI Act, Booking.com messaging and AI guest replies and build vs buy for travel software.
If you run a hotel group, tour operator or DMC and want help choosing and configuring AI tools for staff, writing a data rule your teams will follow, or turning repeated copy-and-paste into a proper integration, start with an AI opportunity review or discuss a project with Kaize. For how guest data moves between your systems, see connecting travel booking systems and CRM.