Agenticum · Thoughts · 8 min read

Can Employees Upload Company Data to ChatGPT? A One-Page Policy That Works

The honest answer is "it depends on which ChatGPT, which data and which country". Here is the short version of that answer, the settings to change today, and a policy you can paste into your handbook. Checked against the vendors' pages on 27 October 2026.

Nikolay Zuykov · Agenticum · 27 October 2026

In April 2023 engineers at Samsung pasted internal material into ChatGPT to get help with their work. Within a month the company restricted generative AI on its devices and told staff not to submit company or personal information to such tools, because, as its memo put it, data on external servers is hard to retrieve and delete. Three years later the same thing happens every day in Dubai and Riyadh offices, mostly through personal accounts that nobody in IT knows about.

The fix is not a ban. People will use the tools anyway, from their phones. The fix is a one-page rule that says which tool is allowed for which kind of data, plus three settings that take ten minutes to change.

Which ChatGPT? The plan decides what happens to the text

"ChatGPT" is at least five different contracts. The same is true of Claude, Gemini and Copilot. On 27 October 2026 the vendors' own pages said this:

ToolPersonal plansBusiness plans
ChatGPT"Content is used to train our models. Opt-out available." Free, Go, Plus and Pro all show this line.Business and Enterprise: "No training on your business data by default." Enterprise adds custom retention and data residency in ten regions, including the UAE.
ClaudePricing page: training "Opt-out". Anthropic's privacy centre describes it as a choice the user makes, and incognito chats are excluded.Team: "No model training on your content by default." Enterprise: "None by default." Anthropic's commercial terms: it "may not train models on Customer Content".
Gemini"Gemini Apps Activity" is on by default. A subset of chats is read by human reviewers and kept for up to three years, and deleting your history does not remove those.Workspace: no training without the customer's prior permission, no human review, interactions stay within your organisation.
CopilotPersonal Copilot has its own terms; not assessed here.Prompts, responses and files reached through Microsoft Graph "aren't used to train foundation LLMs". Stored under the same contract as the rest of your Microsoft 365 data.

So the question in the title has a precise answer. On a business plan the company's own data stays the company's. On a personal plan the default is that the vendor learns from it, and on Gemini a human may read it. A policy that does not make this distinction will be ignored, because people can see that their colleague's Copilot is "allowed" while their own ChatGPT Plus is "banned" and nobody explains why.

Which data? Three classes are enough

Long classification schemes die in the drawer. Three classes cover what most companies in the region actually handle.

Most "can I paste this" questions resolve themselves once people know which class the document is in. The third class is where the law applies, and in both countries it now applies with real penalties.

Which country? Two laws, two regulators

United Arab Emirates. Federal Decree-Law No. 45 of 2021 has applied since 2 January 2022. It prohibits processing personal data without the owner's consent except in defined cases, requires companies holding personal data to keep it secure and confidential, and sets conditions for cross-border transfer. The UAE Data Office is the federal regulator. Companies in DIFC answer to DIFC Law No. 5 of 2020 instead, and ADGM has its own regulations, so check which regime your licence falls under.

Saudi Arabia. The Personal Data Protection Law has been in force since 14 September 2023, with SDAIA as regulator. Consent is the main legal basis for processing. Transfers outside the Kingdom are allowed only under conditions, including adequate protection or appropriate safeguards. A breach that has to be reported goes to SDAIA within 72 hours. Fines reach SAR 5 million, courts can double them for repeat offences, and deliberate disclosure of sensitive data can mean up to two years in prison.

Pasting a customer's passport scan into a personal ChatGPT account is, under both laws, a transfer of personal data to a processor abroad on terms nobody in your company has read. That is the sentence to put in front of a sceptical manager.

Three settings to change this week

  1. Personal ChatGPT accounts that remain in use. Settings, Data controls, turn off the option that lets OpenAI use your content to improve its models. OpenAI's pricing page confirms the opt-out exists on every personal plan. Better: move these people to a Business seat, which showed as AED 73 per month on 27 October 2026.
  2. Personal Gemini accounts. Turn off "Gemini Apps Activity" (Google now also calls it "Keep Activity"). Google's page says future chats then "won't be used to train our AI models", though they are still stored for 72 hours. Note that some connected features stop working when it is off.
  3. Claude personal accounts. Check the model-improvement setting in Privacy settings and use incognito chats for anything sensitive. Or move the user to Claude Team, where the default is no training.

On Copilot and on Workspace the business defaults are already right. What needs attention there is the permission model: Copilot shows a person whatever they can already open, so an over-shared SharePoint folder becomes an over-shared Copilot answer. Clean the sharing settings before the rollout, not after.

The one-page policy

Use of AI assistants at [Company]

1. Approved tools. [ChatGPT Business / Claude Team / Microsoft Copilot / Gemini for Workspace], on company accounts only. Personal accounts, including paid ones, are not to be used for company material.

2. Data classes. Public material may be used with any approved tool. Internal material may be used only with approved tools. Personal and regulated data may be used only with [tool], only when the task requires it, and with identifiers removed where possible.

3. Customer and employee data. Names, ID and passport numbers, salaries, health and bank details are personal data under UAE and Saudi law. Do not paste them into any tool that is not on the approved list. When in doubt, ask [owner of this policy].

4. Output. Anything an assistant writes is a draft. The person who sends it is responsible for it. Check numbers, names and legal wording against the source before it leaves the company.

5. Exceptions. Requests for a new tool or a new use go to [owner]. Decisions are recorded in [place]. This policy is reviewed every six months, because vendor terms change.

6. Incidents. If company or personal data went into a non-approved tool, tell [owner] the same day. Under Saudi law a reportable breach must reach SDAIA within 72 hours, so speed matters more than blame.

Six clauses. Put the owner's real name in it, not a department. Policies with a department as owner have no owner.

Making it stick

Three things make the difference between a policy that lives in the handbook and one that people follow.

First, give people an approved tool that is at least as good as the one they were using on their own. If the company account is slower or weaker, the personal account comes back within a week.

Second, show the setting, not the paragraph. A ten-minute session where everyone opens their own account and flips the training option does more than any memo.

Third, look at the admin console once a month. The business plans show usage per user. If a heavy user suddenly shows no activity, they have probably gone back to a personal account, and that is a conversation, not a disciplinary case.

Sources checked on 27 October 2026. The policy text is a template, not legal advice; have counsel review it against your licence, your regulator and your contracts.