Shadow AI at Work: Data Privacy and AI Governance in the UAE

Shadow AI is what happens when employees use their own AI accounts, such as personal ChatGPT, Claude or Gemini plans, for company work without the company knowing or controlling it. Responsible use of AI in the enterprise starts by giving people an approved AI tool that the company pays for and manages, then adding a short policy, clear rules on which data may go into it, and a way to see what's being used. In the UAE, which has the highest share of AI users of any country, this matters to every management team, and most of all in healthcare, banking, insurance and government supply.

We recently visited a company where almost everyone had their own AI subscription. Sales pasted client emails into one tool, finance pasted spreadsheets into another, and a developer pasted code with a database password still in it. Nobody meant any harm. They used personal plans the company had never seen, and most of them didn't know that on several of those plans their chats can be used to train the vendor's models unless they switch it off.

What shadow AI is and why employees do it

Shadow AI is the AI version of shadow IT. It covers personal chatbot accounts, free AI websites, browser extensions, AI note-takers that join meetings, and AI features switched on inside other apps. The company didn't choose the tool, has no contract for it and can't see what goes into it.

People do it because it saves them time, and a memo won't change that. Microsoft and LinkedIn's 2024 Work Trend Index found that 75% of knowledge workers already used AI at work, and that 78% of those users brought their own AI tools. When a company bans AI without offering anything in its place, the use doesn't stop. It moves to personal phones, where nobody can see it at all.

So the useful question for management is how to give staff AI they're allowed to use, on terms the company controls.

On a personal account, company data goes to the vendor under the employee's own terms; in a company-managed tenant, it stays under the company's contract, settings and logs.

Which AI plans train on your data

Personal plans and business plans from the same vendor follow different rules. Here's what each vendor's own help and privacy pages said in November 2026. They change often, so check them again before you decide.

OpenAI ChatGPT

• Free, Plus and Pro: "Improve the model for everyone" is on by default, so chats can be used for training. Switch it off under Settings, Data controls. Temporary Chats aren't used for training and are kept for up to 30 days. A thumbs up or down can still send that conversation for training.
• Business, Enterprise, Edu and the API: OpenAI says it doesn't train on this content by default.

Anthropic Claude

• Free, Pro and Max: since Anthropic changed its consumer terms in autumn 2025, each user decides under the Model Improvement setting whether chats may be used for training. If allowed, the data is kept for up to five years instead of 30 days.
• Team, Enterprise and the API (including Claude through Amazon Bedrock or Google Vertex AI): commercial terms apply, and inputs and outputs aren't used for training by default.

Google Gemini

• Gemini app on a personal account: for users 18 and over, Keep Activity is on by default, which lets Google use chats to train its AI models. Some chats are read by human reviewers and kept for up to three years. With Keep Activity off, chats are still saved for 72 hours.
• Gemini in Google Workspace (business editions): prompts are customer data under Google's data processing terms, and Google says your content isn't reviewed by people or used for training outside your domain without permission.

Microsoft Copilot

• Personal Microsoft account: "Model training on text" and "Model training on voice" sit under Privacy, and you turn them off to keep conversations out of training. History is kept for 18 months.
• Microsoft 365 Copilot with a work account: prompts, responses and company data aren't used to train the foundation models, and admins set retention in Microsoft Purview. Note for Gulf companies: Microsoft says queries from customers outside the EU may be processed in the US, the EU or other regions. It has announced processing inside the UAE for qualifying UAE organizations, expected by the end of 2026, so ask Microsoft whether your tenant qualifies.

DeepSeek's app

DeepSeek's privacy policy (February 2026) says it stores personal data in the People's Republic of China and uses prompts and uploaded files to train its models, with an opt-out by email. Running DeepSeek's open model on your own servers is different, as we explain below.

Personal plans and business plans differ on training defaults, admin control, single sign-on, audit logs, retention and the contract behind them.

With every vendor, business plans give the company a contract, admin controls and a no-training default, while personal plans leave each choice to an employee who may never open the settings page.

What actually goes wrong

Most AI data leaks come from ordinary people trying to get work done.

• Source code. In 2023, Samsung restricted staff use of generative AI after engineers uploaded internal source code to ChatGPT, as Bloomberg reported.
• Passwords and keys in prompts. A developer pastes a configuration file to fix an error, database password included, into a chat history the company doesn't control. We cover how secrets end up in code and prompts in security risks of AI-generated code.
• Client and patient data. Contracts, salary sheets, insurance claims and medical reports get pasted in to be summarized. In healthcare, that can also break the rules on where health data may be stored.
• Public links. In mid-2025, Fast Company found more than 4,500 shared ChatGPT conversations indexed by Google, some with work details. OpenAI removed the option that allowed it in August 2025.
• Shared accounts. When a team shares one personal login, everyone can read everyone else's history, and access can't be removed when someone leaves.
• Browser extensions. In December 2025, Koi Security reported that Urban VPN Proxy and related extensions, with more than eight million users, were capturing people's conversations with ChatGPT, Claude, Gemini and other chatbots.
• AI note-takers. In August 2025, a US class action accused Otter.ai of recording meeting participants without their consent. A note-taker invited by one person can record clients who never agreed.
• Made-up answers. In October 2025, Deloitte agreed to refund part of its fee to the Australian government after a report prepared partly with generative AI was found to contain references that didn't exist.

AI use in the UAE, the Gulf and worldwide

The UAE uses AI more than anywhere else. Microsoft's AI diffusion report for the first quarter of 2026 found that 70.1% of the UAE's working-age population used AI tools, the highest share of any country, up from 64.0% in the second half of 2025. In KPMG and the University of Melbourne's 2025 study, 97% of UAE respondents said they use AI for work, study or personal purposes, while 66% said their organization had a policy on generative AI.

Across the Gulf, BCG's AI at Work survey (September 2025, more than 1,000 people in Kuwait, Qatar, Saudi Arabia and the UAE) found that 87% use AI at work at least several times a week, against 72% globally. In the same survey, 63% said they would use AI tools even without company authorization, and only 45% were satisfied with their AI training.

Worldwide studies show where the data goes:

• In the 2025 KPMG and University of Melbourne study, 48% of employees who use AI at work said they had uploaded company information into public AI tools, and 57% said they hide their AI use.
• LayerX's 2025 report found that 77% of employees who use generative AI paste data into it, and 82% of those pastes come from unmanaged personal accounts.
• IBM's Cost of a Data Breach Report 2025 found that one in five organizations had a breach due to shadow AI, and high levels of shadow AI added an average of $670,000 to the cost of a breach. 63% of breached organizations had no AI governance policy or were still writing one.

The rules that apply in the UAE and the Gulf

These laws apply because pasting personal data into an outside AI service counts as processing it, and sometimes as sending it abroad. This summary (November 2026) is general information, and your legal adviser should confirm what applies to you.

UAE

• Personal Data Protection Law (Federal Decree-Law No. 45 of 2021): in force since January 2022 for onshore UAE. Its executive regulations still hadn't been published in the Official Gazette when we checked, but the law itself, with its rules on consent, purpose and transfers abroad, is in force. In June 2026 the UAE created a federal authority for AI and data that reports to the Cabinet.
• DIFC and ADGM: the DIFC Data Protection Law (DIFC Law No. 5 of 2020) and the ADGM Data Protection Regulations 2021 apply in the two financial free zones. The DIFC's Regulation 10 (2023) adds rules for personal data processed by autonomous and semi-autonomous systems, including notice to users and extra oversight for high-risk processing.
• Health data (Federal Law No. 2 of 2019): health data from services provided in the UAE may not be stored, processed or transferred outside the country except in cases the health authorities allow. Abu Dhabi's Department of Health (2018) and the Dubai Health Authority (2021) also have policies on AI in healthcare.
• Banks and insurers: the Central Bank of the UAE issued a guidance note on responsible AI and machine learning for licensed financial institutions in February 2026.
• National principles: the UAE Charter for the Development and Use of Artificial Intelligence (2024) includes data privacy, transparency, human oversight and accountability.

Saudi Arabia, Qatar and Bahrain

• Saudi Arabia: the Personal Data Protection Law, overseen by SDAIA, has been fully enforced since September 2024. SDAIA's AI Ethics Principles (2023) and Generative AI Guidelines (2024) show what regulators expect.
• Qatar: Law No. 13 of 2016 on personal data privacy protection.
• Bahrain: Personal Data Protection Law No. 30 of 2018.

In a regulated sector, you need to know which AI tools touch personal data, where that data is processed and under what contract. Shadow AI leaves you unable to answer.

AI governance that works

Good AI governance fits on a few pages and makes the approved tool the easy choice.

AI governance is built from a few blocks: a policy, approved tools, data classes, access control, monitoring, training and regular review.

• Approved tools. Staff use only AI tools the company has chosen and pays for, signed in with their work account.
• Data classes. Decide what may never go into any AI tool (passwords and keys, patient records, card data, anything under a client NDA unless the contract allows it) and what may go only into the approved tool.
• Central control. Single sign-on, training switched off by contract and in the admin settings, a retention period you choose, and audit logs IT can search.
• Blocking and data loss prevention. Once the approved tool is live, block unapproved AI sites and extensions on company devices, and add rules that stop card numbers and keys from being pasted.
• Vendor checks. Read the data processing agreement, the no-training terms, where data is processed and the list of sub-processors before you sign.
• Human review. A person checks AI output before it reaches a client or a regulator, or feeds a decision about someone's job, loan or treatment.
• Training. A short session for every employee, with real examples from your business.
• An owner and a register. A small AI committee with IT, compliance and a business lead keeps the list of approved tools and uses and reviews it every quarter.
• Incident response. Staff report a mistake at once without fear of punishment. You rotate any exposed keys, ask the vendor to delete the data and check whether you must notify anyone.

A sample AI acceptable use policy outline

1. Purpose and who it covers, contractors included.
2. Approved AI tools and how to request a new one.
3. Data classes, with examples of what may never be entered.
4. Work accounts only, with no shared logins and no personal plans for company data.
5. Checking output, and telling clients about AI use where contracts require it.
6. When meeting recorders are allowed and how participants are told.
7. What the company monitors and logs.
8. How to report a mistake.
9. Policy owner and review date.

A quick checklist

• Do we know which AI tools our staff use today?
• Have we bought an approved tool with training off by contract?
• Is sign-in through single sign-on, with access removed when people leave?
• Can IT search the logs and set retention?
• Do we know in which country each AI tool processes our data?
• Are unapproved AI sites and extensions blocked on company devices?
• Is there a named owner and a date for the next review?

Three ways to give your staff AI safely

Each step from enterprise plans to your own cloud tenant to fully local models gives you more control and more work to run it.

1. Enterprise plans from the AI vendors

ChatGPT Business or Enterprise, Claude Team or Enterprise, Microsoft 365 Copilot and Gemini in Google Workspace are the fastest route. You get the no-training default, admin controls, single sign-on and logs, and staff keep a tool they already like. Your data is processed by the vendor, often outside the UAE. For most companies and most data, that's acceptable once the contract and settings are right.

2. AI models in your own cloud account

Here the models run inside your company's own Azure, AWS or Google Cloud account, and you build your own assistant or connect them to your systems. As of November 2026:

• Azure (UAE North): GPT models can run with prompts processed in the UAE on reserved (provisioned) capacity. Pay-as-you-go deployments in the region cover embedding and speech models, and global deployments may process prompts in any Azure region.
• AWS: Amazon Bedrock has been available in the UAE region since September 2025. Check which models run in the region itself, since some are offered there only through cross-Region inference, which can process prompts in other countries.
• Google Cloud: Vertex AI lists Doha and Dammam among its generative AI locations, though not every Gemini model is in every location, and Google says choosing a regional endpoint doesn't on its own guarantee the data is processed in that country.
• Sovereign cloud: Core42, part of G42, offers a Sovereign Public Cloud built on Microsoft Azure with data hosted in the UAE.

This needs engineers to build and run it, and you pay for usage or reserved capacity. You get control over where data is processed and how AI connects to your ERP, documents and customer data.

3. Your own local AI

Open-weight models, such as Meta's Llama, Alibaba's Qwen, Mistral, DeepSeek's open models, TII's Falcon from Abu Dhabi and the Arabic-first Jais, can run on GPU servers you own or rent in a private cloud. Open-source tools such as vLLM and Ollama serve them, and a retrieval layer lets the model answer from your own documents. Nothing leaves your network.

Smaller open models now handle everyday drafting, summaries and document questions well, and their Arabic has improved a lot, so this is far more realistic than it was two years ago. The costs are real, though. You buy or rent GPUs, pay for power and cooling, and need people to install, secure, update and monitor the system as better models come out. The largest commercial models still do better on the hardest tasks.

Local AI fits your most sensitive data and cases where data must stay in the country or inside your building. For general office work, an enterprise plan is usually simpler and cheaper, and many companies end up using both.

Your first 30 and 90 days

First 30 days

• Find out which AI tools people use, with an anonymous survey and a look at network logs.
• Name an owner and form the AI committee.
• Write a one-page list of what may never go into AI.
• Buy one approved tool, switch on single sign-on and set training off.
• Give staff the approved tool before you block anything, and explain why.

Days 31 to 90

• Publish the policy and run the training.
• Block unapproved AI sites and extensions, and add data loss prevention rules.
• Start the AI register and review vendor contracts for data location.
• Pick one or two valuable uses, such as answering questions from your own policies or drafting replies from ERP data, and decide whether they need option 2 or 3. If you run Odoo 20, its built-in AI agents can handle some of this within each user's access rights.

Frequently asked questions

What is shadow AI?

Shadow AI is any AI tool employees use for work without the company's approval or control, such as personal chatbot accounts, free AI websites, browser extensions and AI note-takers. The risk is that company and client data goes to services the company has no contract with and can't monitor.

Does ChatGPT use my company's data for training?

On personal Free, Plus and Pro accounts, the training setting is on by default unless the user switches it off. On ChatGPT Business, Enterprise and Edu, and on the API, OpenAI says it doesn't train on your content by default.

Is it legal in the UAE to paste customer data into an AI tool?

It depends on the data, your sector and the tool. Personal data falls under the UAE PDPL, or the DIFC or ADGM rules in those free zones, and health data has its own law that limits processing outside the UAE. An approved tool under a proper contract, used only with the data your rules allow, is the safe route. Ask your legal adviser to confirm your case.

What should an AI acceptable use policy include?

The approved tools, what data may never be entered, work accounts only, human review of output, rules for meeting recorders, how to report a mistake, and an owner with a review date.

Is running our own AI model cheaper than buying enterprise plans?

Usually not for general office use, once you count hardware, power, staff time and updates. Local models make sense when the data is too sensitive to leave your network or must stay in the country, or when usage is high and steady enough to keep the servers busy.

Get AI your team can use safely

Your staff are probably using AI already. The aim is to bring that use onto tools your company controls, with rules people understand. We can help you choose and set up approved AI tools, build private deployments in your own cloud or on your own servers, and connect AI agents to your ERP and business systems so they work with your data under your rules. It starts with an AI automation assessment, from $1,000, that shows where AI can save your team time and how to do it safely. To talk it through, book a free 30-minute call.