Usage arrived before the policy
In most companies, staff did not wait for an official decision. In the 2024 Work Trend Index from Microsoft and LinkedIn, based on a survey of 31,000 people across 31 countries, 78% of AI users said they were bringing their own AI tools to work. The figure is global rather than Saudi-specific, but it describes something every manager will recognise: an employee pastes an email into a chat tool to rephrase it, a spreadsheet to summarise it, or a contract to pull out its clauses.
That is not bad behaviour in itself; the tools save real time. The problem is that the most important decision — what may be entered, into which tool, under which account — is made individually every time, by someone who does not know what happens to the data after they press send. A blanket ban rarely solves it. It usually moves the usage onto personal phones, where nobody sees it.
What is the main risk of using generative AI tools with work content?
The main risk is sensitive information leaking outside the company’s control. Every piece of text or file pasted into an external tool leaves your systems and may be retained, reviewed or used according to the provider’s terms and the account’s settings, not according to your policy. That includes customer and employee personal data, classified information, contracts and pricing, source code, and passwords and system keys.
That is why the generative AI guidelines published by the Saudi Data and AI Authority (SDAIA) recommend that organisations implement usage policies prohibiting the entry of classified information into third-party tools, setting out acceptable use and the steps that prevent sensitive data leaks. The OWASP project on LLM application security lists sensitive information disclosure among the ten most significant risks in its 2025 edition.
Leakage here needs no attacker and no bad intent. A rushed employee, a personal account and a file nobody checked are enough.
The other risks: confident errors, intellectual property, and bias
NIST’s Generative AI Profile (NIST AI 600-1) lists twelve risks, and four sit closest to everyday business use. The first is what it calls confabulation: wrong content delivered with full confidence — a reference that does not exist, an invented figure, a legal clause with no basis. SDAIA’s guidelines say plainly that verifying generated content is the user’s responsibility, by scrutinising it and checking it against trustworthy sources.
The second is intellectual property: output that resembles protected material, or company content entered without the right to share it. SDAIA’s guidelines call for adherence to the Personal Data Protection Law and intellectual property laws when using these tools. The third is data privacy and information security, covered above. The fourth is harmful bias: output that reflects bias in the training data, which matters especially in first-pass screening of job applicants or customer assessments.
Daily use adds a fifth risk that no list mentions: over-reliance. When the machine draft becomes the final version without review, its mistakes reach customers under your name.
Personal account or business account: the difference that changes everything
Data terms differ sharply between consumer and business accounts from the same provider. Anthropic, for example, states that by default it does not use inputs or outputs from its commercial products to train models, while users of its consumer plans choose for themselves whether their chats may be used to improve the model. Microsoft states that prompts, responses and data accessed through Microsoft 365 Copilot are not used to train its foundation language models.
The practical lesson is not to pick a particular vendor. It is to ask three questions of any tool before approving it. Is our data used for training, and how do we turn that off? Where is it processed and stored, and for how long? Who at the provider can access it? The answers sit in the terms of service and privacy documentation, and a named person in the company should read them rather than every employee alone.
The logical conclusion: give your staff an approved tool under a business account. It works better than a ban that pushes them back to personal accounts.
A two-page usage policy: what it must contain
An effective policy is short enough that everyone reads it and specific enough to answer the employee’s question in the moment. Start by sorting data into three classes everyone understands. Green: public, published information, usable in any approved tool. Amber: internal but not sensitive, usable only in approved tools under the company account. Red: personal data, confidential client information, classified material, contracts, passwords and keys — never entered into any tool that has not been approved for exactly that purpose.
Then list the approved tools and the account type allowed for each, and state the duty to verify: no generated content goes to a customer or outside party without human review. Add a disclosure rule for when AI assistance must be stated; SDAIA’s guidelines recommend notifying users when messages or content have been generated by AI. Add rules for intellectual property and source code, and a clear channel for reporting a mistaken entry without fear of punishment, because fast reporting matters more than hiding the error.
Finally, give the policy a named owner and a review date. The tools and their terms change faster than any document.
The Saudi framework: what to put in front of your counsel
The Personal Data Protection Law applies to any processing of personal data relating to individuals that takes place in the Kingdom, and transfers of personal data outside the Kingdom are governed by a dedicated regulation issued by SDAIA. When an employee pastes customer data into a tool whose provider or servers sit outside the Kingdom, real legal questions arise that no article and no technical team can settle. They belong with your legal counsel before the tool is approved, not after an incident.
On the security side, the National Cybersecurity Authority opened a consultation in July 2026 on draft AI Cybersecurity Guidelines whose scope includes emerging technologies such as generative and agentic AI, across governance, defence, resilience and third-party domains. Watch for the final version; it is likely to become a reference for organisations subject to the Authority’s controls.
SDAIA’s generative AI guidelines are a sound starting point for drafting your own policy: written for the public and for organisations, in plain language, and covering verification, disclosure, intellectual property and data leakage.
From document to behaviour: training and approved tools
A policy nobody knows protects nobody. SDAIA’s guidelines recommend targeted training that shows staff how to reduce legal risk and reminds them that data governance, cybersecurity, personal data protection and intellectual property obligations remain binding when they use these tools. The training that works is short and practical: real examples from the department’s own work of what is and is not allowed.
The second step is to make the safe route the easy one. An approved tool with a business account and single sign-on is easier than a personal account. An internal assistant that answers from the company’s approved documents, within each employee’s permissions, removes the reason to paste those documents into a public tool.
The third step is measurement. Which tools are actually in use, in which departments, for which tasks? A simple quarterly inventory catches individual initiatives before they turn into unmanaged dependencies.
How DataWork can help
DataWork offers responsible AI and governance advisory. It starts with an inventory of AI use across departments, including tools that began as individual initiatives, and classifies each use by its effect on customers, personal data and consequential decisions. We draft with you an acceptable use policy and a data handling policy in Arabic and English, define the roles — who approves, who reviews, who can switch something off — and provide a vendor and tool assessment checklist for use before any new contract, along with human intervention and escalation procedures and a scheduled review cycle.
We work alongside your legal counsel, not in place of them, and we do not issue compliance certificates. If the next step is deciding where AI genuinely makes a difference in your business, our AI strategy and readiness service is the way in. If the need is a safe internal assistant that answers from your approved documents, our AI assistants service builds one with clear permissions.
Key takeaways
- The main risk of using generative AI tools with work content is sensitive information leaving the company’s control.
- Verifying generated content is the user’s responsibility, as SDAIA’s guidelines state; nothing goes to a customer without human review.
- Data terms differ between consumer and business accounts from the same provider; provide an approved business tool instead of a blanket ban.
- An effective policy fits on two pages: three data classes, approved tools, a duty to verify, a disclosure rule, a reporting channel and a named owner.
- Put questions about personal data and transfers outside the Kingdom to your legal counsel before approving any tool.
Frequently asked questions
What is the main risk of using generative AI tools with work content?
Sensitive or confidential information leaking outside the organisation’s control. Whatever is pasted into an external tool leaves its systems and may be retained or used under the provider’s terms. That is why SDAIA’s guidelines recommend policies prohibiting classified information from being entered into third-party tools.
Can we put customer data into ChatGPT or similar tools?
Do not enter customer personal or confidential data into any tool the company has not approved for that specific purpose, and never into a personal account. Processing personal data falls under the Personal Data Protection Law, and a provider outside the Kingdom may raise transfer questions your legal counsel needs to answer.
Are my conversations used to train the model?
It depends on the account type, its settings and the provider’s terms. Some providers do not use business account data for training by default, while others leave the choice to consumer users. Read the privacy documentation for your approved tool and switch training off where the option exists.
Do we have to disclose that AI was used to prepare work?
SDAIA’s guidelines recommend notifying users when messages or content have been generated by AI, and being clear when it is used in dealings with the public. Your policy should define when disclosure is mandatory, especially for customer-facing content.
Who is responsible if the tool gives a wrong answer?
The user and the organisation they work for, not the tool. SDAIA’s guidelines state that verifying generated content is the user’s responsibility, so the policy needs a human review rule before anything is sent to a customer or outside party.
Should we just ban AI tools at work?
A blanket ban usually pushes usage onto personal devices and accounts where there is no oversight. It is more effective to provide an approved tool under a business account, with a clear policy on what may be entered, short training, and periodic review of what is actually used.
Sources
- Guidelines: Generative Artificial Intelligence for Public — Saudi Data and AI Authority (SDAIA) · https://sdaia.gov.sa/en/SDAIA/about/Files/GenerativeAIPublicEN.pdf
- AI Cybersecurity Guidelines — public consultation — National Cybersecurity Authority (NCA) · https://nca.gov.sa/en/news/ai-cybersecurity-guidelines/
- Personal Data Protection Law — Bureau of Experts at the Council of Ministers · https://laws.boe.gov.sa/BoeLaws/Laws/LawDetails/b7cfae89-828e-4994-b167-adaa00e37188/1
- Regulation on Personal Data Transfer Outside the Kingdom — Saudi Data and AI Authority (SDAIA) · https://dgp.sdaia.gov.sa/wps/wcm/connect/e5bbede0-1119-4f70-b4ef-f043ce58d780/Regulation+on+Personal+Data+Transfer+Outside+the+Kingdom..pdf?MOD=AJPERES
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1) — National Institute of Standards and Technology (NIST) · https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
- OWASP Top 10 for LLM Applications 2025 — OWASP GenAI Security Project · https://genai.owasp.org/llm-top-10/
- AI at work is here. Now comes the hard part (Work Trend Index 2024) — Microsoft and LinkedIn · https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part
- Is my data used for model training? — Anthropic Privacy Center · https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training
- Data, privacy, and security for Microsoft 365 Copilot — Microsoft Learn · https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-privacy