Shadow AI: How to Find and Govern the AI Tools Your Staff Already Use

Shadow AI is any AI tool that staff use for work without the organisation's approval: a personal ChatGPT account, a browser extension that summarises email, or an AI feature a vendor added to software you already pay for. You govern shadow AI in four moves: find what is in use, classify the data at stake, publish an approved-tools list with a fast route to request new tools, and add access and data loss prevention (DLP) controls. A blanket ban without approved alternatives tends to move the use to personal phones, where you cannot see it.

The scale is hard to judge from outside. Microsoft and LinkedIn's 2024 Work Trend Index, published in May 2024, reported that 78% of AI users were bringing their own AI tools to work. The survey covered about 31,000 people in 31 markets, so treat it as a global signal and measure your own organisation.

Why does banning AI tools rarely work?

Staff use unapproved AI because it saves time on real tasks: drafting Arabic and English letters, summarising tender documents, cleaning spreadsheets. Remove the tool without a replacement and the need stays. The likely result is the same prompts typed on a personal phone, outside your network, your logs and your data protection controls.

Saudi guidance points to controlling the data rather than the idea of using AI. SDAIA's Generative AI Guidelines for the public (May 2025 edition) tell organisations to adopt usage policies that prohibit users from entering classified information into third-party tools, to monitor and review adherence, and to train employees. Block what is unsafe, approve what is useful, and tell staff which is which.

How do you discover shadow AI in your organisation?

No single source shows everything. Combine four and compare what each one finds. Check that reviewing these logs is covered by your employee privacy notice and IT acceptable-use policy.

Frame the survey as an amnesty: you are collecting use so you can approve good tools, not to discipline anyone. Ask which tasks people use AI for, which tools, and whether they have pasted customer, employee or financial data. Honest answers depend on that promise being kept. The amnesty covers discipline, not legal duties: if an answer shows that personal data has already left your control, handle it through your incident process, which may include breach notification under the applicable data protection law.

Many organisations already own the technical discovery tools. Microsoft's documentation, checked in October 2026, states that the Defender for Cloud Apps catalogue has a Generative AI category with more than a thousand apps and a risk assessment for each. Administrators can sanction or unsanction apps and set policies that keep discovering new AI apps by risk score or daily users. Other cloud access security broker (CASB) and secure web gateway products offer similar categories, so check what your current licences include before buying anything.

Finally, review the admin consoles of the software you already use, such as CRM, office suites and meeting tools. Vendors add AI features over time, and each one needs the same decision as a new tool.

Classify the data before you judge the tools

The risk of a shadow AI tool depends less on the tool than on what goes into it. Use the classification scheme you already have. Saudi public entities have one: the national Data Classification Policy published by SDAIA sets four levels, Top Secret, Secret, Restricted and Public. Many private companies use Public, Internal, Confidential and Restricted. Map each level to the tools it may enter.

The table is an illustration; your own classification policy and sector controls take precedence.

Personal data needs extra care. Saudi Arabia's Personal Data Protection Law, the UAE's Federal Decree-Law No. 45 of 2021 and Jordan's Personal Data Protection Law No. 24 of 2023 each set conditions for processing and for transfers outside the country, so check where each provider processes and stores data. This is general information, not legal advice; your obligations depend on your sector and the data involved.

Build an approved-tools list people will use

The approved-tools list is the answer to the question every employee asks: what may I use? Keep it to one page, in Arabic and English, and record for each tool:

• the tool and the plan (enterprise or consumer), since the same product can carry different data terms;
• the data classes it may receive and the tasks it is approved for;
• a named owner and the date of the next review;
• where to get access, usually through single sign-on.

Pair the list with a request route: a short form, a named reviewer and a published turnaround time. If staff wait months for an answer, shadow AI returns. A regional bank and a family-owned retailer will set different limits, but both need the list and the route.

Access and DLP controls that make the list stick

Policy tells people what to do; controls catch mistakes. Introduce them in monitor mode first, then warn, then block, so you learn what breaks before users do.

• Identity: put approved AI tools behind single sign-on and conditional access, so only managed accounts and devices reach corporate data.
• Edge controls: warn on, then block, unsanctioned AI categories at the gateway or CASB, starting with apps that score worst on risk.
• Data loss prevention: apply DLP to prompts and uploads. Microsoft states that Purview DSPM for AI offers ready-made policies against data loss in AI prompts, and that monitoring third-party AI sites requires devices to be onboarded to Purview.
• Browser extensions: allow-list extensions on managed browsers, since many AI add-ons can read page content.
• Records: keep audit logs for approved tools so you can investigate an incident later.

A 30-day plan to bring shadow AI under control

1. Days 1 to 5: name an owner, announce the amnesty survey, and route new AI purchases through the request form.
2. Days 6 to 12: run the survey and interviews; pull SSO, gateway and expense data; build one inventory.
3. Days 13 to 18: classify the data each tool receives and decide keep, replace or block for each one.
4. Days 19 to 24: publish version 1 of the approved-tools list and the request route, and switch on enterprise alternatives for the most common tasks.
5. Days 25 to 30: move gateway rules for the highest-risk apps from monitor to warn or block, test DLP policies, and report the inventory and decisions to management.

By day 30 you will not have removed every unapproved tool. You will know what is in use, which data is exposed, and what staff should use instead.

How to start

Our free AI Exposure Scan is a 45-minute conversation on where AI is already exposing you, ending in a written next step. For a full inventory and the policy around it, see our shadow AI service page and the AI Governance & Compliance Foundation. When you are ready to lock approved tools to corporate identity and add DLP, AI Access & Data Protection covers SSO, conditional access, edge controls and recorded platform settings. More is on the Governance Lab page.

Related lab: Governance Lab ↗