Shadow AI | Stopping Company Data Leaking Into Chatbots

Somewhere in your organisation this week, someone pasted a customer list into a chatbot to reformat it. Someone else uploaded a contract to get a summary before a meeting. A developer pasted production code containing credentials to debug an error faster.

None of them meant harm. All of them were solving a real problem under time pressure with a tool that worked. And in most organisations, none of them broke a rule, because no rule exists.

This is shadow AI, and it differs from earlier shadow IT in one important respect. When staff adopted unapproved file sharing, the data went somewhere the organisation did not control. When staff paste into AI tools, the data may go somewhere the organisation does not control and may contribute to a system it can never retrieve it from.

Fortinet expanded FortiEndpoint with AI visibility and governance, native data security, and FortiAI-assisted operations, aimed at helping organisations govern AI usage, reduce sensitive data exposure, and enforce risk-aware access across distributed environments. The capability exists because the problem is now general.

Why prohibition fails

The instinctive response is to block AI tools at the perimeter. It does not work, and understanding why matters before designing anything else.

Staff use these tools because they deliver real productivity gains. Removing the tool does not remove the pressure that drove its adoption. It relocates usage to personal devices and personal accounts, where the organisation has no visibility at all. Blocking converts a governable problem into an invisible one.

There is also a competitive dimension. Organisations that prohibit AI outright while competitors adopt it with governance accept a productivity disadvantage in exchange for a control that does not hold.

The workable objective is governed usage: approved tools, clear boundaries on what may be shared, and technical enforcement behind the policy.

What actually leaks

Understanding the categories helps target the response.

Customer and personal data pasted for formatting, analysis, or drafting. This carries direct regulatory consequence under Saudi data protection obligations, because it is a transfer of personal data to a processor the organisation never assessed.

Source code, frequently including embedded credentials, API keys, and connection strings. Developers paste code to debug, and the secrets travel with it.

Contracts and commercial terms uploaded for summarisation before meetings. Confidentiality obligations to counterparties rarely contemplate this.

Internal strategy material used to generate summaries and presentations. Competitively sensitive by definition.

Credentials and configuration pasted in the course of troubleshooting, which is the highest-severity category and the least discussed.

Building governance that holds

Discovery comes first. An organisation cannot govern usage it cannot see. Network and endpoint telemetry establishes which AI services are being reached, from where, and at what volume. The results routinely surprise leadership, both in scale and in which departments lead adoption.

Classification defines the boundary. Policy should state which data categories may be shared with which tools. A blanket prohibition gets ignored because it is unworkable. A clear line — public and internal material acceptable in approved tools, confidential and personal data never — gets followed because staff can apply it without deliberation.

Approved tooling removes the excuse. If the organisation provides a sanctioned AI capability that meets the actual need, compliance costs staff nothing. Enterprise arrangements typically offer data handling commitments that consumer tools do not, which is the substantive reason to route usage through them.

Technical enforcement backs the policy. Endpoint controls that identify sensitive data patterns and prevent transmission to unapproved services turn policy into something operative rather than declarative. Enforcement should distinguish approved from unapproved destinations rather than blocking a category wholesale.

Training explains the reasoning. Staff who understand why a boundary exists apply judgement in cases the policy never anticipated. Staff given rules without reasons apply them literally and stop at the edges.

Monitoring closes the loop. Governance decays without visibility. New tools appear constantly, and usage patterns shift. This is an ongoing operational function, not a project with an end date.

The regulatory angle

For Saudi organisations, shadow AI sits at the intersection of several obligations that already apply.

Pasting personal data into an external AI service is a transfer to a third party, engaging data protection obligations around lawful basis, processor assessment, and in many cases the location where processing occurs. The organisation carries that obligation whether or not it knew the transfer happened.

Under the third-party and cloud components of current cybersecurity control frameworks, an unassessed external service processing company data is precisely the exposure those controls exist to address. An organisation that cannot describe which AI services its staff use cannot evidence third-party governance.

Frequently asked questions

Should the organisation just block AI tools? Blocking pushes usage onto personal devices where no visibility exists, and forfeits productivity gains competitors are capturing. Governed adoption is more effective and more defensible.

How is shadow AI usage discovered? Network and endpoint telemetry showing which services are being reached and by whom. Most organisations already collect the underlying data and have never analysed it for this purpose.

Does enterprise AI tooling solve the problem? It addresses the largest part, by giving staff a sanctioned route with better data handling commitments. It still requires classification policy and enforcement, because staff will continue reaching unapproved tools for tasks the approved one handles poorly.

What is the highest-severity category of leakage? Credentials and configuration data pasted during troubleshooting. It attracts the least attention and creates the most direct exposure.

How ITBuilders supports AI governance

ITBuilders helps organisations establish visibility over AI usage, define workable classification policy, and enforce it through endpoint and network controls. Because the same team operates the underlying security infrastructure, enforcement is designed around what the estate can support and what staff will actually follow.

To discuss AI governance, contact ITBuilders at 920-020-750 or itbuilders.com.sa