<aside> đź“ť
The procurement team requested access to external AI tools to help query, summarize, and analyze consolidated supplier information.
From a business perspective, the request was reasonable. Procurement users wanted to improve efficiency when reviewing supplier profiles, comparing vendor records, summarizing contract-related information, and extracting insights from scattered supplier datasets.
However, from a security perspective, directly allowing external AI access to supplier information introduces significant data exposure risks. Supplier data may include company names, contact persons, email addresses, phone numbers, pricing information, contract terms, negotiation history, payment information, risk ratings, internal comments, and other business-sensitive fields.
Therefore, this is not a simple “allow or block AI” problem. The real question is:
Can users benefit from external AI while preventing sensitive supplier data from being exposed to unmanaged third-party AI services?
</aside>
Actually, my first instinct was to apply the same control mindset used in Shadow AI and Shadow MCP governance. Just as mentioned in 01. Case Note: Shadow AI / Shadow MCP Control Review. However, these controls are important, but they are mostly visibility and audit controls.
They do can answer questions such as:
Who used the AI tool?
Which AI service was accessed?
When did the access happen?
Was there a file upload or abnormal traffic pattern?
Was the usage approved or unapproved?
But they do not fully solve the prevention problem.
If the procurement team has a legitimate need to use AI, simply monitoring traffic does not prevent sensitive supplier information from being pasted into an external model. Blocking all AI access may also be too disruptive, because the business use case itself is valid.
So the control strategy needs to move from:
“Can we detect external AI usage?”
to:
“Can we control what data is allowed to reach external AI?”
At first, my thinking was still close to traditional Shadow AI governance: identify external AI usage, monitor network traffic, review logs, and audit suspicious behavior. However, this procurement use case exposed a different problem. Monitoring can tell us that supplier data may have been sent to an external AI tool, but it does not necessarily prevent the sensitive content from leaving the organization in the first place.