Table Of Content
260 AI Services in One Organization: Turning Shadow AI Discovery Into Policy
-
October 30, 2026
-
An inventory tells you what exists. Policy decides what happens next.
In a 23 October post on its Falcon Guardian product, CrowdStrike shared what early deployments found when they first looked for AI use: hundreds of AI interactions existing controls had not recorded, more than 260 distinct AI services in use at a single organization, and confidential documents and customer records shared with public AI chatbots. Zscaler executives are making a similar point in public: shadow AI is a growing challenge, and organizations struggle to know where AI systems operate and what data they can reach.
Those numbers will not surprise most security teams. The harder question is the one after discovery: with 260 services on the list, what do you do on Monday?
Discovery is the start of the work
A list of AI services is not a policy. Teams that stop at the inventory tend to end up in one of two places: a blanket block that employees route around with personal devices and accounts, or a long allow-list that nobody maintains. Neither reflects how people actually use AI, which depends on who they are, what they are working on and what data is involved.
Tier the list by decision, not by vendor
A workable approach is to sort what you find into a small number of tiers, each with a clear action:
- 1.Sanctioned, enterprise-managed. Allowed for work data, with data classification applied to prompts and uploads, and an audit trail.
- 2.Tolerated, low-risk. Allowed for non-sensitive content only. Block or coach when sensitive data is detected.
- 3.Restricted. Allowed for specific teams or purposes with a documented owner, such as an engineering group using a coding agent.
- 4.Blocked. Services with unacceptable terms, no business need or unknown ownership.
The tiers only work if the decision can be made at the moment of use. That is where endpoint context matters.
What the endpoint adds to the decision
The same service can be fine in one situation and a problem in another. Four pieces of context do most of the work:
- Process and application. A browser tab, a desktop assistant, an IDE extension and a coding agent are different risks even when they call the same model API.
- Account. A corporate login and a personal login to the same service should not receive the same policy. This is one of the most common gaps in browser-only or network-only views.
- Device posture. An unmanaged, out-of-date or jailbroken device should not get the same access to sensitive data or AI tools as a healthy managed one.
- Data class. What is being pasted, uploaded or read by an agent determines the risk far more than which service receives it.
Because these signals sit on the device, they can be evaluated together, locally, before content leaves.
A practical first 30 days
- 1.Week 1: observe. Run in monitor mode. Capture services, processes, accounts and data categories without blocking.
- 2.Week 2: tier. Assign each discovered service to a tier with a named business owner for anything sanctioned or restricted.
- 3.Week 3: enforce the clear cases. Block tier 4. For sensitive data classes such as credentials, source code, customer records and regulated data, begin blocking or coaching on every non-sanctioned destination.
- 4.Week 4: review exceptions and tune. Look at what employees tried to do and why. Many "violations" show a missing sanctioned option rather than bad intent.
Keep it proportionate
Discovery data about how individuals use AI is sensitive in itself. Collect what the policy needs, restrict who can view user-level records and tell employees what is monitored and why. Coaching messages that explain the rule and offer an approved alternative reduce workarounds better than silent blocks.
Where Kitecyber fits
Kitecyber runs Gen AI and data security on the endpoint, so the service, process, account, device posture and data classification (80+ categories) are available in one place at the moment of use. Policies can allow, coach or block by content, destination, account and posture, with an audit trail. The agent uses under 2% CPU, and deployment typically takes around a day with pre-built compliance policies, which makes the “observe first, then enforce” approach above realistic for a small team.
Book a Kitecyber walkthrough at kitecyber.com to see what your own AI inventory looks like.
Ajay Gulati
Ajay Gulati is a passionate entrepreneur focused on bringing innovative products to market that solve real-world problems with high impact. He is highly skilled in building and leading effective software development teams, driving success through strong leadership and technical expertise. With deep knowledge across multiple domains, including virtualization, networking, storage, cloud environments, and on-premises systems, he excels in product development and troubleshooting. His experience spans global development environments, working across multiple geographies. As the co-founder of Kitecyber, he is dedicated to advancing AI-driven security solutions.
