The IT manager at an insurer receives the CASB discovery report every month. This month: 1,200 cloud services seen, 41 of them in the AI category, ChatGPT at the top. The report says how much traffic went to each domain and from how many devices. It does not say whether that traffic belonged to the business account, whether the organisation ever decided anything about those 41, or how much AI use took place outside the office network.
That is not what a CASB promises, so it is not a failing. It is where governance of AI use begins.
A CASB (cloud access security broker) looks at the traffic between your organisation and cloud services and can step in there. Governance of AI use looks at the tool: which AI services are opened, what has been decided about them, and whether the employee sees that decision at the moment of opening the tool. The two overlap at discovering use and then head in different directions.
What does a CASB see?
A CASB collects in three ways. It reads the logs of your firewall or proxy and recognises the domains of cloud services in them. It can use an agent on managed devices, so that traffic outside the office network is included. And through an API connection to the services you have bought yourself, such as Microsoft 365 or Google Workspace, it can see which files are stored there and who they are shared with. Microsoft describes those three layers for its own CASB in the Defender for Cloud Apps documentation.
What comes out is a catalog of thousands of services, each with a risk score, and a switch per service: sanctioned or not. Through a proxy a CASB can then stop uploads, limit downloads to unmanaged devices and cut a session short. It is built for all cloud use, from storage to project tools, and AI is one category within that.
That is a lot, and it is work that governance of AI use does not do. If you have to stop uploads to unknown storage services, or map all SaaS use for a supplier register, you need a CASB and not an AI register.
What does a CASB not see?
The picture stops where the proxy stops. A remote worker without an agent, a contractor on their own laptop, a phone: that traffic never lands in a log. How large that share is varies per organisation, and nobody knows precisely, because it is the part that is not measured. Why every method has a blind spot of its own is covered in how do you find out which AI tools employees actually use.
A traffic log sees a domain, not an account. An employee who logs in to chatgpt.com with a private account that trains on conversations, and a colleague using the business account, are the same line in the log. That distinction decides a large part of the risk, and you cannot read it from the traffic.
And a CASB looks at sessions and files, not at the text someone types into a chat window. That the sentence with a citizen service number in it went to an AI tool is, to a traffic log, a request like any other. Reaching that moment takes something that sits in the text box; that is a different layer, worked out in Purview and BeeSensible.
What does governance of AI use add?
The decision, and the moment at which that decision is visible. A CASB knows two states per service: sanctioned or not. Governance of AI use knows four, and the most interesting one is "no decision yet". That status makes visible which tools are in use but have not been assessed by anyone. In a CASB those tools land in the same bucket as the tools that were deliberately unsanctioned.
The second is the place of the decision. A CASB steps in on the network, and the employee sees a blocked page or nothing. Governance of AI use shows the decision in the browser as soon as someone opens the tool, with the condition the organisation wrote itself, or with the approved alternative as the first button. Continuing is always possible. That keeps the use visible, because an employee who is not stopped has no reason to switch to a phone.
The third is the person. A proxy log contains a device or a user, because a CASB is built to act on a session. For the governance question that is not needed: "this tool was opened four hundred times this month and has no decision" is a complete answer. What that saves under the GDPR and with the works council is in AI monitoring versus employee monitoring.
| CASB | Governance of AI use | |
|---|---|---|
| What it sees | Traffic to cloud services, files in connected services | Which AI tools are opened in the browser |
| Where it looks | Proxy, firewall, agent on managed devices, API connections | The browser the extension is installed in |
| Remote work without proxy or agent | Not visible | Visible, if the extension is there |
| Phones | Not visible | Not visible |
| Private or business account | Cannot be told apart by domain | Cannot be told apart by domain; it is a question for the register |
| The text box | Out of reach | Out of reach for this layer; marking sensitive data is a separate module |
| Statuses per service | Sanctioned or not | Allowed, limited use, no decision yet, not allowed |
| Stepping in | Can block and limit sessions | Shows the decision and the alternative; never blocks |
| Link to a person | Device or user in the log | No user id; only the domain, at most once a day per device |
| Scope | All cloud use | AI tools only |
Two rows stand out. Neither sees the phone. And neither reads the account type from the domain; that is an agreement you record in the AI register and show at the moment of work.
What does this look like in practice?
A municipality has allowed Copilot and runs a CASB on the office network. The discovery report shows Copilot, ChatGPT and Perplexity. When the organisation then looks at what is opened from the browser, including at the remote workers, three specialised tools appear: a transcription tool for meetings, a translation tool and a service that summarises policy memos. Two of the three appear in no register.
The CASB did nothing wrong. The three tools were mostly used at home, outside the proxy, and for the tools that did appear in the report there was no decision to compare the report against. That is not an incident. It shows that a list of domains only becomes something once a decision sits next to each domain, and that the decision has to reach the remote workers too.
