The IT admin at a housing association gets a question from the data protection officer: which AI tools are in use here? He opens the secure web gateway logs and finds chatgpt.com, copilot.microsoft.com and a domain he does not recognise. He sees nothing of the forty colleagues working from home today, because the proxy sits on the office network.
That is the whole comparison in one scene. A network proxy sees the traffic of every device routed through it, to every service, and can stop a connection. A browser extension sees only the browser it is installed in, but it travels with the laptop and sits in the text box at the moment someone opens an AI tool. For network security the proxy remains necessary. For the question of which AI tools are in use and whether staff know what has been decided about them, the extension is better placed.
What does a network proxy see, and what does it miss?
A proxy, a secure web gateway or a DNS filter sits between the device and the internet. Everything that leaves through the office network or an always-on VPN passes through it: every browser, every native app, every script. That is its strength. If you want to know whether a managed device opened a connection to a malware domain, this is where you look.
With encrypted traffic, the proxy sees the domain and the volume, not the content. Seeing the content takes TLS inspection: breaking the traffic open, reading it and encrypting it again. That means a certificate on every device, extra capacity for all traffic and the question of whether you also want banking, health portals and the GP's website to be readable. Many organisations switch inspection off for large categories, and then the domain is all that remains.
A domain tells you less than you might think. chatgpt.com is the same domain for a free account and for a business account with a data processing agreement. An AI feature inside a service you already use has no domain of its own. And the proxy only sees the prompt when it leaves the device, by which time it has been written and sent. A proxy can stop the connection; it cannot help with what goes into it.
What a proxy log always does contain is the device or account the traffic came from. That ties every line in the log to a person, which turns it into processing of personal data and a matter for the works council. Why that is also a measurement problem for governance is covered in AI monitoring versus employee monitoring. When an employee asks whether their employer can see what they do in ChatGPT, a log like this is part of the answer.
What does a browser extension see, and what does it miss?
An extension lives in the browser and sees what happens there: which page opens and, in supported apps, what is typed into the text box. It goes home with the laptop, onto the hotspot on the train and onto the client's wifi, because it sits on the device rather than on the network.
What it does not see is everything outside that browser. A second browser it is not installed in. Native apps. Phones. A script calling an AI service through an API. Devices the organisation does not manage. Covering those brings you back to the proxy or to managed devices.
What an extension records is up to its maker. It can be the entire browsing history, in which case the extension is worse than a proxy. It can also be only the domain of an opened AI tool, with no user ID. Ask before you roll anything out.
| Visibility of | Network proxy | Browser extension |
|---|---|---|
| Remote work without a VPN | No, unless it is a cloud proxy with an agent on each device | Yes, it travels with the device |
| Phones | Only on office wifi or as a managed device | No |
| Other browsers and native apps | Yes | No, only the browser it lives in |
| Encrypted traffic | The domain; content only with TLS inspection | The page itself, with nothing to decrypt |
| Free or business account | Same domain, no difference | Same domain, no difference |
| The text box, before sending | No | Yes, in supported apps |
| Can stop a connection | Yes | Depends on the extension; BeeSensible does not do so of its own accord, but the organisation can set it so that staff cannot continue to a tool that is not allowed |
| What gets recorded | Per device or account | Depends on the extension; with BeeSensible the domain, at most once a day per device, no user ID |
| Rollout | Network configuration, plus certificates on every device for inspection | Centrally, through a browser management policy |
Why does the text box make the difference?
Governance of everyday AI use turns on a decision per tool and on whether people know that decision at the moment it counts. That moment is opening the tool and typing the prompt, and it happens in the browser. It is the same shift we describe in our article on the browser as a security blind spot: work has moved into the browser, and an after-the-fact log only describes what has already happened.
At that moment a proxy can do two things: allow or stop. An extension can do something else: show what the organisation has decided, offer the approved alternative and highlight sensitive data as it is typed. That is not security in the classic sense. It is where a policy moves off paper and into the working day. How that fits into the full cycle is explained in our main article on AI governance and on the AI governance topic page.
When is a proxy the better choice?
More often than extension vendors like to admit.
When the question is network security: malware, phishing domains, connections to attacker infrastructure, categories that must be blocked for everyone. When you need to cover native apps, scripts and API traffic to AI services, for instance developers with an API key. When your devices always run through a managed network or VPN and the logs already exist. And when a regulator or a contract requires a specific service to be technically unreachable, because then a notice is not enough. In that last case, bear in mind what blocking does to the use itself: it moves to the phone.
In none of those cases does a proxy tell you whether anything has been decided about a tool. That lives in a register, not in a log. Which other sources each show part of the picture is covered in our guide to discovering shadow AI.
What does this look like at a housing association?
The admin from the first paragraph rolls out an extension through Intune and, a month later, puts the two lists side by side. The proxy saw chatgpt.com and Copilot. The extension also sees Perplexity, a transcription tool the tenant liaison team uses for meeting notes and a service that rewrites letters in plain language. Two of them are not in the register at all.
That is not an incident, and the proxy was not wrong: it described the office. The list from the browser describes the work, including the days at home. The gap between that list and the register is the backlog: three tools that need a decision.
BeeSensible's browser extension, for Chrome and Edge, is not a proxy. The extension records only the domain of an opened AI tool, at most once a day per device, with no user ID and no page content, and checks it against a catalogue of over 1,000 assessed tools. An admin sees in the Inbox which tools still have no decision and records one. Rollout is central, through Microsoft Intune, and takes less than an hour.
