answers.delivery

How-to · 3 min read

How to find out which AI tools your staff actually use

If you run security or IT, you probably suspect people use far more AI tools than the two your company signed up for, but you cannot name them. Someone asks for an AI policy, and the first honest question is which services are in use today and on which machines. You would like a real list before you write rules. And you would rather not ship that list off to yet another vendor's cloud to get it.

Published
17 Nov 2026
Delivery
No. 12
Topic
AI and agents
Evidence
10 claims quoted from milvago.ai

01
Why it's hard

Surveys and policy acknowledgements only tell you what people are willing to admit. Firewall and proxy logs show host names, but plenty of AI services sit behind generic domains, and remote laptops off the VPN never pass through your network at all. Cloud gateways can read that traffic, but usually by intercepting TLS, and the usage record then lives with the gateway vendor. Banning everything feels safe on paper, yet it mostly pushes the same use somewhere you cannot see.

This part describes the usual way. It makes no claims about Milvago AI, so nothing is quoted here.

02
How Milvago AI does it

Milvago AI puts a browser extension and a light agent on each workstation and keeps the record on a server you host. The extension spots visits to more than 160 AI platforms by address and recognises up to 9 services inside the page, so you see which ones are used and from which machines. Nothing is decrypted in the browser and requests still go straight to the AI service. By default each event stores facts such as time, service and workstation, not the text people typed.

The extension reports visits to 160+ AI platforms by their address and recognises up to 9 services inside the page.Quoted from milvago.ai
For each platform you see the workstations that reached it and how many OS accounts were behind the visits.Quoted from milvago.ai
Self-hosted editions never send your usage data to Milvago, and send no telemetry unless an organisation opts in.Quoted from milvago.ai

03
Step by step

  1. 1Read the free Community edition's docs and decide where the server will run, since it lives on your own infrastructure.
  2. 2Deploy the server in containers with PostgreSQL and connect your own OpenID Connect identity provider.
  3. 3Roll out the agent to Windows or Linux workstations and the extension to the browsers your teams use.
  4. 4Leave every service on observe for a while so you collect a picture of real use before blocking anything.
  5. 5Open the usage map and the activity log to see which AI platforms were reached and from which workstations.
  6. 6Turn what you found into rules per service, such as allow, redirect to the approved tool, mask or block.

04
Tips

Click around the live demo before you install anything. It runs on synthetic data, so you can see what the usage view looks like without touching your own fleet.

The live demo is a real Milvago instance with synthetic data.Quoted from milvago.ai

Know what Community covers. It recognises ChatGPT and Claude in depth inside the page, and every edition still detects the 160+ other platforms by address. Desktop tools like Claude Code and Codex, plus installed AI apps per machine, need Enterprise.

Enterprise adds the blocked attempts, and the AI tools installed on each workstation.Quoted from milvago.ai

Plan for macOS separately. The agent is listed for Windows and Linux, so Mac users will not show up in the workstation view yet.

An agent for Windows and Linux and an extension for 6 browsers recognise up to 9 AI servicesQuoted from milvago.ai

Tell staff what is being recorded before rollout. Saying plainly that the tool logs which service was used and when, and not what they typed, makes the first discovery report far less tense. People are much more likely to keep using approved tools in the open when they know the rules and the limits of what is watched.

By default an event records facts, not the text of prompts.Quoted from milvago.ai

05
Questions

Does it need a proxy or TLS interception?

No. Requests go straight from the browser to the AI service, with no proxy in your network and nothing decrypted in the browser.

Requests go straight from the browser to the AI service, with no proxy in your network and nothing decrypted in the browser.Quoted from milvago.ai

Does it replace our CASB or DLP?

No. It works on the workstation and is meant to sit next to your network security tools, not replace them.

Milvago complements a network security stack; it does not replace it.Quoted from milvago.ai

How is Enterprise priced?

Per workstation, on quote. There is a 30-day trial before you commit.

Enterprise is licensed per workstation, on quote, with a 30-day trial.Quoted from milvago.ai

More from Milvago AI

All 2 answers