What Proxon captures
The three capture levels, what each one includes, and how an admin turns capture up or down.
Proxon measures how AI tools are used across your organization by observing activity on the surfaces you connect. This page explains what that observation can include, so you and your security team can decide whether to keep the default or lower it.
Proxon collects data at one of three levels, and an admin chooses a single level that applies to every observer. Each level includes everything in the level below it, so more detail builds on less. The most detailed level, Content aware, is the default, and an admin can lower it at any time.
Metadata only
Which AI tools each person uses, and when and how often. Proxon captures no content of any kind at this level, so nothing that was typed, named, or copied. It records the site someone visited, but never the specific page.
Proxon never records which keys you press, the contents of what you copy or paste, or the individual commands you run, at this level or any other. The Windows Desktop Observer counts how many keys and clicks happen in each short interval, at every capture level including Metadata only, which shows how actively the device is in use. That is volume and rhythm only, never which keys, where you clicked, or any text you type. The macOS Desktop Observer does not count keys or clicks at all.
Behavioral data
Everything in Metadata only, plus the shape of each interaction. That includes how big a prompt, file, or paste was, which tools ran and for how long, and which provider it went to. It still captures none of the words themselves, so no prompts, replies, titles, or page addresses.
Content aware
Everything in Behavioral data, plus the text itself. That includes prompts, replies, the model's reasoning, document and tab titles, full page addresses, and what tools were given and returned. A tool result can include the contents of a file, up to a size limit. Document and tab titles and full page addresses sit at this level because they can reveal the substance of the work, such as a customer or file name, not just the fact that AI was used. The observers capture the words you type into AI tools, such as your prompts, only at the Content aware level, and only on the supported AI chat sites and apps listed below.
Content aware is on by default. The one setting applies to all three observers at once. Admins set it in Data Collection under org settings and can change it at any time.
What each observer adds at the Content aware level
The client-side observers watch different places, so each one contributes something different at the Content aware level.
| Observer | What the Content aware level adds |
|---|---|
| Browser | The full page address and tab title, and the prompt text on supported AI chat sites |
| Desktop | The title of the foreground window, which can include a document name |
| Network | The prompts, the replies, the model's reasoning, and what tools were given and returned, for supported apps |
The other sources, meaning AI vendor integrations, the Capture Proxy, the Proxon SDK for Agents, and OpenTelemetry, capture whatever the connected account or instrumented system exposes. Each one's own page documents the specifics.
Microsoft 365 Copilot sensitivity labels
In Microsoft 365 Copilot Agent Builder activity, the tool details Copilot sends can carry the Microsoft Purview sensitivity label of the message. At the Content aware level the Network Observer keeps those tool details like any other, so it keeps the label too. That includes the label's display name and id, not only its priority. Label names can be revealing, such as a project codename or a legal matter, so treat them as content.
Below Content aware, Proxon never keeps the label's name or id. At most it keeps two plain values about the label: whether it encrypts the content, and its priority number.
Proxon does not remove label names on its own. If your organization does not want them kept, keep capture below Content aware.
How Proxon protects content
Content aware is the default because Proxon's deeper insights, such as where AI spend goes to waste, need to see the work itself and not only its size and shape. These safeguards apply to what it captures.
Secrets are removed on the device. Before captured text leaves a Mac or Windows machine, the Network Observer strips credentials out of it. That covers API keys and access tokens from the major AI and cloud providers, passwords, private keys, authorization headers and cookies, and secret lines in configuration and .env files. It happens at every capture level, so a secret pasted into a prompt or returned by a tool is removed before Proxon receives it. The Browser Observer removes AI provider API keys and authorization headers from the prompts it captures.
Only AI services are inspected. On the network, Proxon looks only at calls to known AI services, not your general traffic. Other traffic, such as regular browsing and email, passes by untouched, and apps that encrypt their traffic end to end cannot be inspected.
Lowering the level takes effect within minutes. When an admin lowers the capture level, devices pick up the change within about five minutes and stop capturing content from then on. Content captured before the change is kept until it ages out.
Content is deleted after 13 months. Proxon keeps captured content for up to 13 months and then deletes it automatically.
What Proxon keeps in order to work out the plan
Knowing what a call cost is not the same as knowing who paid for it. Two things Proxon keeps are there to answer the second question, and neither one depends on the capture level.
From Claude Code over OpenTelemetry. Claude Code stamps its own identity on every record it sends, and Proxon keeps those fields: the signed-in email address, Claude Code's own user id, the Anthropic account and organization ids, and the session id. They arrive with the telemetry rather than from an observer, so they are kept at every capture level. Prompt and reply text over this path is a separate switch on the tool itself, and it stays off unless you turn it on. Other tools' OpenTelemetry exporters are read for model and token counts only. See OpenTelemetry for the list and what the ids are used for.
From the Claude config file on the machine. The Network Observer reads the account section of the local Claude config file to learn which Claude plan the person is on. That section holds the plan along with the account's name and email address, and Proxon reads nothing else from the file. It happens at every capture level and even when network capture is switched off, because without it Proxon cannot tell whether a Claude Code call was covered by a subscription somebody already pays for or billed as cash.