Three of Four Platforms Can Capture Your Prompts, and the UI Never Tells You
The audit trail that protects the buyer is, by construction, the most detailed per-employee work record your company has ever held. Everything the rest of this section builds (the audit trail property, the four-axis query, the per-actor attribution that frameworks reward) reads differently from the other side of the desk: you, the developer being governed, not the buyer assembling the audit plane. So the question you keep asking on vendor forums, "does the admin see what I'm working on?", deserves a straight answer.
The answer on all four major platforms is the same: not by default. On three of the four (all but GitHub Copilot), full prompt capture is one employer setting away, and the tool's UI never changes when the setting is flipped. You cannot tell from inside the tool whether your org turned it on. Only your employer's disclosure answers the question, and no vendor publishes a developer-addressed "what your admin sees" page.
Copilot stops at metadata; the other three can capture prompts
Last verified: June 2026. Vendor docs churn fast; three of the GitHub citations below are 2026 changelogs, so treat this as a snapshot.
| Platform | What the admin sees per developer | Prompt content visible? | Developer-facing disclosure |
|---|---|---|---|
| GitHub Copilot (Business/Enterprise) | Per-user usage metrics: engagement, acceptance counts, request and token counts, per-user CLI sessions, 1-day and 28-day reports | No. Metrics are counts and rates; admins have asked for per-user prompt data and it isn't offered | Metrics schema is public docs; no in-product "you are being measured" notice |
| Cursor (Teams/Enterprise) | Admin analytics per user: tab completions accepted, agent lines accepted, chat message counts, AI-attributed lines committed, active days | Not via the dashboard. But Enterprise hooks can forward prompt and codegen events to the employer's own infrastructure, invisible in the UI | Docs exist; forum threads show developers can't tell what admins see |
| Claude Code (enterprise OTel) | Whatever the employer's collector keeps. Default telemetry: per-session metrics with user.email, session.id, prompt length, tool-call counts, cost. Managed settings "cannot be overridden by users" | Off by default; one env var away. Prompts are <REDACTED> unless OTEL_LOG_USER_PROMPTS=1; further flags escalate to full conversation capture. Data goes to the employer, never to Anthropic | Flags are publicly documented, but you cannot see from inside the tool whether your org enabled them |
| OpenAI Codex / ChatGPT Enterprise | Analytics dashboard: daily active users, threads, turns, usage per user | Yes, via the Compliance API, which "returns every user input, output, and system-generated message"; Microsoft Purview displays prompts to compliance roles | The clearest of the four: OpenAI's help pages separate analytics (counts) from compliance (content). Whether you're told it's running is up to your employer |
Privacy mode protects you from the vendor, not from your employer
These are two different switches, and forum threads routinely conflate them. Cursor's Privacy Mode governs whether Cursor and model providers retain or train on your code; it says nothing about what your employer's dashboard or hooks see. Claude Code is the sharpest case: its telemetry goes to your employer's own collector, never to Anthropic, which makes your employer, not the vendor, the controller of that data.
The enforcement signal is telemetry, not output quality
The mandate trend is real, and the enforcement signal is telemetry. Microsoft's June 2025 memo made AI usage "core to every role and every level," with usage weighed in performance reflections. Shopify put AI questions into performance and peer reviews. Coinbase fired engineers who hadn't onboarded to Copilot or Cursor by an end-of-week deadline; the trigger was onboarding telemetry. One engineer's account on Hacker News: "they look at your Bedrock and LLM API calls as well as Claude Code history."
Concretely, the measured set is: license activation, active days and session counts, acceptance counts and AI-attributed lines, API-call volume, and review-cycle "AI nativeness" ratings. All of it is producible from the table above. See Mandates and measurement for the management side of this.
Whether any of this is lawful depends on where you sit. In Germany, a system merely capable of monitoring you triggers works-council codetermination before rollout, and template agreements bar individual-level evaluation outright; France's CNIL fined Amazon €32M for per-event worker telemetry of exactly this shape. In the US, constraints are mostly notice-based. Works councils and monitoring law covers the legal detail. One gap: no regulator has ruled on coding-assistant telemetry specifically; the analogues are inference, not precedent.
The same trail that worries you is the evidence that protects you
A per-actor record showing the AI authored the change, a separate gate approved it, and your instruction was X is exactly what answers "who broke prod" without throwing you under the bus. That is the audit trail property and binding to the human working in your favor: attribution that names the agent and the approver, not just you. The trail is a threat only when it is secret, individually scored, and purpose-unbound. Those three properties are exactly what the rest of this section is built to remove.
Five questions force the disclosure no vendor will write
The field has no developer-addressed disclosure standard, so make your org write one. Ask:
- Is prompt content captured, or only metadata (length, counts, cost)? Under what setting, and who approved it?
- Can managers query per-actor, or only team aggregates? GitHub already suppresses aggregates below 5 members; the SPACE framework's first guardrail is "review all metrics at the team level, never individual."
- Is performance evaluation a declared purpose of the trail, or is it audit-only with individual access as logged break-glass?
- Is every per-actor query against the trail itself an audit event you can see? No vendor ships this today; it's this handbook's ask.
- What are the retention periods, per purpose?
If your org can publish its answers, you have the disclosure no vendor does. If it can't, you've learned what the trail is actually for.