Is Your Team Using AI? What Admin Dashboards Actually Show

Asked whether the AI spend is working, most executives open the admin dashboard of whichever product they pay for and read out an active-user number. This page is about what that number is, and is not. We read the documentation for seven admin consoles, from OpenAI, GitHub, Anthropic, Microsoft, Google and Cursor, and catalogued 85 named metrics. All seven count seats and active users. No two define an active user the same way. None measures whether the work was correct.

This post is for the executive or IT lead choosing which numbers to put in front of a board. It sets out what every console counts, what only some count, and what none count; a table of twelve words that mean different things in different consoles; and the two new measures that reached dashboards in the week before this page's date. It does not argue that usage is the wrong question; our post on completed work as evidence makes that case and we do not repeat it here. Every number in every console is vendor-computed from vendor telemetry; nothing in this census has been audited by a third party. This blog runs on Anthropic's models, and Anthropic's consoles are read by the same rule as the others.

Three more counts sit between those. Six of seven report spend or credits; Google Workspace's Gemini reports are the exception, and their answer to heavy use is a licence upgrade rather than a cost figure. Three of seven measure agent actions rather than messages: GitHub's cloud-agent and third-party-agent fields, Anthropic's Cowork dispatch turns and actions, and Cursor's cloud-agent pull requests and bot automation level; Microsoft counts agent responses and sessions but not what the agent changed. And exactly one console reports a cycle time: GitHub's median minutes to merge, with Copilot-authored and Copilot-reviewed variants. Every other console stops at how much happened, not how fast.

One console asks the humans. OpenAI's impact survey is the only built-in instrument that collects a self-reported outcome, and OpenAI's own page says it is directional and not a causal measure of return. Microsoft's dashboard can ingest survey data, but from a separate product and only above a licence threshold.

One row per console, four questions: does it count seats and active users, does it count spend, does it classify the work, and does it measure what agents did. "Partly" means the console has a weaker form of the measure than the column asks for. The 85 underlying metrics, with their exact field names, windows and export paths, are in the research file this page is built from and summarised in the methodology.

This is the table the post exists for. Every row is a word that appears on more than one dashboard and means something different on each. The definitions are the vendors' own, condensed; the last column is why a board deck built from two consoles will not reconcile.

Two consoles changed in the days before this page's date, and both changes replace a volume count with a judgement. On September 16, OpenAI published a walkthrough of its Admin Console analytics: a Usage view of active users, credits and tokens across ChatGPT Work and Codex; an Insights view whose task classifier groups a sample of messages into use cases and tasks, with the share of credits by model, reasoning and speed setting, a plugin leaderboard and a skills view; and an Outcomes view with the Codex share of merged commits and lines of code. The post's worked example of return on investment is, in its own words, hypothetical, and we do not print its figures. Two limits matter: the classifier works on a sample, and OpenAI's help centre says the task insights cannot be exported.

On September 17, GitHub added feature engagement to the Copilot impact dashboard and its usage-metrics API. A user counts as engaged with a feature only when they used it on at least two distinct days in the 28-day window, across code completion, agent edits, passive and active code review, the cloud agent, the CLI and the app. GitHub's documentation adds two cautions: the per-feature counts overlap and must not be summed, and a user who chats with Copilot every day but never produces a completion or an agent edit is not in the adoption cohort at all. The same day GitHub put agentic CLI customisations into the API, with top-five counts for skills, custom agents, MCP servers, slash commands and plugins, and the customer-defined names hidden.

How to read them: OpenAI's classifier answers what the work was; GitHub's engagement bar answers who actually adopted a feature rather than who touched it once. Both are more useful than a message count and both are fenced by their own vendors. Neither says whether the work was any good, which is the question our post on Anthropic's oversight metrics approaches from the supervision side.

Every item below exists as a named metric in at least one console, so it can be sourced rather than estimated. The rule that makes the list honest is the last one.

1. Seats assigned against seats active in the last 28 days, with one window across vendors and the window stated.
2. Adoption rate with its denominator printed. Anthropic ships it as a field; everyone else needs the division done by hand.
3. Returning users. Microsoft’s “returning users”, active in the prior and the current 28-day period, is the only named retention metric in the census; rebuild the equivalent elsewhere from per-user daily rows.
4. Depth, not only breadth: GitHub’s two-days-in-28 engagement bar or Microsoft’s habitual and power bands, with the bar named.
5. Spend per active user, available from Cursor, Anthropic, OpenAI and, with its caveat, GitHub. Not available from Google Workspace.
6. AI-attributed share of merged code with the attribution method beside it. Never the percentage alone. The harness matters as much as the model, as our post on harness cost showed, and our census of what companies disclose about their own AI share shows how far definitions drift.
7. One cycle-time number. GitHub ships median minutes to merge. If you are not on GitHub, this row comes from your own version control, not the AI vendor.
8. One thing the dashboard cannot tell you, named explicitly: whether the work was correct. No console in this census measures it, and a board that hears that sentence once will ask the right question next quarter. If you want the deck built, our analytics service starts from these eight rows.