Scenario: rolling out AI spend visibility

A team is spending real money on AI tools and nobody can say what it buys. This guide is the rollout sequence that fixes that: metering on, identities clean, attribution working, and (the part that decides whether the rest sticks) the conversation with the people whose usage is being metered. It links out to the setup and concept docs rather than repeating them.

1. Tell the team before you turn anything on. Metering is a change to how people's tools are configured, so it lands better announced than discovered. Say what is measured (requests: cost, tokens, model, who), what is not (no source code, no prompts in reports), who sees what (org and team views for admins; each person can see their own usage), and why (budgeting and tooling decisions). The governance model is the reference for the privacy stance: it's easier to point at a stated posture than to improvise one in a team channel.

2. Get the foundation right. Attribution rides on people and identities: every person added, identities linked, and (if you want the AI Tax) hourly rates in place, since the metric needs an engineering cost to divide by.

3. Enable metering and connect the team. Set up AI usage metering: an admin enables it in Settings, everyone receives their personal access URL and per-tool instructions by email, and each connection is one base-URL change. Usage starts appearing within minutes of the first request, so early adopters validate the pipeline before the long tail connects.

4. Give attribution its evidence. Sessions attach to tickets through ticket references: keys in branch names, commit messages and the session itself. Teams with branch-per-issue habits get strong attribution immediately; teams without will see a large unticketed share at first. That's the number telling you about the habit, not a reason to distrust the metering: how attribution works explains the buckets.

5. Read the first reports for hygiene, not verdicts. After a week or two, the spend report shows where the money goes by model, team and person; the attribution view shows how much of it ties to delivery. Expect the first read to surface gaps: people not yet connected, unlinked identities, a high unticketed share. Fix those before drawing any conclusion about the spend itself.

6. Close the loop with the team. Share the first readout, including what's unattributed and why. Reports that show their own gaps are easier to trust, and the people closest to the work are the fastest route to fixing attribution habits.

7. Then decide about controls. Once a full period of trusted data exists, budgets and alerts are decisions rather than guesses: budgets, limits and enforcement covers that step, starting with alert-only rules.

What good looks like

A settled rollout has every person's tools routed through their access URL, an attributed share trending up as ticket-reference habits bed in, an unticketed remainder the team can explain, and the AI Tax reading as a routine line item beside labour cost. The product framing for the pillar is on the AI Spend overview; the honest-numbers posture the rollout leans on is on the AI Governance overview.