Skip to main content

Catch spend drift before the invoice does

Enterprise Tier

Hard limits stop catastrophes; they do not catch the slower drifts. An agent framework that starts making ten calls where a human would have made one stays comfortably under every rate limit while it doubles the bill. The control for drift is observation on a cadence: a usage review short enough to be sustainable and regular enough that a surprise is at most a week or two old when it surfaces.

Persona: Platform operator working in the Admin Dashboard.

Estimated time: 10--15 minutes per review; weekly is typical, daily for environments with high turnover.

Outcomes​

By the end of this guide:

  • Usage Analytics is read on a regular cadence, with per-key attribution as the unit of analysis.
  • The by-key, by-user, by-model, and by-provider breakdowns are each mapped to the question they answer.
  • Anomalies found in the review have a follow-up path: drill-down, the audit trail, or the owning team.

Prerequisites​

  • Administrator access to the Admin Dashboard: typically the super_admin or billing_admin role.
  • API keys that follow the per-purpose convention established in Onboard developers and issue keys. A single organization-wide key that everyone uses makes per-purpose attribution effectively impossible.

Step 1: Run the review on a cadence​

  1. Open Usage Analytics in the Admin Dashboard on a regular cadence: weekly is typical, daily for environments with high turnover.
  2. Apply a time range that matches the budgeting cadence (last 7 days, last 30 days, last billing cycle).
  3. Switch to the By API Key breakdown.
  4. Sort by cost descending.
  5. Review the top consumers. The shape of the list should match the operator's mental model of Agent Router; surprises in this view are the most common signal that something is worth investigating.

Step 2: Read the breakdowns against their questions​

The breakdowns by user, by model, and by provider all support the same workflow at different levels of aggregation. By User answers "which team is spending the most"; By Provider answers "which contract is bearing the load"; By Model answers "which models are doing the actual work". A row that moved sharply between reviews is the drift signal this guide exists to catch: drill into the row to see the keys, models, and costs behind it, then take the finding to the owning team.

Both units of measure matter, and they answer different questions. Token totals describe the workload, how much a consumer is actually asking the models to do, and are stable across pricing changes. Cost translates that workload into money and is the unit finance thinks in.

Step 3: Route findings to the right follow-up​

For longer-horizon reporting (chargeback statements, quarterly reviews, or compliance attestations), the Export function returns the underlying data as CSV; the reporting workflow is covered in Bill AI spend back to the teams that incur it. For always-on visibility with alerting instead of a manual cadence, see Get alerted to cost spikes as they happen.

Where to go next

Get alerted to cost spikes as they happen

Replace dashboard-watching with alert rules in the observability stack.

Keep each team's spend inside an agreed ceiling

Turn the baselines read here into enforced budgets.