Introduction
Cost Explorer answers one question — where is the money going — from the same traces everything else in AI Governance reads. It is not a billing system and it is not your invoice; it is an attribution view, and the difference matters when the numbers do not match.
Where to find it
AI Governance > Cost Explorer.
The default range is the last 15 days, with presets from Today to Lifetime. Search, Filter (Owner, Platform, Department and the numeric columns) and Sort are available.
View | Columns |
| Agent, Total Spend, LLM, Tools, Eval, Governance, Share, Traces, Owner |
| Model, Platform, Total Spend, Share, Spans, Owner |
Note: Cost Explorer has no budgets or alerts. For spend caps and alerts, see
AI Setup>Spending.
What the view gives you
Element | Answers |
| Whose spend is this — the thing doing the work, or the thing being called? |
| What did the window cost in total? |
| What is this costing per day? The window total ÷ the days in the window. |
| What does one run cost on average? |
| If nothing changes, what does the month look like? The daily rate × 30. |
Spend over time | Is it steady, growing, or spiking? |
| Which agent, model or dimension is responsible? |
Each tile and each row shows its % change against the previous window of the same length.
Spend per trace is the number worth watching
Total spend rises when you succeed, so it is a poor health signal on its own. Spend per trace is the one that separates growth from waste, because it holds volume constant.
Total spend | Spend per trace | Reading |
Up | Flat | You are doing more of the same. Usually good |
Up | Up | Each run is getting more expensive. Investigate |
Flat | Up | Volume fell and unit cost rose. Investigate urgently |
Down | Up | Fewer, costlier runs — often retries on a failing path |
When spend per trace rises, the causes are few and findable: a longer context, more tool calls per run, retries, or a switch to a more expensive model. All four are visible in an individual trace. See Reading a Trace.
Read the projection as arithmetic, not a forecast
A 30-day projection extrapolates the current rate. It knows nothing about your roadmap, your seasonality, or the pilot that ends next week.
Its value is as a scale check — turning a per-day figure into one a budget holder recognises — and as an early warning: a projection that jumps sharply means the recent rate changed, and that is worth understanding whether or not the projection turns out to be right.
"The projection is not a prediction. It is today's rate, multiplied, so you notice today's rate."
Why this will not match your provider bill
Cost Explorer derives spend from observed token usage and applied rates. Your provider's invoice is computed by the provider, from their own metering, under your commercial terms. Several things legitimately differ:
Uninstrumented traffic is invisible here. Anything not passing through a traced agent produces no spend in this view and appears on the bill.
Commercial terms are not reflected. Committed-use discounts, negotiated rates and credits are between you and the provider.
Some provider-side pricing is not visible here — batch pricing and similar mechanisms change what you are actually charged. Caching is reflected for UnifyApps-traced calls: cached tokens are priced separately (cache read / cache write).
Periods differ. A billing month and a rolling window are not the same range.
Use Cost Explorer for attribution and trend; use the invoice for the total. They are answering different questions and neither is wrong. See How Cost & Tokens Are Calculated.
It can only see instrumented traffic
Every figure here comes from traces. An agent in Setup Pending contributes nothing, and traffic that is never traced contributes nothing at all — so a reassuringly small number can mean low spend or low coverage, and those look identical from this screen. Spend from traced agents that are not registered does appear, as Unregistered agent (<id>).
Before treating a total as your AI spend, check how much of your estate is actually registered and sending. That is a registry question, not a cost one.