Introduction
Grouping does not change how much you spent. It changes who you can hold responsible for it — and the choice of dimension is really a choice about which conversation you intend to have.
Where to find it
AI Governance > Cost Explorer. The grouping control sits below the chart; the Agent / Model toggle above it already splits the data one way before you group at all.
Two decisions, in order
Which lens: agent or model — Agent attributes spend to the thing doing the work; model to the thing being called. Same money, two indexes — see Agent View vs Model View.
Which grouping within that lens — How the chosen total is divided.
Getting the first wrong makes the second pointless. If the question is "should we renegotiate with this provider", no grouping of the agent view will answer it.
Pick the dimension from the question
The question | Group by | Because the answer names |
"Who should pay for this?" |
| A team or a person with a budget |
"Where is the money going inside an agent?" |
| LLM, tools, evaluation or governance spend |
"What are we spending with this provider?" |
| A commercial relationship |
"Which part of the business consumes AI?" | Filter by | An organisational unit — if the field is populated |
Group by offers Agent, Owner and Cost Type in the Agent view, and Model, Provider and Owner in the Model view. Department is a filter in the Agent view, not a grouping, and there is no Workspace grouping or filter.
Note: Filtering by department only works if somebody filled the field in. Department is optional at registration and cannot be edited afterwards, so a department filter across a registry that never populated it finds little. Set it when you register — see Ownership, Workspace & Department.
Filtering versus grouping
They answer different shapes of question, and reaching for the wrong one produces a chart that cannot say what you want:
Group when you want to compare parts of a whole — "which of these costs most". The total stays the same and gets divided.
Filter when you want to study one part in isolation — "how has this one behaved". The total shrinks to that subset.
The pairing that produces most findings is filter to a period of interest, then group by agent. That isolates when and reveals who in one step.
The part no grouping shows
Every breakdown covers the spend it can attribute. Spend that belongs to no agent and no outcome is still real money, and it will not appear in any slice.
So a grouping that sums to less than you expect is not necessarily wrong. It may be telling you how much of your spend is attributable at all — which is a finding in itself, and the same caveat that makes a return-on-spend figure only as good as its measured spend. See Reading the ROI View.
"Grouping tells you how the money you can see divides up. It says nothing about the money you cannot."
Two habits worth adopting
Group before you conclude — A total that rose is not yet a finding. A total that rose because one agent tripled is.
Say which grouping produced a number you quote — "AI spend was X" is ambiguous between lenses and between filters. A figure without its slice is not reproducible, and someone will try to reproduce it.