Introduction
Every cost figure in AI Governance — on an agent, on a model, on an outcome, in Cost Explorer — comes from one place: the traces your agents produce.
Understanding that derivation matters more than knowing where a particular chart lives, because it answers the questions people actually ask. Why is this agent expensive. Why is there a governance line on my bill. Why does this number not match my provider invoice.
Where to find it
AI Governance > Cost Explorer
Access to AI Governance is granted through platform roles. Ask your workspace administrator, or see Settings > Governance > Roles for how permissions are assigned.
The derivation
Cost is not reported to UnifyApps by your providers. It is computed from what your agents did.
A span records a unit of work — One model call, one tool call. It carries what was consumed.
Spans roll up into a trace — One end-to-end agent run. Its cost is the sum of its spans.
Traces roll up into everything else — Sessions, agents, models, outcomes and any window you select are aggregations of the same trace data.
So the chain is span → trace → agent → report, and the whole of Cost Explorer sits at the far end of it. See Sessions, Traces & Spans.
"No traces, no cost. An agent that has never run is not free — it is unmeasured."
What spend decomposes into
Spend is broken into cost types, all derived from the same trace data. Group the Cost Explorer table by Cost Type to see the split, each with its total, its per-day rate and its share.
Cost type | What it is | Driven by |
| Model calls the agent made | Tokens in and out, per model |
| Tool invocations during a run | How often the agent reaches outside itself |
| Running your evaluators | How much traffic you score, and at what sampling rate |
| Running your controls | How many LLM-based checks you run, and on how many inputs |
| Spend that could not be assigned to a type | Activity whose trace data lacks what is needed to classify it |
Two of those deserve attention.
Eval and Governance are separate, and both are yours
Neither is a charge from UnifyApps. Both are the cost of model calls made on your behalf by things you configured.
Eval is what your evaluators cost: an LLM-judged evaluator makes a model call per scored record. Governance is what your controls cost: a custom LLM check makes a call per input it inspects. They are separated because they are tuned by different dials — sampling rate for the first, check design and applied-on scope for the second.
Eval also appears per agent as Eval spend on its Evaluation tab — the same figure, scoped to one agent.
Unattributed is the honesty line
Not everything can be classified. Unattributed is real spend that the trace data did not carry enough information to assign, and it is shown rather than absorbed into the other buckets.
Read its share before quoting any breakdown. A clean-looking split with a large unattributed slice is a less complete picture than it appears — and knowing that is more useful than a tidy chart that hides it.
Controlling what governance costs
Sample rather than score everything — Evaluators carry a sampling rate. Scoring every trace on a busy agent is rarely worth what it costs; a representative sample answers the same question.
Prefer a prebuilt check to a custom LLM check — A rule-based check costs nothing per input. An LLM check costs a model call every time it runs.
Prefer rule verification to an AI judge — The product says it plainly: a rule reads a signal the session already carries, at no added cost; an AI judge adds spend every session. Use the judge where no signal exists, not by default.
Narrow where checks apply — A control watching four points costs four times a control watching one.
Watch backfills — Evaluating past records is a one-off cost proportional to how many you select.
Reading cost figures honestly
Every figure is window-scoped — Two numbers from two windows are not comparable. Check the range before quoting one.
Cost is attributed to what produced it — The same underlying model usage appears against an agent when you look by agent, and against a model when you look by model. It is one set of spans read two ways, not two costs.
Coverage bounds the total — UnifyApps can only cost what it can see. An agent that is registered but not instrumented reports nothing, and its real spend is invisible here.
It will not equal your provider invoice — And should not. This measures governed agent activity in a window; an invoice measures everything on the account for a billing period. Reconcile the direction of change, not the absolute figure.
Refer: Spend on its own is half the picture. To weigh it against what the agents returned, see Outcomes & ROI.
Best Practices
Instrument before you budget — An uninstrumented agent contributes nothing to these figures and everything to your real bill.
Track spend per trace, not just total — Total spend rising with volume is expected. Spend per trace rising is a design change worth investigating.
Budget governance deliberately — Decide what proportion of spend you are willing to spend on knowing what your agents did. Then set sampling rates to match.
Review the expensive tail — Cost concentrates. A few agents usually account for most of it, and they are where optimisation pays.