Introduction
A deterministic outcome counts a session when a signal the session already carries says it happened — a tool succeeded, a score cleared a bar, a trace contained something.
The product states the trade-off plainly: a rule adds no cost and is only as strong as that signal. Where a signal exists, this is the route to use.
Where to find it
AI Governance > Outcomes & ROI > Outcome (New outcome when the list is empty)
Before you begin
Identify the signal — What in the session proves the result happened? A tool call is usually the strongest — an agent that created the ticket did the thing, and the tool result proves it.
Agree the value — With someone who can defend it. See Set Value per Outcome.
Define the outcome
Step 1: Name the result
Under Define, enter an Outcome name. Name the result, not the activity — Ticket resolved without escalation, not Ticket handled.
Step 2: Add an agent
Under 2. Measure outcome, click Add agent and pick the agent. For each agent, choose How it is verified and What it is worth — A set amount (every counted session books the same figure) or An automation prices it.
Step 3: Choose Rule
Set How it is verified to Rule.
Step 4: Build the condition
In 2. Measure outcome, use the When all / When any toggle, then add conditions.
If your signal is | Condition |
The agent completed an action in another system |
|
The agent avoided an action |
|
The result appears in what the tool returned |
|
Quality cleared a bar |
|
The reply said something specific |
|
The session was tagged |
|
wasn't called is worth knowing about. An outcome like Resolved without escalation is best expressed as the escalation tool not being called — counting an absence rather than a presence.
Step 5: Save
The source summarises as Counts when …. Read that line back — it is the rule in plain English, and it is where a wrong condition becomes obvious.
Combining conditions
When all — Every condition must hold. Use for a result with more than one requirement — the ticket was created and quality cleared a bar.
When any — One is enough. Use where an agent can reach the same result several ways.
Start narrow. An over-broad rule inflates the count, and an inflated count multiplied by a value is exactly the ROI claim that gets taken apart in review.
Verify it counts correctly
Before reporting anything, open a handful of sessions and check them against the rule by hand.
A session that should count — Confirm it did.
A session that should not — The more important test. An outcome counting things it should not is worse than one counting nothing.