What if two agents contribute to one outcome?

Add both with Add agent, each with its own verification and value. The Value by score table then shows the split.

Does an AI judge cost money?

Yes, per session. It is model work and appears in governance spend.

Why does an agent show no outcomes?

Either no session met its condition, or the condition reads a signal the sessions do not carry. Check a session manually against the rule.

Can an outcome be negative?

What it is worth is set per agent on the outcome. If the spend exceeds it, the verdict reads Net negative — which is a legitimate and useful finding.

Introduction

Cost Explorer tells you what your agents spend. Outcomes tells you what they returned, in the same units, so the two can be subtracted.

It is the part of AI governance nobody documents, and the only part your CFO will ask about.

Where to find it

AI Governance > Outcomes & ROI

Access to AI Governance is granted through platform roles. Ask your workspace administrator, or see Settings > Governance > Roles for how permissions are assigned.

What an outcome is

A named result worth a specific amount of money each time it happens. "Track the value your agents deliver and measure it against what they cost to run."

Defining one takes two halves: Define the result, then Measure it — add the agents that produce it, how each is verified, and what each counted session is worth.

Define

The outcome itself is a name. Value is set per agent, in each agent's What it is worth card.

Field	What it means
`Outcome name`	"Name the result, not the activity" — Lead marked qualified, not Lead handled.
`What it is worth` (per agent)	`A set amount` — every counted session books the same figure — or `An automation prices it`.

The naming rule is doing real work. An outcome named for an activity counts effort; an outcome named for a result counts value. Only the second can be multiplied by a number and defended.

There is no Owner field; the outcome shows the person who created it. The value is still an assumption someone in finance should be willing to stand behind, not a number a platform team invented.

Measure

A source is an agent credited with producing the outcome. Each source needs its own check for when a session counts, because two agents can reach the same result differently.

Three ways to verify — the How it is verified field

Method	How it works	Trade-off
`Rule`	Reads a signal the session already carries.	No added cost, only as strong as that signal.
`AI judge`	A model scores the session against your criterion.	Adds spend every session; captures what no signal records.
`Automation`	Reads it from the system of record.	Free; as complete as the system that records it.

Prefer a rule where a signal exists. Reach for the judge where the result is real but nothing in the session records it — which is common for quality-shaped outcomes.

Rules: when a session counts

A rule counts a session When all or When any of its conditions hold, built from three things a session carries: a Tool, an Evaluator, or the Trace itself.

Tool conditions — Whether a tool succeeded, failed, was called or wasn't called, and whether its result contains something.
Evaluator conditions — Whether a score is at least or is at most a value, or is present — the bridge between quality and value.
Trace conditions — On Output text, Input text or a Trace tag.

A tool call is usually the strongest available signal: an agent that actually created the ticket did the thing, and the tool result proves it.

The AI judge

Takes a judge model (text-generation models only), an evaluation criterion in plain language, and a Pass score threshold — a slider from 0 to 1, default 0.85.

Remember it charges per session. An AI judge on a high-volume agent is a standing cost, and it shows up in governance spend.

Automation

Pick a Recorded by automation built on the outcome callable interface. Optionally tick Also record the outcome value from this automation so the automation supplies the value as well as the count.

Reading the return

Figure	Means
`Total outcomes`	How many times it happened in the window.
`Total value`	Counted sessions × what each is worth.
`Total spend`	What the crediting agents cost.
`Net value`	Value minus spend.
`Return on spend`	Value ÷ spend, as a multiple.
`Verdict`	`Net positive`, `Break even` or `Net negative`.

The Value by score table breaks the total down by agent — Source, Return, Owner, Total value, Total spend, Net value — which is how you find out that one agent carries an outcome and another was never really contributing.

The honest caveat: unmeasured spend

Return on spend divides value by tracked spend — the spend of agents credited to outcomes. Everything else is unmeasured.

The Unmeasured spend tile sits beside Total spend precisely so the figure is not mistaken for all spend. A five-times return on a fifth of your estate is a much weaker claim than a five-times return on all of it, and the label is your reminder to check which you are looking at.

Compare it against total spend in Cost Explorer over the same window. The difference is the spend no outcome is claiming, and it belongs in any ROI figure you report upward.

"Return on spend answers a question about the agents you measured. Unmeasured spend tells you how much of the estate that question ignored."

Reporting both is what makes the figure credible to a finance reviewer, and quoting the multiple alone is how ROI claims get taken apart.

On the agent

Each agent's Outcomes tab shows a Returns card and a Value by outcome table. Rows offer View and Edit only — outcomes are shared across agents, so they are deleted from Outcomes & ROI, not from the agent.

Best Practices

Name results, not activities — If the name describes work rather than a result, the value figure will not survive scrutiny.
Get the value agreed, not estimated — Have someone who can defend it agree the figure before you enter it in each agent's What it is worth card.
Prefer rules to judges — Free, deterministic, auditable. Use a judge only where no signal exists.
Start with one outcome — The one your agents were funded to produce. Breadth after credibility.
Report tracked spend against total spend — Volunteering the gap is what makes the multiple believable.
Review sources when agents change — A rule reading a tool that has been renamed silently stops counting.