Four Questions. Four Sprints. One Stack.

Your AI agents are running. They're billing. The governance platforms tell you which ones are active and whether their security posture is clean.

What they don't tell you is whether any of it is producing economic value.

That gap is the one PromptKing has been closing since June 4. Four sprints. Four questions. One stack.

Question one: what do your agents actually produce?

Not how many tokens. Not how many sessions. What verified business outcome did each agent produce, and what did it cost to produce it?

The Outcome Cost Ratio answers that. Total agent session cost divided by verified outcomes. Five categories: Resolution, Generation, Transformation, Decision Support, and Orchestration. Each with its own verification method, its own confidence score, and its own benchmark band.

The market benchmark for a Resolution agent is $1.50 per closed ticket — anchored to Zendesk, Fin, and HubSpot's published pricing. An agent running at $4.80 per closed ticket isn't a cost problem. It's a structural problem. Wrong model, wrong prompt architecture, or wrong task assignment. The number surfaces that before it shows up as a surprise on the invoice.

FOCUS 1.4 has zero outcome columns. PromptKing is proposing OutcomeType, OutcomeCount, and CostPerOutcome for FOCUS 1.5 — the first attempt to make outcome economics part of an open standard.

Question two: can you trust the number?

A metric without integrity rules is an inflatable KPI.

If a bot writes a file that gets immediately overwritten, that is not an outcome. If the same agent produces the same output twice within 30 minutes, that is one outcome with a duplicate entry. If a session costs $0.009, it is below the minimum cost threshold and should be suppressed from the calculation.

Three verification states: auto_verified, system_verified, human_verified. Decision Support is always proxy — analysis delivered to a human is an indirect signal regardless of what the downstream data shows. It never surfaces as high confidence. Orchestration tracks but doesn't benchmark yet — A2A protocols aren't stable enough for a number you'd put in a board deck.

After a policy fires and an agent is restricted, the platform measures the before and after. The delta is stored. The result surfaces as a single line: "This policy improved Cost per Business Outcome by 38%." That is a board-level statement. The integrity engine is what makes it defensible.

Question three: what do you do about it?

Two things fire when an agent's economics are wrong.

The first is the Agentic Cost Ceiling — a per-session spend limit that triggers at 70% with a soft alert and at 90% with a block or escalation to human review. It fires before the invoice arrives, not after. That distinction is the entire point.

The second is a set of four OCR-driven policy triggers. Ghost by Outcome: an agent that has cost more than $0 and produced zero verified outcomes in 30 days. That is not an unused seat. It is an agent actively running, actively billing, and producing nothing. OCR = ∞. The strongest deprovisioning signal in AI FinOps. The other three — above benchmark, declining trend, 2× anomaly from baseline — catch structural inefficiency before it compounds.

None of them enforce automatically. Every trigger flows to a Decision Log, then human approval, then action. PromptKing is a visibility and control layer. It is not a runtime proxy.

Question four: what would happen before you act?

This is the one no other platform answers.

Microsoft Agent 365 has a behavioral simulation framework — does the agent comply with policy. Salesforce AgentForce has an AWU count — tasks completed. Neither answers: if I apply this economic policy to my agent fleet, what would happen to Cost per Business Outcome before I turn it on?

The Policy Simulator does. Set the policy type, the scope, the time window, the threshold. Run it against your live agent data. The output tells you how many agents would trigger, what percentage of your fleet that represents, the estimated annual cost impact, and — critically — the actionability breakdown by platform.

Not every platform supports enforcement. A Copilot Studio agent can be blocked. A Claude API integration can only receive a notification. An open-model agent on Bedrock may require manual intervention. That distinction determines whether your policy creates operational change or a notification backlog. The simulator surfaces it before you commit.

The action bridge — "Create Draft Policy from Simulation" — pre-fills the Policy Console and routes to human approval. The simulator is read-only. The board statement that comes after enforcement is not.

---

The Outcome Cost Ratio, five-category outcome taxonomy, seat archetypes, Agentic Cost Ceiling, and Policy Simulator are PromptKing-original frameworks. Ghost Seat, Underutilized, Normal, Power User, and At-Risk seat archetypes are patent pending (USPTO Application No. 2-01025019).