Before the Invoice

The invoice arrived. It was three times the forecast.

Nobody triggered it deliberately. An AgentForce orchestration chain ran longer than expected. A Copilot Studio agent retried a failing tool call eleven times. A Claude batch job processed a document set that had quietly grown by 40% since the last audit. By the time finance saw the number, the spend had already happened.

This is the agentic cost problem in one paragraph. Agents don't wait for approval to spend money. They execute autonomously, and your bill reflects every decision they make.

The standard FinOps response — look at last month's invoice and investigate — doesn't work for agentic AI. By the time the invoice arrives, the overspend has already happened. You can understand it. You cannot prevent it.

The Agentic Cost Ceiling

A Cost Ceiling is a per-agent, per-session maximum spend threshold. It fires before the invoice — at 70% of the ceiling with a soft alert, and at 90% with a block or escalation to human review.

It is not a token limit. It is not a credit cap. It is an economic control primitive that sits between the agent and the billing event.

The logic is simple:

• Set a ceiling of $3.00 per session for your AgentForce Resolution agents
• At $2.10 (70%), an alert fires to the IT Director
• At $2.70 (90%), the session is flagged for human approval before continuing
• The invoice never surprises you

Four OCR-driven policy types

The Cost Ceiling is the most immediate control. But v3.47.0 also introduces four economic policy triggers based on the Outcome Cost Ratio:

Ghost by Outcome — an agent that has cost more than $0 and produced zero verified outcomes in 30 days. Cost per Business Outcome = ∞. This is the strongest deprovisioning signal in AI FinOps. Not unused seats — agents actively running, actively billing, producing nothing.

OCR above benchmark — an agent whose cost per verified outcome exceeds the category benchmark. A Resolution agent at $4.80 per closed ticket when the market benchmark is $1.50 means something structural is wrong — wrong model, wrong prompt architecture, or wrong task assignment.

OCR declining trend — an agent whose outcome efficiency is worsening over consecutive measurement periods. Not a one-off spike. A structural degradation that compounds monthly.

OCR anomaly — an agent whose cost per outcome deviates more than 2× from its own baseline. Catches model version changes, prompt drift, and upstream data quality issues before they show up as unexplainable cost growth.

The economic loop closes

The most important capability in this sprint is the one that feels smallest: the Policy Impact panel.

After a policy fires and an agent is restricted or deprovisioned, PromptKing measures the before and after. Pre-enforcement cost per outcome vs post-enforcement cost per outcome. The delta is stored. The result surfaces in the dashboard.

"This policy improved Cost per Business Outcome by 38%."

That is a board-level statement. It means the governance layer is not just a cost center — it is an ROI generator. The platform doesn't just identify waste. It proves that removing waste improved economic efficiency.

What Microsoft and Salesforce ship

Microsoft Agent 365, generally available since May 1, 2026, governs agent identity, access, and lifecycle. It answers: which agents are running, what they can access, and what their security posture is.

Salesforce AgentForce, generally available June 15, 2026, governs agent execution and workflow orchestration. It counts Agentic Work Units — tasks completed.

Neither measures cost per verified business outcome. Neither simulates what an economic policy would do before enforcement. Neither closes the loop between governance action and economic improvement.

PromptKing governs the economic layer. Agent 365 governs who agents are. AgentForce governs how agents run. These are genuinely complementary.

The question they cannot answer is the one your CFO is asking: "Are our agents producing economic value efficiently — and what would happen if we enforced against the ones that aren't?"

That question now has an answer.

---

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