AI FinOps

The Decision Layer: Why PromptKing Doesn't Compete With Microsoft

4 min read

Every enterprise AI conversation in 2026 eventually reaches the same question: "Are you trying to replace Microsoft?"

No. And the architecture makes that explicit.

The six-primitive control loop

Microsoft validated PromptKing as a complete control loop across six primitives:

Observe — see every seat, session, and agent across vendors Attribute — know who is spending what, on which model, in which session type Govern — classify agents, score maturity, map Annex III risk Enforce — evaluate policy, route to approval, execute via vendor APIs Export — FOCUS 1.4 conformant billing data for CFO and FinOps tools Audit — immutable Decision Log and Action Log with full reversal path

Microsoft builds the platform — Copilot, Copilot Studio, Azure OpenAI, Graph, Entra. PromptKing does not sit in that stack. PromptKing sits between behaviour and control.

What "decision layer" means in practice

When an HR onboarding agent in Copilot Studio triggers a spend spike policy, this is what happens:

  1. PromptKing's Policy Engine evaluates the session against ten rules
  2. The violation writes a policy outcome to the Decision Log
  3. An admin reviews the violation in the Approval Queue
  4. On confirmation, PromptKing calls the Copilot Studio Quarantine API — Microsoft's own management endpoint
  5. The agent is isolated via Microsoft's governance model
  6. The Action Log records the approval, the API call, and the vendor response
  7. If reversed, a separate restore entry is appended — never modified

PromptKing influences execution using native governance APIs. It does not intercept prompts. It does not block traffic. It does not sit inline between the user and the model.

The trust statement is unchanged: metadata only, zero prompt visibility.

Why this matters for Microsoft reviewers

Microsoft's architecture review asks three questions that monitoring tools cannot answer:

Can you undo a quarantine? Yes — DELETE to the same Power Platform endpoint, logged as a separate Action Log entry.

Which log is the decision and which is the action? Decision Log (pending human review) and Action Log (executed enforcement) are separate, labelled sections.

Does this sit in our traffic path? No. Explicit annotation on every enforcement surface: "No inline interception · All enforcement via vendor management APIs."

The positioning that holds

PromptKing is not a Copilot competitor. It is not a network appliance. It is not a proxy that returns 429 before the model fires.

It is the decision layer that enterprise FinOps, IT, and compliance teams use to close the gap between "we can see the problem" and "we can prove we acted on it" — through the vendor's own APIs.

Review your enforcement log →

Configure policy rules →

See your organization's AI spend data

PromptKing connects to your AI vendors and surfaces exactly this analysis — for your seats, your vendors, your budget.

← Back to Insights