Every governed agent run gets a receipt.
Every receipt gets a hash.
Every CFO gets an answer.
The Governance OS for AI Agents
Out-of-path · Metadata-only · Agent-readable · MCP-native
PromptKing governs every AI agent across every vendor — out-of-path, metadata-only. Agent census. Policy simulation. Evidence receipts. Cost reconciliation. MCP-readable governance signals across six vendor surfaces.
PromptKing decides. Your stack enforces. The receipt proves it.
Proof Run — July 15, 2026
Six vendors. Six receipts. One law.
One identical governed prompt. Six different AI vendors. Every run got a receipt. Every receipt got a hash. The watsonx run was denied — and the agent halted immediately.
Policy halted the agent before the vendor was called
Every receipt is publicly verifiable. Click any verify link — the hash is recomputed in your browser, not by our server.
For the CFO
THE MOMENT BEFORE
Your AI invoice arrives. It's bigger than last month. Nobody in the room can name the team, the agent, or the workflow that caused it.
For the IT Director
THE MOMENT BEFORE
The Copilot renewal lands on your desk. Finance wants to know what it's buying. You have utilisation data for none of it.
For the Platform Lead
THE MOMENT BEFORE
An agent loops. Nobody sees it. By the time the bill arrives, the waste has already compounded.
USPTO Provisional 2-01025019 · Evidence Receipts · SHA-256 Verified · 6 Vendors · 18 MCP Tools · Proof Run July 15, 2026 · Zero-Context Stranger Test Passed July 31, 2026 · SOC 2 Type II target Q2 2027
COMPLEMENTARY INFRASTRUCTURE
What Agent Identity Platforms Don't Capture
Agent identity platforms answer who an agent is and what it's allowed to touch. PromptKing answers what it's costing, whether it's wasting spend, and what to do before that compounds.
PromptKing runs alongside your existing identity and security infrastructure — it does not replace it.
Architecture
Out-of-path by design
Agents talk directly to vendor APIs. PromptKing stands beside the flow — consuming metadata, issuing decisions and evidence receipts your stack enforces.
AI Agents
Vendor APIs
← consumes metadata | emits decisions ↓
SIEM · IT governance · vendor-native controls
PromptKing does not touch agent traffic. It issues governance decisions and evidence receipts your stack enforces.
Built for the buying committee
Different questions. Same system.
CISO
Evidence receipts. Halt compliance. EU AI Act deployer evidence.
“When an agent misbehaves, can you prove what governance decision was made and when?”
Every governed run produces an append-only receipt with SHA-256 integrity. Public verify URLs let auditors check every hash: receipt hashes recompute in the auditor’s own browser, and v2 obligation artifacts publish their complete envelope for independent recomputation.
CFO
Economic authorization. Waste recovery. Variance accountability.
“Your AI invoice is growing 200%+ year over year. Nobody can explain it.”
Declared vs actual cost on every run. Seat-level rightsizing. Recoverable spend surfaced before the invoice arrives.
CIO / IT Director
Agent census. Six-vendor seat fleet. Shadow agent discovery.
“You're managing AI licenses across five vendors with no unified view.”
One dashboard. Every seat. Every vendor. Ghost seats flagged. Rightsizing recommendations with confidence scores.
Agents don't just spend. They chain actions.
A single agent session can spawn sub-agents, loop, retry, and drift across tools. Each step carries cost and risk. Without mapping behavior → trajectory → outcome → projection → decision → evidence, that cost and risk stays invisible until it compounds.
The Governance OS for agent economic outcomes.
PromptKing governs every agent — out-of-path, metadata-only. It maps the full flow — behavior → trajectory → outcome → projection → decision → evidence — detects behavior patterns in real time, flags runs that breach economic thresholds, and surfaces the economic payload at every decision point — before the cost compounds.
Behavior → Trajectory → Outcome → Projection → Decision → Evidence
What PromptKing governs
Ghost seats and wasted licenses
Detects licensed AI seats with zero or near-zero activity across Claude, Copilot, Gemini, Bedrock, and Watsonx. Classifies every seat into five archetypes — Ghost, Underutilised, Normal, Power User, At-Risk — and surfaces rightsizing recommendations with confidence scores.
See Seat Fleet →
The layer that keeps agents from becoming financial liabilities.
Maps behavior → trajectory → outcome → projection → decision → evidence in real time. When an agent loops — same tool, same result, no progress — PromptKing quantifies the waste, flags it with a dollar figure before it compounds, and issues the decision your stack enforces.
See Loop Detection →Six vendors. One governance view.
Anthropic Claude, Microsoft 365 Copilot, GitHub Copilot, Google Gemini, AWS Bedrock, IBM Watsonx — unified in a single dashboard. Seat-level utilisation, cost attribution, and governance signals across every vendor your organisation runs.
See Vendor Dashboard →
EU AI Act — August 2, 2026
EU AI Act transparency and employment-context obligations bind August 2, 2026; Annex III high-risk obligations follow (deferred to Dec 2027 under the AI Omnibus). PromptKing generates the deployer evidence and human-oversight records both require.
See Compliance Dashboard →CFO, COO, and CRO reports
One-click reports formatted for each executive persona. The CFO sees spend, waste, and recoverable budget. The COO sees utilisation and efficiency. The CRO sees ROI and cost avoidance. No spreadsheets. No manual assembly.
See Executive Reports →
Simulate before enforce
Every governance action runs in simulate mode first — recording what the system would do and why, with full explainability and audit trail, before any real enforcement is applied. No black box. No surprise actions.
See Policy Engine →Two tracks. One system.
Governance and economics aren't separate problems — they're the same agent behavior viewed through different lenses.
Governance Track
Agent Census
Shadow discovery, autonomy tiers, governed coverage %
Policy Decisions
Simulate-before-enforce, explainability, audit trail
Evidence Receipts
SHA-256 integrity, append-only, public verify
MCP Tools
18 tools, agent-readable, one port for every model
Decision Log
Every verdict, every reason, every timestamp
Economics Track
Seat Intelligence
Ghost seats, archetypes, rightsizing with confidence
Spend Forecast
30-day projection, variance bands, burn rate
Cost Reconciliation
Declared vs actual, per-run attribution
Executive Reports
CFO / COO / CRO one-click personas
Vendor Health
Six vendors unified, connector status
How it works
Discover → Register → Tier → Decide → Prove → Consult
Agent census scans for shadow agents across six vendors
Each agent gets a SPIFFE identity and autonomy tier
Policy profiles map to inform / optimize / operate
Every run gets a policy verdict with full explainability
Evidence receipt issued, SHA-256 hashed, append-only
18 MCP tools let any agent read governance state
Simulate before enforce
Show, don't tell — the same trajectory, two outcomes
Toggle between simulate and enforce modes. Demo metrics (story data ◔): 34.5% cycle waste, $14,200 flagged spend.
Behavior → Trajectory → Outcome → Projection → Decision → Control
Behavior
Trajectory
Outcome
Projection
Decision
Control
Cycle waste
34.5%
Projected spend
$18,400
Same trajectory — simulate mode records the full audit trail and projected impact without touching live vendor credentials.
Open demo trajectory →Client Zero
We validated on our own tenant first.
Client Zero — completed June 30, 2026. We ran PromptKing against our own M365 Copilot tenant before asking anyone else to. No external logos. No fabricated case studies. Our connector, our seats, our utilisation data — the same pipeline we ship to customers.
Behavior → Trajectory → Outcome → Projection → Decision → Control — tested on PromptKing's own Microsoft 365 Copilot deployment.
What agent behavior caused this cost?
Should it have been allowed to run?
What would it have cost if we hadn't stopped it?
These are the questions most platforms can't answer. PromptKing was built for exactly this.
DEMO TRAJECTORY — synthetic-cycle-test-google-001
We don't measure AI spend.
We govern how it's created.
Behavior → Trajectory → Outcome → Projection → Decision → Evidence
Download the CFO waste-report template