August 22, 2026 · Company
From What It Costs to Whether It Was Authorized
How the question changed — and how the product followed it.
Every product starts with the question it can answer. Ours was: what is this AI spend doing?
In early 2026, enterprises were buying AI seats across Claude, Copilot, Gemini, Bedrock, and watsonx — and nobody could tell the CFO what those seats were producing. We built the answer: seat-level cost intelligence, five behavioural archetypes (Ghost, Underutilised, Normal, Power User, At-Risk), multi-vendor dashboards, and the first governed receipts that gave every AI agent run an auditable evidence trail.
That was the right question at the time. Then the problem moved.
Agents stopped being tools that a human invoked. They became autonomous workers: chaining actions, calling other agents, crossing vendor boundaries, retrying on failure, and reporting completion — all without a human reconstructing what happened. The question stopped being about cost. It became about closure.
The workflow finished. Can you prove it was the work that was authorized?
This is not the same question asked louder. It is a structurally different problem. Cost visibility tells you what was spent. Evidence tells you whether the thing that was spent on was the thing that was supposed to happen — across every system it touched.
The composition gap
Every native platform can answer its own part of the story. An identity system knows who was authorized. A runtime knows what policy was enforced. An orchestrator knows what steps completed. An observability tool knows what events it observed.
None of them answer the whole-work question: was the complete population of autonomous work that returned across agents, tools and vendors the work that was originally declared — and does the available evidence support that claim?
This is the composition gap. Local permission is not whole-work completeness. Local completion is not whole-work authorization. The gap exists because no single system was designed to prove a cross-system autonomous-work claim independently.
Enterprises fill this gap today with humans. Operations, finance, risk or audit manually reconstruct what happened across systems before the organization treats the outcome as closed. The verification scales linearly with agent activity.
How the product followed the problem
The receipts we built for cost governance turned out to be the primitive for something larger. A receipt that records what an agent was allowed to do, what it actually did, what it cost, and whether the evidence holds — that is not a FinOps artifact. That is an evidence object.
The evolution was not a pivot. It was a recognition that the receipt, the hash, the out-of-path architecture, and the stranger-verifiable verification surface were already doing evidence work. We were just calling it cost governance.
Declared Work Contract
The predeclared scope, authority, and completion conditions. What "done" and "authorized" mean, recorded before execution.
Declared Authority Envelope
The delegation boundaries: which agents, tools, data domains, and escalation paths are authorized.
AWIR
The Autonomous Work Integrity Record — the portable evidence object that binds the work claim to its underlying evidence.
Claim discipline
PromptKing knows when not to claim proof. Receipts explicitly reflect the limits of the available evidence.
The original capabilities — seat intelligence, cost reconciliation, loop detection, executive reporting — are still live. They are applications of the evidence plane, not the product definition.
Where we are
PromptKing is now the Independent Evidence Plane for consequential autonomous work. Native platforms govern and execute; PromptKing produces a portable record showing which parts of the cross-system work claim are supported, incomplete, or bounded by the available evidence.
The receipts still get a hash. The hash is still recomputable in your browser. The architecture is still out-of-path. But the question they answer is no longer about cost. It is about whether the work that finished was the work that was authorized — and whether the evidence to support that claim survives the system that produced it.
That is the problem we own now.