The conversation took place in the boardroom of a mid-sized UK professional services firm. The CFO had asked to meet the newly appointed Head of AI Governance (a role that hadn't existed twelve months earlier) about what she called "the measurement problem."
Fourteen months into AI deployment, the productivity story looked great: output volume up substantially, junior time on first drafts down significantly. But the CFO couldn't make the numbers add up. She knew, from conversations with senior partners, that the time they spent reviewing AI-generated work had risen sharply since deployment. She suspected it was quietly eating the productivity gain the dashboard was celebrating. She had no metric, no framework, and no data to prove it either way.
"We can tell the board we've saved fifteen thousand junior hours in the past year. We cannot tell the board what we've spent in senior hours to verify the outputs those hours produced. That asymmetry is a governance problem."
That asymmetry is what I built the Cost-per-Validated-Asset (CVA) to close. It's the foundational metric of the governance model in my book The Verification Economy, and it exists because conventional productivity metrics were never designed to measure verification: they were designed to measure production, and AI has broken the correlation between the two.
Before AI, a document that took three days to produce was, on average, more reliable than one that took three hours, since time invested was a crude proxy for quality. AI destroys that correlation. A document generated in three hours can look identical in surface quality to one built over three days, while being simultaneously less accurate and less defensible, because the surface quality comes from the model and the reliability depends entirely on the verification layer sitting above it.
I call what happens next the Invisible Verification Economy: the portion of an organisation's AI economics that's real, consequential, and completely absent from the management information driving strategic decisions. It includes the fully-loaded cost of senior time spent reviewing AI output, the cost of compliance events when errors slip through under-scrutinised review, and the opportunity cost of senior people doing verification instead of the strategic work that actually grows the firm. None of it shows up on a standard AI productivity dashboard. All of it is real, and it routinely dwarfs the generation savings the dashboard is celebrating.
In our research, organisations that don't measure CVA show a consistent pattern: AI gets treated as a pure cost saving, governance gets framed as a cost rather than an investment, and senior time gets systematically undervalued in how volume targets are set: three distortions that compound each other and get worse the longer they go unmeasured.
Cost-per-Validated-Asset is the true, fully-loaded cost of producing one client-ready, human-verified deliverable, not the cost of generating a first draft. It accounts for generation time, the time spent in junior first-pass verification, the time spent in senior review and authorisation, the infrastructure cost allocated per batch, and the compliance cost attributable to verification failures in the period, all divided by the number of validated assets actually produced. Three supporting metrics (Review Dwell Time, Edit Delta, and Interception Rate) sit underneath CVA and tell you why it's moving in a given direction, so a rising CVA is diagnosable rather than just alarming.
An organisation that can measure its generation costs but not its verification costs is not managing its AI economics. It is misreading them.
Once you can see CVA, it becomes a governance instrument, not just an accounting figure: it improves with real governance investment and worsens without it, which makes it one of the few AI metrics that tells executives the truth about whether their deployment is actually working.
The book gives you the complete technical specification: the full formula, every component defined, exactly what data your existing systems can already capture, and the monthly operational routine (the CVA Intelligence Cycle) that turns it from a single number into an ongoing governance practice.