The enterprise spent three years answering one question: should we deploy AI? That question is closed. The technology is in the budget, in the workflow, and in the renewal conversation. The question that replaced it is harder and more expensive to get wrong: did it work? The reason that question has bad answers is not that AI is hard to measure. It is that the parties measuring it are the parties selling it. A vendor-neutral measurement layer changes who holds the pen, by sitting above every tool, reading the behavioral signals the work already generates, and reporting a return the tools cannot tune.
A vendor that reports the return on its own product faces the same conflict that stops an auditor from auditing a client it also advises. It defines the metric, controls the instrument, and ships a number tuned to support the next purchase order. That is not a character flaw in any company. It is a structural fact about who built the tool and why. And it means the numbers most enterprises rely on to justify AI spend are produced by the one party with the least incentive to report a disappointing result.
The stakes behind that conflict just rose. In the Futurum Group's 1H 2026 Enterprise Software Decision Maker Survey of 830 IT leaders, hard financial impact, top-line revenue and bottom-line profitability, nearly doubled to 21.7 percent of primary ROI responses, while productivity gains fell from 23.8 percent to 18.0 percent as the leading success metric. The buyer has stopped accepting "saves four hours a week" and started asking for a number tied to the P&L. That is precisely the number a selling vendor is least equipped to produce without a conflict, and precisely the number a CFO has to be able to defend.
If the Human Capital Operating System is the category, this is the measurement engine inside it. The hub explains what the intelligence layer is and why it sits above the stack. This piece opens the hood on how the number is actually produced, so the framework underneath it can be scrutinized rather than taken on faith.
Six signal families, one substrate
Work does not happen inside one tool, so measurement cannot either. A claim is drafted in one system, reviewed in a second, approved in a third, and only becomes an outcome in a fourth. Levos organizes the signals from across that stack into six families.
| Signal family | What it captures |
|---|---|
| Activity | How AI tools are actually used in the flow of work |
| Quality | Whether the output meets the bar the work requires |
| Delivery | Throughput and cycle time of work that ships |
| Revenue | Movement in commercial outcomes |
| OKR | Progress against goals declared before deployment |
| AI Impact | The residual that survives once the others are controlled for |
The OKR family is what stops anyone from defining success backward from whatever the AI happened to do, because the goals were declared first. AI Impact, the family we believe is first of its kind, is the number that asks whether AI changed the outcome, distinct from changes the team would have shown anyway. The six-family taxonomy is a Levos framework, not a peer-reviewed standard, and we say so. The individual signals map to established external frameworks for engineering, to public labor-productivity baselines for the macro check, and to the published controlled-comparison literature for the attribution method. The framework is original. The components are accountable to work that already exists. You can read the full set on the signal families page.
Controlled cohort analysis, not a vendor claim
The discipline that makes the number defensible is controlled cohort analysis. A single team's change after AI deployment is not a measurement, because the market, the hiring, and the roadmap all moved in the same window. Levos reports the matched delta: the change in the AI-using team against the change in a comparable team, matched on tenure, role mix, type of work, and tool stack. This is observational, and we never overstate it as full attribution. It approximates a controlled comparison under stated assumptions, and that honesty is the difference between a number a CFO can take to a board and a number a CFO has to caveat into meaninglessness. The method is documented in the Levos measurement methodology.
The reason this matters is in the public record. METR's July 2025 randomized trial found experienced developers were 19 percent slower with AI on familiar work while reporting themselves 20 percent faster. Brynjolfsson, Li, and Raymond, studying roughly 5,000 customer-support agents, found gains of about 14 percent on average, with the largest gains, over 30 percent, among less-experienced workers. The effect is real, large, and wildly heterogeneous by experience and context. A method that reports a single company-wide average for an effect that swings from minus 19 to plus 34 is not measuring. Cohort matching is how you stop averaging away the truth.
A reliability score honest about what it cannot see
Every Levos measurement ships with a Confidence Score from 0 to 100. It is a composite measurement-reliability index, not a statistical confidence interval, and we are precise about the difference because that is where most dashboards quietly overclaim. The score decomposes into how much of the six-family picture was available, how well each family met its sample threshold, and how strongly the families agreed. Every published score arrives with the decomposition and the inputs, so a reviewer in FP&A or people analytics can recompute it without calling us. A number you can reproduce is a measurement. A number you cannot is a dashboard.
Why this is built to last
Frameworks are easy to copy. The defensibility is not in the six family names or the matching logic, both of which any competent team could reconstruct. It is in two things that compound.
Public honesty about limits
The first is the discipline of being publicly honest about limits, which is why we are publishing the methodology openly rather than holding it as a proprietary black box. Measurement standards form when a method is documented and adopted, not when it is hidden.
A longitudinal behavioral record
The second is the longitudinal behavioral record. A benchmark built on one quarter is a guess. A benchmark built on years of matched behavioral signal across many organizations is an asset that sharpens every time it is used, and it cannot be reverse-engineered from a pricing page. The order of operations is deliberate: get the measurement honest and continuous first, then let the benchmarks and the AI Impact signal family compound on top of it.
What compounds, and what does not
The framework does not compound, because anyone can copy six family names. The record does. Each matched measurement adds to a behavioral baseline that a newcomer cannot reconstruct from a feature list, which is why the durable advantage is the data discipline, not the diagram.
The enterprise does not need another tool that tells it how much its tools are being used. It needs a layer that tells it what those tools are actually returning, in a number the vendor cannot tune and the CFO can defend. That layer was skipped in the rush to deploy. It is the one worth building now, and it is the core of the Levos platform.
Frequently asked questions
What is vendor-neutral AI measurement? Measuring the return on AI from a layer above the tools, with no product to grade. It removes the conflict that arises when the party selling the tool also reports its return.
How does a cross-tool measurement layer work? It organizes signals from across the whole stack, captures real AI use, joins it to output, and isolates the part of the change attributable to AI from the change a team would have shown anyway.
What is controlled cohort analysis? The change in an AI-using team measured against a comparable team matched on tenure, role mix, work type, and tool stack. Observational, never overstated as full attribution, with a confidence score attached.
What is a Confidence Score? A composite measurement-reliability index from 0 to 100, not a statistical confidence interval. It ships with its decomposition and inputs so a reviewer can recompute it.
See the measurement layer in your own stack
Levos is opening early access to a small cohort of enterprise leaders who want a return on AI that no single vendor can tune, read from the behavioral signal their work already generates.
