New Framework Ranks AI Use Cases for Pilots and Scale
Research Lab proposes a risk‑and‑value framework that scores AI use cases, defining screening, funding and expansion decisions to guide pilots.
Source: Grafy Research LabWhat the paper introduces
Research Lab released a paper titled “Which AI Use Cases Deserve a Pilot? A Risk and Value Framework for Funding, Scoping, and Scaling AI” (arXiv 2611.20894, 17 Nov 2026). The authors propose a systematic way to decide which AI projects should move from idea to pilot, and how to manage them as they grow. By turning qualitative judgments into a composite score and an explicit decision table, the framework aims to reduce wasteful experiments while still allowing high‑impact use cases to be explored.
How the framework scores use cases
The core of the method is a composite score that ranks candidate use cases. The score is calculated from quantitative and qualitative inputs that reflect the expected net value of the AI application at a specific operating scope. The scope is defined along six dimensions: authority (who can approve the system), required human mediation, operational limits, volume of transactions, environmental constraints, and target population. By keeping the scoring formula separate from the definition of the scope, the authors can compare very different projects on a common scale.
The three decision points
The authors treat each “use case at a proposed scope” as a decision unit and define three sequential decisions:
- Screen the scope – determine whether the proposed operating parameters are acceptable for a pilot.
- Fund the experiment – allocate resources if the screened scope passes a minimum score threshold.
- Authorize expansion – allow the system to move beyond the pilot only when residual risk is judged acceptable and incremental net value is demonstrated.
Only when all three steps succeed does a project advance to a larger deployment.
Risk and value tracks
Risk is handled through an explicit “R‑track” (R1‑R4) that functions as a decision table. Each risk tier (from low‑impact R1 to high‑impact R4) includes veto conditions and a documented precedence order, ensuring that a single critical risk can halt progression regardless of the composite score. The value side, referred to as the “V‑track,” requires evidence across four pillars (the paper calls them “evi”). The exact nature of the pillars is not detailed in the summary, but they are meant to prove that each incremental expansion delivers measurable net benefit.
Applying the framework in practice
All experiments reported in the paper use synthetic data, and the authors provide a companion Python package that reproduces the scoring tables and decision logic. Organizations can plug their own metrics into the package, run the risk‑value tables, and generate a ranked shortlist of use cases ready for pilot testing. The approach is deliberately model‑agnostic; it does not prescribe a specific AI technology, only a governance structure around its deployment.
How Grafy can help
Grafy users can explore the framework without leaving the platform. The Grafy Reader lets you upload the arXiv PDF and ask natural‑language questions about the methodology, the risk table, or the scoring formula. If you want to run the accompanying Python package, the Grafy Code studio provides a ready‑to‑use notebook environment where you can install the package, feed in your own data, and visualise the decision tables. For teams that prefer a visual overview, the Grafy Graph app can be used to map the three decision points and link them to concrete project proposals, creating a living roadmap that updates as scores change.
Frequently asked questions
Does the framework tell you which AI model to use?
No. The framework focuses on governance—scoping, risk, and value—rather than model selection. It can be applied to any model or technology stack.
Can the risk track veto a project even if it has a high composite score?
Yes. The R‑track includes explicit veto conditions for each risk tier, so a single high‑impact risk (R4) can block funding regardless of the score.
Is the Python package ready for production use?
The authors released it as a research prototype and demonstrated it with synthetic data. It is suitable for experimentation and can be integrated into internal tooling, but organizations should validate it against real‑world metrics before relying on it for high‑stakes decisions.


