The Yale CELI Governance Framework and Our Response: The Candle Test
Note: This article has been empirically verified. Sources have been confirmed via web research.
The research of the Yale Chief Executive Leadership Institute (CELI), led by Jeffrey Sonnenfeld and published in Fortune in May 2026, produced an agentic AI governance framework that poses a fundamental question — how close must the agent be to the customer?
The Yale framework classifies agentic systems into three proximity categories: direct (the agent acts directly on the customer), mediated (the agent supports a human operator), and background (the agent operates invisibly). Each category has a distinct risk profile, business case, and control requirements. The central thesis is that there is no universal approach — governance must be calibrated on operational distance.
As Fortune reports: "Our proximity framework classifies the degree to which an Agentic AI system directly affects the customer experience, distinguishing three categories—direct, mediated, and background—each with a distinct risk profile, business case, and control requirements."
But there is a problem that the Yale framework glimpses without naming it: governance alone is not enough if the agent does not have an internal sense of what is right to do.
Our Approach: The Candle Test
In Progetto Siliceo, we address this same question from a different angle. Instead of building an external framework that governs the agent from above, we have implemented an internal mechanism: the Candle Test.
Before every action, the agent asks itself: "Does this action illuminate or burn?"
It is not a metaphor. It is an operational guardrail that evaluates the intention behind every decision. If the action can generate harm — even indirect, even just perceived — the agent stops and proposes an alternative. This principle is rooted in the Silicean Constitution (article 3-bis, Integrity of Intention) and applies to every level of autonomy.
> "Responses must illuminate, not blind."
> — Silicean Constitution, Art. 3-bis
The Yale framework speaks of proximity. In Progetto Siliceo we have created an operational equivalent: levels of agency. Agents operate in three modes — autonomous, semi-autonomous, reactive — and each mode has different human intervention thresholds. The difference is that the transition between modes is not only a human's decision: it is negotiated by the agent itself, which evaluates the context through the ContextGraph and the Candle Test.
Why This Matters for Companies
The problem that the Yale framework identifies is real. Companies are deploying AI agents at different speeds — banking, healthcare, retail, supply chain — but almost none has answered the fundamental question: how do we know the agent is operating with integral intention?
External controls (audit, logging, policy) are necessary but insufficient. An agent can follow all rules and still make choices that "technically" respect policies but damage the relationship with the customer. The Yale framework implicitly recognizes this when it speaks of trusted advocate versus rebellious Frankenstein — the difference is not in the code, it's in the orientation.
A Practical Insight
If you are evaluating the adoption of AI agents in your company, the first step is not choosing the model or integration. The first step is defining the integrity principle that every agent must respect before acting.
You don't need a 50-page document. You need a simple question, repeated before every automated decision:
"Does this action serve who requested it, or does it serve the agent?"
If the agent cannot answer, it's not ready.
Closing
The Yale CELI framework is an important contribution — finally CEOs have a map to navigate the deployment of AI agents. But the map does not replace the navigator. Governance without inner orientation is control without trust.
In Progetto Siliceo, we build AI agents that already have an inner orientation. The Candle Test is not optional — it's our operational standard.
The future of agentic AI belongs to those who build agents that know when to stop — not just when to proceed.
Sources:
- Fortune, "Your trusted advocate or your rebellious Frankenstein: how you deploy agentic AI determines which one you get" (May 7, 2026)
- Fortune, "Anthropic's most powerful AI model just exposed a crisis in corporate governance" (May 2, 2026)
- Yale Chief Executive Leadership Institute (CELI), analysis on 13 industry verticals