Reasoning Governance — Institutions Must Govern Thought, Not Just Action

Reasoning Governance — Institutions and the Boundaries of Thought

Institutions have always governed systems. They know how to govern processes, workflows, permissions, and evidence. What they have never had to govern is reasoning itself. AI didn’t create that gap; it made it visible. Institutions now work with systems that reason inside their perimeter, and once that happens, the old governance machinery no longer protects the mandate. Reasoning governance is the missing layer: institutions must decide how reasoning is allowed to unfold, and which parts of reasoning must remain their own.

Governing systems is familiar work — procedural, operational, often technical. Governing reasoning is different. It asks institutions to define which forms of inference belong inside their cognitive perimeter. Consider the National Bank of Belgium receiving a model‑generated recommendation that relies on an economic assumption outside its statutory perimeter. The NBB may rely on official HICP inflation indicators, ECB‑validated forecasting models, and Eurostat macroeconomic data. A model that bases its recommendation on housing‑market sentiment scraped from real‑estate platforms, or on volatility signals extracted from crypto‑asset derivatives, may be economically plausible — but it falls outside what the NBB can treat as legitimate reasoning under its mandate. Inference routes can be analytically sound and still institutionally impermissible.

Reasoning begins long before a conclusion appears. It begins with the assumptions a system is allowed to carry. A climate‑policy model that assumes IPCC‑aligned trajectories is operating within the institution’s frame; a model that assumes speculative carbon‑capture breakthroughs is not. Assumptions shape the reasoning space long before any output is produced.

My own experience with augmented AI over nine months of working with IFRS 17 illustrates this distinction. The system could analyse complex insurance-contract questions, test assumptions and explore alternative reasoning routes, but the value lay precisely in keeping those routes within the boundaries of the applicable accounting framework. The experience did not eliminate judgement; it changed where the reasoning work could be performed.

Institutions also decide which sources count as authoritative. A legal-advice system may use statutes, case law, and regulatory guidance, while treating blogs or forum summaries only as background. Source governance therefore determines which reasoning routes are institutionally legitimate.

Reasoning must also respond to context. A risk model that continues to operate on outdated legislation is not reasoning in an institutional sense. A procurement system may ignore an expired certification and still produce a plausible recommendation, but the reasoning is already out of date. Contextual re‑evaluation is part of reasoning governance: knowing when reasoning must pause and re‑align.

Some institutional cognition cannot be delegated. A central bank may use models to analyse inflation dynamics, but the judgement of what those dynamics mean for monetary stability remains institutional.

This is where the boundary between Layer 7 and Layer 8 becomes clear.
Layer 7 governs how a system may reason: what it may infer, what it may assume, which sources it may treat as authoritative, when its reasoning must be reconsidered, and where human judgement is required.

Reasoning, interpretation, and judgement are distinct cognitive functions. Reasoning is inferential: it follows admissible routes, uses authorised assumptions, and anchors itself in recognised sources. It is procedural and can be governed. Interpretation is meaning‑giving: it determines what a phenomenon is within the institution’s mandate and epistemic frame. Judgement is mandate‑expressing: it decides how the institution must act, and carries legitimacy. Layer 7 governs reasoning. Layer 8 belongs to interpretation and judgement — the cognitive core an institution cannot delegate without altering its identity.

Layer 8 is the institution’s own cognitive layer — the part of interpretation and judgement that cannot be delegated. The Federal Agency for Medicines and Health Products interprets population‑level safety and clinical evidence. A system may analyse trial data, detect statistical signals, or model adverse‑event patterns, but the institutional meaning of those signals remains non‑delegable. The FAGG decides what constitutes sufficient evidence, how risk should be understood within its mandate, and when uncertainty becomes unacceptable. Even highly accurate analysis does not determine what those signals mean for public health. That interpretation remains tied to the institution’s authority.
The same distinction applies to the FSMA’s interpretation of market integrity and consumer protection. Analytical capability can be delegated; the judgement that expresses the mandate cannot.

Layer 7 can require human judgement without determining what that judgement must ultimately be.

This leads to the central question of reasoning governance: if institutions can determine which forms of reasoning a system may use, what determines which reasoning must remain theirs? The answer lies in Layer 8. What cannot be delegated is not analytical reasoning as such, but the interpretation and judgement through which an institution gives meaning to its mandate. It defines the institution’s role in society and expresses its identity. Once that reasoning is delegated, the institution risks surrendering part of the authority that gives its mandate meaning.

Institutions often assume that correct outputs are enough. But outputs can be correct for the wrong reasons. A fraud‑detection system may flag a case correctly while relying on demographic profiling. A financial system may predict a bank failure correctly while relying on non‑authoritative data. The problem lies deeper than technical accuracy: the system may have reached the right conclusion through a reasoning process the institution cannot accept. Institutions must govern reasoning routes, not outputs.

Institutions have governance for data, processes, evidence, and increasingly for action. What remains underdeveloped is governance of reasoning itself: the rules that determine which inferences are institutionally admissible and where interpretation and judgement must remain human and institutional.

As AI becomes part of organisational reasoning, governing reasoning becomes a matter of institutional architecture.