
Writing
On decision systems, AI governance, human-in-the-loop oversight, and the risk frameworks that bind AI to a defensible record. No vendor pitches. No predictions.
Editorial stance
Working definitions, working patterns, and the references that hold up under audit.
All Essays
20 essays — search by title, topic, or tag.
AI agent vs AI assistant comes down to autonomy, scope, tools, and persistence. Here is the four-dimension test and what each implies for oversight.
An AI audit produces four artefacts: decision log, model card, risk register, refusal log. Internal, external, and continuous modes, mapped to NIST.
AI for decision making works inside a bounded scope, refuses on irreversible commitments, and operates through a constraint gate. Here is the framework.
AI guardrails come in four types: input, output, behavioural, and policy. Where each fits in the request lifecycle, how they fail, and how to test them.
AI red teaming is structured pre-deployment testing for failure. Four test classes, the five-stage engagement, and how findings feed refusal.
Automated decision making works in three contexts and fails in three. The operational test, GDPR Article 22, and the four artefacts.
Decision intelligence is a discipline, not a dashboard. The working definition, the three pillars, and how it differs from BI, AI, and DSS.
A decision intelligence platform enforces rules before commitment. The structural test that separates real platforms from rebadged BI.
A decision support system is the classical software category. Anatomy, four DSS types, and how DSS differs from decision intelligence and decision systems.
Human-in-the-loop AI keeps a person in the decision path. The three patterns, four conditions for meaningful oversight, and the EU AI Act.
Contextual AI governance replaces static rules with situational oversight. The three-tier model, its primary focuses, and how to apply it.
A practical guide to assessing AI readiness across 7 governance dimensions: the framework, the scoring method, and the 90-day action plan.
An AI policy is not a list of prohibitions. Approved uses, data handling, refusal conditions, HITL requirements, and incident response.
Most AI risk templates are spreadsheets nobody reopens. A 12-dimension scoring model that produces actionable HITL placement instead.
Generic prompt libraries optimise for productivity. A governance prompt library optimises for accountability. Here is the distinction.
Accountability, transparency, and control are the three primary focuses of AI governance frameworks. What each requires in practice.
A decision system enforces constraints, order, and verification before commitment. The definition, examples, and how it differs from AI agents.
A practical guide to building an AI risk management framework: NIST AI RMF (Govern, Map, Measure, Manage), ISO/IEC 42001, and a rollout plan.
An AI system that cannot refuse is a liability. This essay explains why refusal is a feature, what conditions should trigger it, and how to design it in.
AI agents complete tasks. AI employees hold a bounded role with authority and accountability. Compared across scope, authority, and oversight.
Topic cluster
These essays sit around a central pillar — what a decision system is. Each one is readable independently and links to the others where the topic naturally crosses over.