Feasibility
“Can it work reliably?”
Determines whether the system form is technically justified over baselines and evaluated under real decision conditions.
Is the consequential decision explicitly bounded and justified over non-AI or simpler methods?
A lab demo or impressive generation is mistaken for capability under production uncertainty.
No project proceeds without an explicit decision boundary and a credible baseline comparison.
Feasibility: Decision Statement & Non-AI Baseline Comparator
Before building an AI system, the organization must explicitly bound the consequential decision and prove that proposed AI models outperform simpler, faster, and more explainable baselines.
1. Bounded Consequential Decision Statement
2. System Form Baseline Evaluation
Comparing 4 Implementation OptionsDecision Quality Insight: If a classical XGBoost model provides 93.4% calibrated accuracy at 35ms latency and $0.08 cost, adopting a Generative LLM creates an operational burden that fails the Feasibility Gate unless non-tabular conversational unstructured inputs are explicitly required.
Evidence Conditions for Feasibility
Is the decision the system will influence explicitly defined, including the action that changes and the consequence of being wrong?
Has the proposed AI form been compared with a credible non-AI or simpler alternative, and is there evidence that AI is justified?
Is there a credible evaluation path that can test the claimed capability under conditions resembling the intended decision?
Are the principal technical assumptions and known failure modes explicit enough that they can be tested rather than assumed?
Companion Book Chapters (6)
A New Reality for AI
From Technical Ambition to Decision Discipline
Which AI Should We Build?
Problem Framing and Non-AI Baselines
The Anatomy of a Decision
Uncertainty, Regimes, and Action Windows
Evaluation Integrity, Uncertainty and Model Reality
From Benchmark Theatre to Decision-Relevant Evidence
Method Selection Under Constraints
Rules, Trees, Neural Models, and Ensembles
Retraining and Model Change
Governing System Transitions and Candidate Replacement
Operational Artefacts (3)
AI Release Case
Build a defensible case establishing that an AI system has satisfied all 6 independent evidence gates before production release.
Release Evidence Pack (AI-REP)
Workbook for recording temporal holdout validation, sliced cohort metrics, uncertainty bounds, and leakage audits.
Consequential Decision Statement
Define the specific operational decision an AI system will influence, the action that changes, and the cost of being wrong.