01
The model is downstream of the decision
Teams often begin an AI initiative by comparing models. That choice matters, but it is rarely the first architecture decision. The first decision is what the system is allowed to decide, recommend, explain, or draft—and who remains accountable.
A business rule is more than an if-statement. It expresses ownership, evidence, timing, exceptions, risk tolerance, and the operating consequence of being wrong.
02
Separate stable rules from ambiguous judgment
Deterministic software is usually preferable when inputs and outcomes can be specified reliably. AI becomes useful where language, interpretation, incomplete context, or generative assistance creates measurable value.
The boundary should be explicit. A model can classify or explain while a rule validates eligibility, enforces a limit, or requires approval.
- What evidence is required?
- Which conditions are non-negotiable?
- Where is ambiguity acceptable?
- Who can approve an exception?
- What must be recorded for later review?
03
Architecture turns policy into behavior
Governance becomes credible when it appears in authorization, workflow states, evaluation gates, audit records, escalation paths, and fallback behavior. A policy document alone cannot control a production system.
Starting with rules produces better model evaluations because the team knows what successful behavior means and which failures are consequential.
Decision guide
Questions to carry into the work
- Name the business outcome and accountable owner.
- Map the decision and its evidence before selecting technology.
- Keep deterministic constraints outside probabilistic generation.
- Define review, escalation, and audit behavior.
- Evaluate the complete decision system, not only model quality.