AI ARCHITECTURE

Choose the architecture.
Not the hype.

A visual operating map for deciding when to use retrieval, deterministic workflows, bounded agents—or no generative AI at all.

REVIEWED SEP 06, 2026PRACTITIONER-LED
SYSTEM
INTENTDATAMODELAGENTTOOLSCONTROL

Choose the least complex pattern that can earn trust.

Start with task variability, verification, reversibility, privilege and failure cost. Model choice is downstream of those operating requirements.

Run the recommender ↗
01

Variability

How much does the path change from case to case?

02

Verifiability

Can the result be checked mechanically?

03

Reversibility

Can a wrong action be safely rolled back?

04

Privilege

What systems and data can the workflow reach?

01 / KNOWLEDGE

RAG

Ground responses in retrieved information when answers depend on changing or proprietary knowledge.

USE WHENKnowledge freshness mattersMEASURERetrieval + citation fidelityWATCHStale or inaccessible sources

A prototype becomes a system only when the control boundaries are explicit.

IDENTITYAUTHORIZATIONEVALUATIONOBSERVABILITYHUMAN CONTROLRECOVERY