Before approving LLM use in any project, run it through four questions — one per architectural level. All four must pass. One failure means the LLM does not belong there.
The four levels come directly from the DBJ Enterprise Taxonomy: Conceptual, Logical, Physical, Implementation. The taxonomy defines what each level is responsible for. This model applies that structure to the LLM adoption decision.
Level 1 — Business Capability (Conceptual)
Does the problem actually need language reasoning?
LLM fits when the capability requires:
- Understanding unstructured text
- Contextual or fuzzy reasoning
- Synthesizing knowledge from many sources
- Handling ambiguous intent
LLM does not fit when the capability is:
- Rule execution
- Data transformation
- Structured workflow
- Search or lookup
Decision rule: If you can write the business rules down exhaustively — do not use an LLM.
Level 2 — Service Design (Logical)
Is the LLM a contained service, or is it taking over?
Acceptable:
- LLM is one bounded service with a defined input and output
- Its results are measurable (accuracy, error rate, user acceptance)
Not acceptable:
- LLM as the orchestrator of other services
- LLM as the “brain” that coordinates cross-domain decisions
- LLM as the source of business truth
Decision rule: LLM is a specialist, not a manager.
Level 3 — Operations (Physical)
Can you run this in production without surprises?
You must have answers to:
- What is the monthly cost and how does it scale?
- What is the acceptable response time?
- What happens when the LLM fails or returns bad output?
- Who monitors it and how?
Decision rule: No LLM in production without a defined cost ceiling, SLA, and fallback path.
Level 4 — Implementation (Implementation)
Can you remove or replace the LLM without rewriting the system?
Required:
- Model can be swapped (no vendor lock baked into domain logic)
- Business rules live in code, not in prompts
- A clear boundary separates the LLM from the rest of the system
Decision rule: If pulling out the LLM would collapse the architecture, it was overengineered.
Summary
LLM is the right choice only when all four conditions are met:
| Taxonomy Level | Decision Level | Condition |
|---|---|---|
| Conceptual | Business Capability | Semantic reasoning is genuinely required |
| Logical | Service Design | LLM is a bounded, replaceable service |
| Physical | Operations | Cost, SLA, and failure paths are defined |
| Implementation | Implementation | Clean boundary, no logic hidden in prompts |
One “no” anywhere — and you are building Fred Brooks’ second system.