Executive Summary
- Authority comes before model choice: who may finalize a refund, access grant, or disposition?
- Deterministic rules own policy that must be explainable and stable.
- Workflows own multi-step orchestration, timers, and human tasks.
- RAG owns grounded retrieval and assisted generation—not silent policy changes.
- Hallucination, prompt drift, and permission leaks are operational risks requiring evaluation and fallback.
- Normal search or a rule engine is often enough; AI is not mandatory.
Suitability decision
Teams often ask whether to 'add AI.' The better question is which component should own the decision: a deterministic rule, a workflow engine, a retrieval-augmented assistant, or a human. Those owners can collaborate, but only one should have final authority for each irreversible step.
Existing guides cover how to implement RAG or compare agents and chatbots. This guide owns suitability boundaries—where AI belongs versus where rules must remain in charge.
The answer depends on risk class, explainability needs, data quality, and whether wrong answers create legal, financial, or safety harm. Do not overstate model reliability.
Authority before model choice
If you cannot name the authority for a refund, access grant, or clinical/administrative disposition, do not attach an LLM to that button.
Rules, workflows, and RAG
Deterministic decision logic encodes policy that must be stable, testable, and auditable: eligibility thresholds, entitlement checks, and compliance gates. Rule engines or plain code both count when behavior is explicit.
Business workflow orchestration sequences steps, waits, escalations, and human tasks. It should not invent policy; it should apply policy and route exceptions.
RAG combines retrieval with generation to answer from approved corpora with citations. It assists operators and users; it should not silently rewrite business rules.
Owner matrix
| Feature | Concern | Deterministic rules | Workflow | RAG / generation |
|---|---|---|---|---|
| Irreversible money/access | Primary owner | Routes to approver | Advise only | |
| Multi-step process | Guards | Primary owner | Draft content | |
| Knowledge Q&A | N/A | May trigger | Primary assist | |
| Explainability | High when coded clearly | High for path | Needs citations + limits |
Retrieval, generation, citations
Retrieval must respect permissions: users should only see chunks they are allowed to see. Mixing corpora without ACL filters creates data leaks.
Generation is uncertain. Require citations where claims matter, show uncertainty, and provide fallback to human review or conventional search when confidence is low.
Prompt and model changes are production changes. Version them, evaluate them, and do not hot-swap models on regulated paths without regression checks.
Permissions, evaluation, fallback
Sensitive data needs minimization, redaction, and retention rules before it enters prompts or vector indexes. Vendor and hosting choices affect this boundary.
Evaluation should include golden questions, refusal tests, and permission tests—not only demo transcripts. Cost and latency matter operationally; measure them honestly without invented benchmarks.
Fallback paths: conventional search, rule-based answers, or human queues. Automation may draft; humans approve high-risk outcomes.
Safe placement pattern
AI proposes; policy and humans dispose when harm potential is high.
When AI must not decide
AI must not be the final authority for irreversible money movement, access grants, safety-critical dispositions, or regulated certifications without a formal control plan.
RAG is unsuitable when the corpus is empty, stale, or unpermissioned—or when the user only needs keyword search.
Workflow automation is unsuitable as a substitute for missing policy. Agents that 'figure it out' without rules create un-auditable behavior.
A conventional rule engine or search UI is often enough. Choosing simpler tooling is a valid architecture decision.
Failure modes
Hallucinated citations, prompt injection, permission bypass via retrieval, silent prompt drift, and workflows that auto-approve based on model text are common failures.
What must remain deterministic: policy gates and entitlement checks. What can be automated: draft summaries and suggested next steps. What requires human approval: high-risk dispositions. What must be audited: model/prompt versions and override actions.
Implementation checklist
Treat missing authority maps as a blocker to AI rollout.
AI placement checklist
1Authority
- Final owner named per irreversible action
- Human approval gates listed
- Rules encoded for policy
2RAG
- Permissioned corpus
- Citation requirement
- Evaluation set
3Operations
- Prompt/model versioning
- Fallback path
- Audit logging
How Digital Elliptical helps
Digital Elliptical helps product and engineering leaders place AI appropriately: deterministic policy gates, workflow orchestration, and RAG or agent assists with approval and audit—aligned to risk class.
We do not guarantee accuracy rates, autonomy levels, or compliance certification. Suitability depends on data quality, policy clarity, and governance capacity.
Main-Agent ownership (Prompt 6): Re-authored to center decision authority, hallucination/permission controls, and explicit cases where AI must not finalize outcomes.