FDE Foundations · Module 6: AI Decisions
Rules, ML, or LLM?
Rule-based systems when the logic is known and auditable; classic ML when patterns exist in labeled data; LLMs when language understanding or flexible generation is the bottleneck. Cost and auditability break ties.
12 min reading
Objectives
- Classify problems into rules, classic ML, and LLM territory
- Justify a choice with cost, accuracy, and auditability
- Explain the decision to a non-technical sponsor
The decision table
| Approach | Use when | Strengths | Watch out for |
|---|---|---|---|
| Rules | Logic is explicit, auditable, stable | Predictable, cheap, explainable | Brittleness, rule sprawl |
| Classic ML | Labeled history exists, patterns are statistical | Cheap at inference, measurable | Needs data, drifts |
| LLM | Language understanding or flexible output needed | Handles messiness, fast to build | Cost per call, hallucination, non-determinism |
The hybrid reality
Production systems are usually hybrid: an LLM extracts structured fields from free text, rules enforce business invariants on the result, and a model score ranks priority. The LLM does what it is uniquely good at (reading messiness); rules do what they are uniquely good at (guaranteeing invariants).
What the sponsor hears
Translate to their terms: accuracy target, cost per document, what happens when it is wrong, and who can audit a decision. "The model reads the invoice and proposes fields; hard rules verify totals and tax math, so arithmetic errors cannot ship" is a sentence a CFO can approve.
Write it down
The choice and its reasons belong in the brief: options considered, the one chosen, and the trigger to revisit (volume, cost, or accuracy numbers). Decisions documented at framing time do not get relitigated at delivery.
Quick check
An optional 2-3 question self-check. Answers never leave your device, are not stored, and never count toward any assessment.
Exercise
For three fictional problems (approving invoice totals under 10k, classifying support emails into 12 queues, extracting line items from PDFs), pick the approach and write the one-sentence sponsor justification for each.
Pass criteria
Each problem has a choice consistent with cost and auditability, and each justification names accuracy target, cost behavior, and the failure path in plain language.