Operating guides / Workflow architecture

Rules, extraction and reasoning are different jobs

A legal workflow often needs more than one technical method. Treating every step as generated reasoning makes the system harder to test and supervise.

by Hamza Suleman. Published . Operational guidance, not legal advice.

Direct answer

How should a legal AI workflow choose its method?

Use deterministic rules for stable instructions, extraction for bounded facts and generated reasoning only where language and context genuinely require it.

When should a legal workflow use rules?

Use deterministic logic when the instruction is stable: required fields, deadlines, routing, template selection and straightforward calculations. The same input should produce the same result.

When should a legal workflow use extraction?

Use extraction when the task is to locate parties, dates, clauses or defined terms. Retain the source location and confidence or failure state beside each value.

Reasoning support

Use generated analysis only where language and context require it. Define what the system may propose, what evidence must accompany it and which person decides whether it is accepted.

The architecture test

For every step, ask whether a rule can do it, whether extraction can structure it and whether generated output is actually necessary. Then define the human review and escalation path.

Design principle: use the least variable method that can complete the step safely and make failure visible.

For the wider context, read our legal AI workflow answers. Explore Margo by ClickoAI, currently in development, or use the Workflow Value Workshop to examine one repeated workflow.

Make the boundaries inspectable

Use the AI Essentials for Lawyers course and evidence index together.

The course explains the operating method. The evidence index shows how ClickoAI retains a source, population and limitation beside each public number.