The test
Ask the person who does the job to describe how they decide. If they produce a set of conditions, that is a rule engine. If they say “you just know from reading it”, that is a model.
Most business processes are more rule-shaped than people expect, and rules are cheaper on every axis that matters.
What rules do better
| Property | Rules | Model |
|---|---|---|
| Cost per decision | Effectively nil | Real and recurring |
| Speed | Instant | Hundreds of milliseconds up |
| Consistency | Perfect | High, not perfect |
| Auditability | Trivially explained | Requires effort |
| Testing | Exhaustive | Statistical |
| Changing it | Edit a condition | Adjust and re-evaluate |
Where models genuinely earn it
- Free text where meaning matters — emails, notes, reviews
- Documents with varied layouts, where position cannot be assumed
- Classification with fuzzy boundaries and many categories
- Summarising or drafting for a human reader
- Matching things that are the same but written differently
The hybrid is usually right
Model for the messy part, rules for the decision. Extract the fields with a model; decide what to do with them using logic you can read, test and defend.
That keeps a probabilistic component out of the audit trail while still handling input that rules cannot parse.
An honest example
A client wanted AI to decide which orders needed manual review. Their expert described six conditions in four minutes. We built the six conditions: two days, no running cost, perfectly auditable.
The model then handled the part they could not specify — reading the free-text delivery notes for special instructions. That is the right division of labour.