When Business Rules Become Too Complicated, Teach AI From Examples

01

Most automation starts with a rule

It might look like:

IF description contains "refund"
THEN category = "Billing"

Then another condition gets added. And another. Eventually the application contains dozens—or hundreds—of keyword lists, regular expressions, exceptions and special cases. That doesn't mean rules are bad. Rules are excellent when the condition is clear.

02

The problem is interpretation

Rules become difficult when people can easily understand a record but struggle to describe exactly why. Consider: “Customer says their renewal amount increased even though their number of licenses decreased.” A person might immediately recognize that as a billing issue. The word billing never appears. You can continue adding keywords. Or you can provide examples of previous requests and let a model learn the patterns.

03

Train instead of programming every variation

Suppose your historical data looks like:

Instead of writing logic for every wording variation, use those examples as training data.

04

Rules and AI can work together

This doesn't need to be an all-or-nothing decision. A good system might use:

• Incoming record → check for an exact business rule.
• Yes → use the rule result.
• No → use the StayCharted AMT model → prediction.

Use deterministic rules where the answer should always be deterministic. Use a trained model where interpretation is required.

05

Replace maintenance, not control

Reduce the rules you maintain for interpreting new records. Your application still decides:

• Which model to call
• What confidence is required
• Whether a prediction is automatic
• When a person must review it
• What action happens next

StayCharted AMT handles the prediction. Your application remains in control.

06

A practical migration path

You don't need to remove your existing rules. Start with the records that currently fall through to manual review. Train a StayCharted AMT model using previously resolved examples. Compare the predictions against your current process. If the results are useful, gradually insert the model into the workflow.