Machine learning or plain rules? Choosing the right tool honestly

Some problems described as AI projects are better solved with explicit rules. If invoices above a threshold need approval, that's an IF statement, and pretending otherwise wastes your money.

Rules win when...

• The logic is known and stable: thresholds, categories, policy checks
• Decisions must be identical every time and trivially explainable
• You have few examples of the outcome you care about
• The cost of a wrong decision is high and appeals are formal

ML wins when...

• The pattern is real but too complex to write down: fraud, demand, churn, failure
• Inputs are messy: free text, images, sensor streams
• The world drifts, and you need the system to relearn from new data
• Ranking and probability matter more than yes/no

Combining rules and models

Rules and models can work together: rules enforce policy, while models rank uncertain cases for review. For example, a support workflow can route messages using predicted topics while applying explicit access restrictions. Evaluate error rates, monitor drift and retain a fallback.

The question isn't 'can we use ML here?' It's 'what's the simplest system that makes this decision well — and keeps making it well next year?'

Ask vendors — including us — to justify the complexity they're selling. A partner willing to say 'this part is just rules' is a partner you can trust with the parts that aren't.