Test the AI Idea Against a Simple Baseline

Before adding a model to a workflow, compare it with the simplest approach that could meet the need.

For a fictional support inbox, that might mean testing an AI classifier against a small set of explicit routing rules. Use the same held-out examples for both. Record which cases each handles correctly, where each fails, and the effort needed to review and maintain it. Include edge cases, not just the easy examples.

Agree in advance what would make the AI option worth its added complexity. The decision owner should explain the trade-off and consult the people who handle the work. If the baseline is good enough, that is a useful result—not a failed AI experiment.

What workflow would you test this way, and what would count as “good enough” there?