We are an AI and ML consultancy, and roughly a third of our projects ship without a model. That is not a failure; it is the point of starting with an assessment.
The control limit
A packaging plant's rejection problem came down to two machine settings. A model we built in week six added almost nothing to the control limits from week four. We kept the limits.
The report
A distributor wanted demand forecasting. The assessment showed the real loss was stock expiring in the warehouse while the same item was reordered elsewhere. A daily ageing report fixed it; forecasting came a year later.
The form
A university wanted a conversion model. Call outcomes in the CRM were unreliable, so a model would have learned noise. A call-verification form came first.
How we tell
If a human can state the rule in one sentence after seeing the data, build the rule. If the pattern needs many weak signals combined, build the model. Either way, measure it.